PDFMinerLoader

This notebook provides a quick overview for getting started with PDFMiner document loader. For detailed documentation of all DocumentLoader features and configurations head to the API reference.

Overview

Integration details

Class	Package	Local	Serializable	JS support
PDFMinerLoader	langchain_community	✅	❌	❌

---

Loader features

Source	Document Lazy Loading	Native Async Support	Extract Images	Extract Tables
PDFMinerLoader	✅	❌	✅	❌

Setup

Credentials

No credentials are required to use PDFMinerLoader.

If you want to get automated best in-class tracing of your model calls you can also set your LangSmith API key by uncommenting below:

# os.environ["LANGSMITH_API_KEY"] = getpass.getpass("Enter your LangSmith API key: ")
# os.environ["LANGSMITH_TRACING"] = "true"

Installation

Install langchain_community.

%pip install -qU langchain_community pdfminer-six pillow
 

Note: you may need to restart the kernel to use updated packages.
Note: you may need to restart the kernel to use updated packages.

Initialization

Now we can instantiate our model object and load documents:

from langchain_community.document_loaders import PDFMinerLoader

file_path = "./example_data/layout-parser-paper.pdf"
loader = PDFMinerLoader(file_path)
## Load

API Reference:PDFMinerLoader

Load

docs = loader.load()
docs[0]

Document(metadata={'author': '', 'creationdate': '2021-06-22T01:27:10+00:00', 'creator': 'LaTeX with hyperref', 'keywords': '', 'moddate': '2021-06-22T01:27:10+00:00', 'ptex.fullbanner': 'This is pdfTeX, Version 3.14159265-2.6-1.40.21 (TeX Live 2020) kpathsea version 6.3.2', 'producer': 'pdfTeX-1.40.21', 'subject': '', 'title': '', 'trapped': 'False', 'total_pages': 16, 'source': './example_data/layout-parser-paper.pdf'}, page_content='1\n2\n0\n2\n\nn\nu\nJ\n\n1\n2\n\n]\n\nV\nC\n.\ns\nc\n[\n\n2\nv\n8\n4\n3\n5\n1\n.\n3\n0\n1\n2\n:\nv\ni\nX\nr\na\n\nLayoutParser: A Unified Toolkit for Deep\nLearning Based Document Image Analysis\n\nZejiang Shen1 ((cid:0)), Ruochen Zhang2, Melissa Dell3, Benjamin Charles Germain\nLee4, Jacob Carlson3, and Weining Li5\n\n1 Allen Institute for AI\nshannons@allenai.org\n2 Brown University\nruochen zhang@brown.edu\n3 Harvard University\n{melissadell,jacob carlson}@fas.harvard.edu\n4 University of Washington\nbcgl@cs.washington.edu\n5 University of Waterloo\nw422li@uwaterloo.ca\n\nAbstract. Recent advances in document image analysis (DIA) have been\nprimarily driven by the application of neural networks. Ideally, research\noutcomes could be easily deployed in production and extended for further\ninvestigation. However, various factors like loosely organized codebases\nand sophisticated model configurations complicate the easy reuse of im-\nportant innovations by a wide audience. Though there have been on-going\nefforts to improve reusability and simplify deep learning (DL) model\ndevelopment in disciplines like natural language processing and computer\nvision, none of them are optimized for challenges in the domain of DIA.\nThis represents a major gap in the existing toolkit, as DIA is central to\nacademic research across a wide range of disciplines in the social sciences\nand humanities. This paper introduces LayoutParser, an open-source\nlibrary for streamlining the usage of DL in DIA research and applica-\ntions. The core LayoutParser library comes with a set of simple and\nintuitive interfaces for applying and customizing DL models for layout de-\ntection, character recognition, and many other document processing tasks.\nTo promote extensibility, LayoutParser also incorporates a community\nplatform for sharing both pre-trained models and full document digiti-\nzation pipelines. We demonstrate that LayoutParser is helpful for both\nlightweight and large-scale digitization pipelines in real-word use cases.\nThe library is publicly available at https://layout-parser.github.io.\n\nKeywords: Document Image Analysis · Deep Learning · Layout Analysis\n· Character Recognition · Open Source library · Toolkit.\n\n1\n\nIntroduction\n\nDeep Learning(DL)-based approaches are the state-of-the-art for a wide range of\ndocument image analysis (DIA) tasks including document image classification [11,\n\x0c2\n\nZ. Shen et al.\n\n37], layout detection [38, 22], table detection [26], and scene text detection [4].\nA generalized learning-based framework dramatically reduces the need for the\nmanual specification of complicated rules, which is the status quo with traditional\nmethods. DL has the potential to transform DIA pipelines and benefit a broad\nspectrum of large-scale document digitization projects.\n\nHowever, there are several practical difficulties for taking advantages of re-\ncent advances in DL-based methods: 1) DL models are notoriously convoluted\nfor reuse and extension. Existing models are developed using distinct frame-\nworks like TensorFlow [1] or PyTorch [24], and the high-level parameters can\nbe obfuscated by implementation details [8]. It can be a time-consuming and\nfrustrating experience to debug, reproduce, and adapt existing models for DIA,\nand many researchers who would benefit the most from using these methods lack\nthe technical background to implement them from scratch. 2) Document images\ncontain diverse and disparate patterns across domains, and customized training\nis often required to achieve a desirable detection accuracy. Currently there is no\nfull-fledged infrastructure for easily curating the target document image datasets\nand fine-tuning or re-training the models. 3) DIA usually requires a sequence of\nmodels and other processing to obtain the final outputs. Often research teams use\nDL models and then perform further document analyses in separate processes,\nand these pipelines are not documented in any central location (and often not\ndocumented at all). This makes it difficult for research teams to learn about how\nfull pipelines are implemented and leads them to invest significant resources in\nreinventing the DIA wheel.\n\nLayoutParser provides a unified toolkit to support DL-based document image\nanalysis and processing. To address the aforementioned challenges, LayoutParser\nis built with the following components:\n\n1. An off-the-shelf toolkit for applying DL models for layout detection, character\n\nrecognition, and other DIA tasks (Section 3)\n\n2. A rich repository of pre-trained neural network models (Model Zoo) that\n\nunderlies the off-the-shelf usage\n\n3. Comprehensive tools for efficient document image data annotation and model\n\ntuning to support different levels of customization\n\n4. A DL model hub and community platform for the easy sharing, distribu-\ntion, and discussion of DIA models and pipelines, to promote reusability,\nreproducibility, and extensibility (Section 4)\n\nThe library implements simple and intuitive Python APIs without sacrificing\ngeneralizability and versatility, and can be easily installed via pip. Its convenient\nfunctions for handling document image data can be seamlessly integrated with\nexisting DIA pipelines. With detailed documentations and carefully curated\ntutorials, we hope this tool will benefit a variety of end-users, and will lead to\nadvances in applications in both industry and academic research.\n\nLayoutParser is well aligned with recent efforts for improving DL model\nreusability in other disciplines like natural language processing [8, 34] and com-\nputer vision [35], but with a focus on unique challenges in DIA. We show\nLayoutParser can be applied in sophisticated and large-scale digitization projects\n\x0cLayoutParser: A Unified Toolkit for DL-Based DIA\n\n3\n\nthat require precision, efficiency, and robustness, as well as simple and light-\nweight document processing tasks focusing on efficacy and flexibility (Section 5).\nLayoutParser is being actively maintained, and support for more deep learning\nmodels and novel methods in text-based layout analysis methods [37, 34] is\nplanned.\n\nThe rest of the paper is organized as follows. Section 2 provides an overview\nof related work. The core LayoutParser library, DL Model Zoo, and customized\nmodel training are described in Section 3, and the DL model hub and commu-\nnity platform are detailed in Section 4. Section 5 shows two examples of how\nLayoutParser can be used in practical DIA projects, and Section 6 concludes.\n\n2 Related Work\n\nRecently, various DL models and datasets have been developed for layout analysis\ntasks. The dhSegment [22] utilizes fully convolutional networks [20] for segmen-\ntation tasks on historical documents. Object detection-based methods like Faster\nR-CNN [28] and Mask R-CNN [12] are used for identifying document elements [38]\nand detecting tables [30, 26]. Most recently, Graph Neural Networks [29] have also\nbeen used in table detection [27]. However, these models are usually implemented\nindividually and there is no unified framework to load and use such models.\n\nThere has been a surge of interest in creating open-source tools for document\nimage processing: a search of document image analysis in Github leads to 5M\nrelevant code pieces 6; yet most of them rely on traditional rule-based methods\nor provide limited functionalities. The closest prior research to our work is the\nOCR-D project7, which also tries to build a complete toolkit for DIA. However,\nsimilar to the platform developed by Neudecker et al. [21], it is designed for\nanalyzing historical documents, and provides no supports for recent DL models.\nThe DocumentLayoutAnalysis project8 focuses on processing born-digital PDF\ndocuments via analyzing the stored PDF data. Repositories like DeepLayout9\nand Detectron2-PubLayNet10 are individual deep learning models trained on\nlayout analysis datasets without support for the full DIA pipeline. The Document\nAnalysis and Exploitation (DAE) platform [15] and the DeepDIVA project [2]\naim to improve the reproducibility of DIA methods (or DL models), yet they\nare not actively maintained. OCR engines like Tesseract [14], easyOCR11 and\npaddleOCR12 usually do not come with comprehensive functionalities for other\nDIA tasks like layout analysis.\n\nRecent years have also seen numerous efforts to create libraries for promoting\nreproducibility and reusability in the field of DL. Libraries like Dectectron2 [35],\n\n6 The number shown is obtained by specifying the search type as ‘code’.\n7 https://ocr-d.de/en/about\n8 https://github.com/BobLd/DocumentLayoutAnalysis\n9 https://github.com/leonlulu/DeepLayout\n10 https://github.com/hpanwar08/detectron2\n11 https://github.com/JaidedAI/EasyOCR\n12 https://github.com/PaddlePaddle/PaddleOCR\n\x0c4\n\nZ. Shen et al.\n\nFig. 1: The overall architecture of LayoutParser. For an input document image,\nthe core LayoutParser library provides a set of off-the-shelf tools for layout\ndetection, OCR, visualization, and storage, backed by a carefully designed layout\ndata structure. LayoutParser also supports high level customization via efficient\nlayout annotation and model training functions. These improve model accuracy\non the target samples. The community platform enables the easy sharing of DIA\nmodels and whole digitization pipelines to promote reusability and reproducibility.\nA collection of detailed documentation, tutorials and exemplar projects make\nLayoutParser easy to learn and use.\n\nAllenNLP [8] and transformers [34] have provided the community with complete\nDL-based support for developing and deploying models for general computer\nvision and natural language processing problems. LayoutParser, on the other\nhand, specializes specifically in DIA tasks. LayoutParser is also equipped with a\ncommunity platform inspired by established model hubs such as Torch Hub [23]\nand TensorFlow Hub [1]. It enables the sharing of pretrained models as well as\nfull document processing pipelines that are unique to DIA tasks.\n\nThere have been a variety of document data collections to facilitate the\ndevelopment of DL models. Some examples include PRImA [3](magazine layouts),\nPubLayNet [38](academic paper layouts), Table Bank [18](tables in academic\npapers), Newspaper Navigator Dataset [16, 17](newspaper figure layouts) and\nHJDataset [31](historical Japanese document layouts). A spectrum of models\ntrained on these datasets are currently available in the LayoutParser model zoo\nto support different use cases.\n\n3 The Core LayoutParser Library\n\nAt the core of LayoutParser is an off-the-shelf toolkit that streamlines DL-\nbased document image analysis. Five components support a simple interface\nwith comprehensive functionalities: 1) The layout detection models enable using\npre-trained or self-trained DL models for layout detection with just four lines\nof code. 2) The detected layout information is stored in carefully engineered\n\nEfficient Data AnnotationCustomized Model TrainingModel CustomizationDIA Model HubDIA Pipeline SharingCommunity PlatformLayout Detection ModelsDocument Images The Core LayoutParser LibraryOCR ModuleStorage & VisualizationLayout Data Structure\n\x0cLayoutParser: A Unified Toolkit for DL-Based DIA\n\n5\n\nTable 1: Current layout detection models in the LayoutParser model zoo\n\nDataset\n\nBase Model1 Large Model Notes\n\nPubLayNet [38]\nPRImA [3]\nNewspaper [17]\nTableBank [18]\nHJDataset [31]\n\nF / M\nM\nF\nF\nF / M\n\nM\n-\n-\nF\n-\n\nLayouts of modern scientific documents\nLayouts of scanned modern magazines and scientific reports\nLayouts of scanned US newspapers from the 20th century\nTable region on modern scientific and business document\nLayouts of history Japanese documents\n\n1 For each dataset, we train several models of different sizes for different needs (the trade-off between accuracy\nvs. computational cost). For “base model” and “large model”, we refer to using the ResNet 50 or ResNet 101\nbackbones [13], respectively. One can train models of different architectures, like Faster R-CNN [28] (F) and Mask\nR-CNN [12] (M). For example, an F in the Large Model column indicates it has a Faster R-CNN model trained\nusing the ResNet 101 backbone. The platform is maintained and a number of additions will be made to the model\nzoo in coming months.\n\nlayout data structures, which are optimized for efficiency and versatility. 3) When\nnecessary, users can employ existing or customized OCR models via the unified\nAPI provided in the OCR module. 4) LayoutParser comes with a set of utility\nfunctions for the visualization and storage of the layout data. 5) LayoutParser\nis also highly customizable, via its integration with functions for layout data\nannotation and model training. We now provide detailed descriptions for each\ncomponent.\n\n3.1 Layout Detection Models\n\nIn LayoutParser, a layout model takes a document image as an input and\ngenerates a list of rectangular boxes for the target content regions. Different\nfrom traditional methods, it relies on deep convolutional neural networks rather\nthan manually curated rules to identify content regions. It is formulated as an\nobject detection problem and state-of-the-art models like Faster R-CNN [28] and\nMask R-CNN [12] are used. This yields prediction results of high accuracy and\nmakes it possible to build a concise, generalized interface for layout detection.\nLayoutParser, built upon Detectron2 [35], provides a minimal API that can\nperform layout detection with only four lines of code in Python:\n\n1 import layoutparser as lp\n2 image = cv2 . imread ( " image_file " ) # load images\n3 model = lp . De t e c tro n2 Lay outM odel (\n\n" lp :// PubLayNet / f as t er _ r c nn _ R _ 50 _ F P N_ 3 x / config " )\n\n4\n5 layout = model . detect ( image )\n\nLayoutParser provides a wealth of pre-trained model weights using various\ndatasets covering different languages, time periods, and document types. Due to\ndomain shift [7], the prediction performance can notably drop when models are ap-\nplied to target samples that are significantly different from the training dataset. As\ndocument structures and layouts vary greatly in different domains, it is important\nto select models trained on a dataset similar to the test samples. A semantic syntax\nis used for initializing the model weights in LayoutParser, using both the dataset\nname and model name lp://<dataset-name>/<model-architecture-name>.\n\x0c6\n\nZ. Shen et al.\n\nFig. 2: The relationship between the three types of layout data structures.\nCoordinate supports three kinds of variation; TextBlock consists of the co-\nordinate information and extra features like block text, types, and reading orders;\na Layout object is a list of all possible layout elements, including other Layout\nobjects. They all support the same set of transformation and operation APIs for\nmaximum flexibility.\n\nShown in Table 1, LayoutParser currently hosts 9 pre-trained models trained\non 5 different datasets. Description of the training dataset is provided alongside\nwith the trained models such that users can quickly identify the most suitable\nmodels for their tasks. Additionally, when such a model is not readily available,\nLayoutParser also supports training customized layout models and community\nsharing of the models (detailed in Section 3.5).\n\n3.2 Layout Data Structures\n\nA critical feature of LayoutParser is the implementation of a series of data\nstructures and operations that can be used to efficiently process and manipulate\nthe layout elements. In document image analysis pipelines, various post-processing\non the layout analysis model outputs is usually required to obtain the final\noutputs. Traditionally, this requires exporting DL model outputs and then loading\nthe results into other pipelines. All model outputs from LayoutParser will be\nstored in carefully engineered data types optimized for further processing, which\nmakes it possible to build an end-to-end document digitization pipeline within\nLayoutParser. There are three key components in the data structure, namely\nthe Coordinate system, the TextBlock, and the Layout. They provide different\nlevels of abstraction for the layout data, and a set of APIs are supported for\ntransformations or operations on these classes.\n\x0cLayoutParser: A Unified Toolkit for DL-Based DIA\n\n7\n\nCoordinates are the cornerstones for storing layout information. Currently,\nthree types of Coordinate data structures are provided in LayoutParser, shown\nin Figure 2. Interval and Rectangle are the most common data types and\nsupport specifying 1D or 2D regions within a document. They are parameterized\nwith 2 and 4 parameters. A Quadrilateral class is also implemented to support\na more generalized representation of rectangular regions when the document\nis skewed or distorted, where the 4 corner points can be specified and a total\nof 8 degrees of freedom are supported. A wide collection of transformations\nlike shift, pad, and scale, and operations like intersect, union, and is_in,\nare supported for these classes. Notably, it is common to separate a segment\nof the image and analyze it individually. LayoutParser provides full support\nfor this scenario via image cropping operations crop_image and coordinate\ntransformations like relative_to and condition_on that transform coordinates\nto and from their relative representations. We refer readers to Table 2 for a more\ndetailed description of these operations13.\n\nBased on Coordinates, we implement the TextBlock class that stores both\nthe positional and extra features of individual layout elements. It also supports\nspecifying the reading orders via setting the parent field to the index of the parent\nobject. A Layout class is built that takes in a list of TextBlocks and supports\nprocessing the elements in batch. Layout can also be nested to support hierarchical\nlayout structures. They support the same operations and transformations as the\nCoordinate classes, minimizing both learning and deployment effort.\n\n3.3 OCR\n\nLayoutParser provides a unified interface for existing OCR tools. Though there\nare many OCR tools available, they are usually configured differently with distinct\nAPIs or protocols for using them. It can be inefficient to add new OCR tools into\nan existing pipeline, and difficult to make direct comparisons among the available\ntools to find the best option for a particular project. To this end, LayoutParser\nbuilds a series of wrappers among existing OCR engines, and provides nearly\nthe same syntax for using them. It supports a plug-and-play style of using OCR\nengines, making it effortless to switch, evaluate, and compare different OCR\nmodules:\n\n1 ocr_agent = lp . TesseractAgent ()\n2 # Can be easily switched to other OCR software\n3 tokens = ocr_agent . detect ( image )\n\nThe OCR outputs will also be stored in the aforementioned layout data\nstructures and can be seamlessly incorporated into the digitization pipeline.\nCurrently LayoutParser supports the Tesseract and Google Cloud Vision OCR\nengines.\n\nLayoutParser also comes with a DL-based CNN-RNN OCR model [6] trained\nwith the Connectionist Temporal Classification (CTC) loss [10]. It can be used\nlike the other OCR modules, and can be easily trained on customized datasets.\n\n13 This is also available in the LayoutParser documentation pages.\n\x0c8\n\nZ. Shen et al.\n\nTable 2: All operations supported by the layout elements. The same APIs are\nsupported across different layout element classes including Coordinate types,\nTextBlock and Layout.\n\nOperation Name\n\nDescription\n\nblock.pad(top, bottom, right, left) Enlarge the current block according to the input\n\nblock.scale(fx, fy)\n\nblock.shift(dx, dy)\n\nScale the current block given the ratio\nin x and y direction\n\nMove the current block with the shift\ndistances in x and y direction\n\nblock1.is in(block2)\n\nWhether block1 is inside of block2\n\nblock1.intersect(block2)\n\nblock1.union(block2)\n\nblock1.relative to(block2)\n\nblock1.condition on(block2)\n\nReturn the intersection region of block1 and block2.\nCoordinate type to be determined based on the inputs.\n\nReturn the union region of block1 and block2.\nCoordinate type to be determined based on the inputs.\n\nConvert the absolute coordinates of block1 to\nrelative coordinates to block2\n\nCalculate the absolute coordinates of block1 given\nthe canvas block2’s absolute coordinates\n\nblock.crop image(image)\n\nObtain the image segments in the block region\n\n3.4 Storage and visualization\n\nThe end goal of DIA is to transform the image-based document data into a\nstructured database. LayoutParser supports exporting layout data into different\nformats like JSON, csv, and will add the support for the METS/ALTO XML\nformat 14 . It can also load datasets from layout analysis-specific formats like\nCOCO [38] and the Page Format [25] for training layout models (Section 3.5).\nVisualization of the layout detection results is critical for both presentation\nand debugging. LayoutParser is built with an integrated API for displaying the\nlayout information along with the original document image. Shown in Figure 3, it\nenables presenting layout data with rich meta information and features in different\nmodes. More detailed information can be found in the online LayoutParser\ndocumentation page.\n\n3.5 Customized Model Training\n\nBesides the off-the-shelf library, LayoutParser is also highly customizable with\nsupports for highly unique and challenging document analysis tasks. Target\ndocument images can be vastly different from the existing datasets for train-\ning layout models, which leads to low layout detection accuracy. Training data\n\n14 https://altoxml.github.io\n\x0cLayoutParser: A Unified Toolkit for DL-Based DIA\n\n9\n\nFig. 3: Layout detection and OCR results visualization generated by the\nLayoutParser APIs. Mode I directly overlays the layout region bounding boxes\nand categories over the original image. Mode II recreates the original document\nvia drawing the OCR’d texts at their corresponding positions on the image\ncanvas. In this figure, tokens in textual regions are filtered using the API and\nthen displayed.\n\ncan also be highly sensitive and not sharable publicly. To overcome these chal-\nlenges, LayoutParser is built with rich features for efficient data annotation and\ncustomized model training.\n\nLayoutParser incorporates a toolkit optimized for annotating document lay-\nouts using object-level active learning [32]. With the help from a layout detection\nmodel trained along with labeling, only the most important layout objects within\neach image, rather than the whole image, are required for labeling. The rest of\nthe regions are automatically annotated with high confidence predictions from\nthe layout detection model. This allows a layout dataset to be created more\nefficiently with only around 60% of the labeling budget.\n\nAfter the training dataset is curated, LayoutParser supports different modes\nfor training the layout models. Fine-tuning can be used for training models on a\nsmall newly-labeled dataset by initializing the model with existing pre-trained\nweights. Training from scratch can be helpful when the source dataset and\ntarget are significantly different and a large training set is available. However, as\nsuggested in Studer et al.’s work[33], loading pre-trained weights on large-scale\ndatasets like ImageNet [5], even from totally different domains, can still boost\nmodel performance. Through the integrated API provided by LayoutParser,\nusers can easily compare model performances on the benchmark datasets.\n\x0c10\n\nZ. Shen et al.\n\nFig. 4: Illustration of (a) the original historical Japanese document with layout\ndetection results and (b) a recreated version of the document image that achieves\nmuch better character recognition recall. The reorganization algorithm rearranges\nthe tokens based on the their detected bounding boxes given a maximum allowed\nheight.\n\n4 LayoutParser Community Platform\n\nAnother focus of LayoutParser is promoting the reusability of layout detection\nmodels and full digitization pipelines. Similar to many existing deep learning\nlibraries, LayoutParser comes with a community model hub for distributing\nlayout models. End-users can upload their self-trained models to the model hub,\nand these models can be loaded into a similar interface as the currently available\nLayoutParser pre-trained models. For example, the model trained on the News\nNavigator dataset [17] has been incorporated in the model hub.\n\nBeyond DL models, LayoutParser also promotes the sharing of entire doc-\nument digitization pipelines. For example, sometimes the pipeline requires the\ncombination of multiple DL models to achieve better accuracy. Currently, pipelines\nare mainly described in academic papers and implementations are often not pub-\nlicly available. To this end, the LayoutParser community platform also enables\nthe sharing of layout pipelines to promote the discussion and reuse of techniques.\nFor each shared pipeline, it has a dedicated project page, with links to the source\ncode, documentation, and an outline of the approaches. A discussion panel is\nprovided for exchanging ideas. Combined with the core LayoutParser library,\nusers can easily build reusable components based on the shared pipelines and\napply them to solve their unique problems.\n\n5 Use Cases\n\nThe core objective of LayoutParser is to make it easier to create both large-scale\nand light-weight document digitization pipelines. Large-scale document processing\n\x0cLayoutParser: A Unified Toolkit for DL-Based DIA\n\n11\n\nfocuses on precision, efficiency, and robustness. The target documents may have\ncomplicated structures, and may require training multiple layout detection models\nto achieve the optimal accuracy. Light-weight pipelines are built for relatively\nsimple documents, with an emphasis on development ease, speed and flexibility.\nIdeally one only needs to use existing resources, and model training should be\navoided. Through two exemplar projects, we show how practitioners in both\nacademia and industry can easily build such pipelines using LayoutParser and\nextract high-quality structured document data for their downstream tasks. The\nsource code for these projects will be publicly available in the LayoutParser\ncommunity hub.\n\n5.1 A Comprehensive Historical Document Digitization Pipeline\n\nThe digitization of historical documents can unlock valuable data that can shed\nlight on many important social, economic, and historical questions. Yet due to\nscan noises, page wearing, and the prevalence of complicated layout structures, ob-\ntaining a structured representation of historical document scans is often extremely\ncomplicated.\nIn this example, LayoutParser was\nused to develop a comprehensive\npipeline, shown in Figure 5, to gener-\nate high-quality structured data from\nhistorical Japanese firm financial ta-\nbles with complicated layouts. The\npipeline applies two layout models to\nidentify different levels of document\nstructures and two customized OCR\nengines for optimized character recog-\nnition accuracy.\n\nAs shown in Figure 4 (a), the\ndocument contains columns of text\nwritten vertically 15, a common style\nin Japanese. Due to scanning noise\nand archaic printing technology, the\ncolumns can be skewed or have vari-\nable widths, and hence cannot be eas-\nily identified via rule-based methods.\nWithin each column, words are sepa-\nrated by white spaces of variable size,\nand the vertical positions of objects\ncan be an indicator of their layout\ntype.\n\nFig. 5: Illustration of how LayoutParser\nhelps with the historical document digi-\ntization pipeline.\n\n15 A document page consists of eight rows like this. For simplicity we skip the row\n\nsegmentation discussion and refer readers to the source code when available.\n\x0c12\n\nZ. Shen et al.\n\nTo decipher the complicated layout\n\nstructure, two object detection models have been trained to recognize individual\ncolumns and tokens, respectively. A small training set (400 images with approxi-\nmately 100 annotations each) is curated via the active learning based annotation\ntool [32] in LayoutParser. The models learn to identify both the categories and\nregions for each token or column via their distinct visual features. The layout\ndata structure enables easy grouping of the tokens within each column, and\nrearranging columns to achieve the correct reading orders based on the horizontal\nposition. Errors are identified and rectified via checking the consistency of the\nmodel predictions. Therefore, though trained on a small dataset, the pipeline\nachieves a high level of layout detection accuracy: it achieves a 96.97 AP [19]\nscore across 5 categories for the column detection model, and a 89.23 AP across\n4 categories for the token detection model.\n\nA combination of character recognition methods is developed to tackle the\nunique challenges in this document. In our experiments, we found that irregular\nspacing between the tokens led to a low character recognition recall rate, whereas\nexisting OCR models tend to perform better on densely-arranged texts. To\novercome this challenge, we create a document reorganization algorithm that\nrearranges the text based on the token bounding boxes detected in the layout\nanalysis step. Figure 4 (b) illustrates the generated image of dense text, which is\nsent to the OCR APIs as a whole to reduce the transaction costs. The flexible\ncoordinate system in LayoutParser is used to transform the OCR results relative\nto their original positions on the page.\n\nAdditionally, it is common for historical documents to use unique fonts\nwith different glyphs, which significantly degrades the accuracy of OCR models\ntrained on modern texts. In this document, a special flat font is used for printing\nnumbers and could not be detected by off-the-shelf OCR engines. Using the highly\nflexible functionalities from LayoutParser, a pipeline approach is constructed\nthat achieves a high recognition accuracy with minimal effort. As the characters\nhave unique visual structures and are usually clustered together, we train the\nlayout model to identify number regions with a dedicated category. Subsequently,\nLayoutParser crops images within these regions, and identifies characters within\nthem using a self-trained OCR model based on a CNN-RNN [6]. The model\ndetects a total of 15 possible categories, and achieves a 0.98 Jaccard score16 and\na 0.17 average Levinstein distances17 for token prediction on the test set.\n\nOverall, it is possible to create an intricate and highly accurate digitization\npipeline for large-scale digitization using LayoutParser. The pipeline avoids\nspecifying the complicated rules used in traditional methods, is straightforward\nto develop, and is robust to outliers. The DL models also generate fine-grained\nresults that enable creative approaches like page reorganization for OCR.\n\n16 This measures the overlap between the detected and ground-truth characters, and\n\nthe maximum is 1.\n\n17 This measures the number of edits from the ground-truth text to the predicted text,\n\nand lower is better.\n\x0cLayoutParser: A Unified Toolkit for DL-Based DIA\n\n13\n\nFig. 6: This lightweight table detector can identify tables (outlined in red) and\ncells (shaded in blue) in different locations on a page. In very few cases (d), it\nmight generate minor error predictions, e.g, failing to capture the top text line of\na table.\n\n5.2 A light-weight Visual Table Extractor\n\nDetecting tables and parsing their structures (table extraction) are of central im-\nportance for many document digitization tasks. Many previous works [26, 30, 27]\nand tools 18 have been developed to identify and parse table structures. Yet they\nmight require training complicated models from scratch, or are only applicable\nfor born-digital PDF documents. In this section, we show how LayoutParser can\nhelp build a light-weight accurate visual table extractor for legal docket tables\nusing the existing resources with minimal effort.\n\nThe extractor uses a pre-trained layout detection model for identifying the\ntable regions and some simple rules for pairing the rows and the columns in the\nPDF image. Mask R-CNN [12] trained on the PubLayNet dataset [38] from the\nLayoutParser Model Zoo can be used for detecting table regions. By filtering\nout model predictions of low confidence and removing overlapping predictions,\nLayoutParser can identify the tabular regions on each page, which significantly\nsimplifies the subsequent steps. By applying the line detection functions within\nthe tabular segments, provided in the utility module from LayoutParser, the\npipeline can identify the three distinct columns in the tables. A row clustering\nmethod is then applied via analyzing the y coordinates of token bounding boxes in\nthe left-most column, which are obtained from the OCR engines. A non-maximal\nsuppression algorithm is used to remove duplicated rows with extremely small\ngaps. Shown in Figure 6, the built pipeline can detect tables at different positions\non a page accurately. Continued tables from different pages are concatenated,\nand a structured table representation has been easily created.\n\n18 https://github.com/atlanhq/camelot, https://github.com/tabulapdf/tabula\n\x0c14\n\nZ. Shen et al.\n\n6 Conclusion\n\nLayoutParser provides a comprehensive toolkit for deep learning-based document\nimage analysis. The off-the-shelf library is easy to install, and can be used to\nbuild flexible and accurate pipelines for processing documents with complicated\nstructures. It also supports high-level customization and enables easy labeling and\ntraining of DL models on unique document image datasets. The LayoutParser\ncommunity platform facilitates sharing DL models and DIA pipelines, inviting\ndiscussion and promoting code reproducibility and reusability. The LayoutParser\nteam is committed to keeping the library updated continuously and bringing\nthe most recent advances in DL-based DIA, such as multi-modal document\nmodeling [37, 36, 9] (an upcoming priority), to a diverse audience of end-users.\n\nAcknowledgements We thank the anonymous reviewers for their comments\nand suggestions. This project is supported in part by NSF Grant OIA-2033558\nand funding from the Harvard Data Science Initiative and Harvard Catalyst.\nZejiang Shen thanks Doug Downey for suggestions.\n\nReferences\n\n[1] Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado,\nG.S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A.,\nIrving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg,\nJ., Man´e, D., Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J.,\nSteiner, B., Sutskever, I., Talwar, K., Tucker, P., Vanhoucke, V., Vasudevan, V.,\nVi´egas, F., Vinyals, O., Warden, P., Wattenberg, M., Wicke, M., Yu, Y., Zheng,\nX.: TensorFlow: Large-scale machine learning on heterogeneous systems (2015),\nhttps://www.tensorflow.org/, software available from tensorflow.org\n\n[2] Alberti, M., Pondenkandath, V., W¨ursch, M., Ingold, R., Liwicki, M.: Deepdiva: a\nhighly-functional python framework for reproducible experiments. In: 2018 16th\nInternational Conference on Frontiers in Handwriting Recognition (ICFHR). pp.\n423–428. IEEE (2018)\n\n[3] Antonacopoulos, A., Bridson, D., Papadopoulos, C., Pletschacher, S.: A realistic\ndataset for performance evaluation of document layout analysis. In: 2009 10th\nInternational Conference on Document Analysis and Recognition. pp. 296–300.\nIEEE (2009)\n\n[4] Baek, Y., Lee, B., Han, D., Yun, S., Lee, H.: Character region awareness for text\ndetection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and\nPattern Recognition. pp. 9365–9374 (2019)\n\n[5] Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-Scale\n\nHierarchical Image Database. In: CVPR09 (2009)\n\n[6] Deng, Y., Kanervisto, A., Ling, J., Rush, A.M.: Image-to-markup generation with\ncoarse-to-fine attention. In: International Conference on Machine Learning. pp.\n980–989. PMLR (2017)\n\n[7] Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation.\nIn: International conference on machine learning. pp. 1180–1189. PMLR (2015)\n\x0cLayoutParser: A Unified Toolkit for DL-Based DIA\n\n15\n\n[8] Gardner, M., Grus, J., Neumann, M., Tafjord, O., Dasigi, P., Liu, N., Peters,\nM., Schmitz, M., Zettlemoyer, L.: Allennlp: A deep semantic natural language\nprocessing platform. arXiv preprint arXiv:1803.07640 (2018)\n(cid:32)Lukasz Garncarek, Powalski, R., Stanis(cid:32)lawek, T., Topolski, B., Halama, P.,\nGrali´nski, F.: Lambert: Layout-aware (language) modeling using bert for in-\nformation extraction (2020)\n\n[9]\n\n[10] Graves, A., Fern´andez, S., Gomez, F., Schmidhuber, J.: Connectionist temporal\nclassification: labelling unsegmented sequence data with recurrent neural networks.\nIn: Proceedings of the 23rd international conference on Machine learning. pp.\n369–376 (2006)\n\n[11] Harley, A.W., Ufkes, A., Derpanis, K.G.: Evaluation of deep convolutional nets for\ndocument image classification and retrieval. In: 2015 13th International Conference\non Document Analysis and Recognition (ICDAR). pp. 991–995. IEEE (2015)\n[12] He, K., Gkioxari, G., Doll´ar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the\n\nIEEE international conference on computer vision. pp. 2961–2969 (2017)\n\n[13] He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.\nIn: Proceedings of the IEEE conference on computer vision and pattern recognition.\npp. 770–778 (2016)\n\n[14] Kay, A.: Tesseract: An open-source optical character recognition engine. Linux J.\n\n2007(159), 2 (Jul 2007)\n\n[15] Lamiroy, B., Lopresti, D.: An open architecture for end-to-end document analysis\nbenchmarking. In: 2011 International Conference on Document Analysis and\nRecognition. pp. 42–47. IEEE (2011)\n\n[16] Lee, B.C., Weld, D.S.: Newspaper navigator: Open faceted search for 1.5\nmillion images. In: Adjunct Publication of the 33rd Annual ACM Sym-\nposium on User\nInterface Software and Technology. p. 120–122. UIST\n’20 Adjunct, Association for Computing Machinery, New York, NY, USA\n(2020). https://doi.org/10.1145/3379350.3416143, https://doi-org.offcampus.\nlib.washington.edu/10.1145/3379350.3416143\n\n[17] Lee, B.C.G., Mears, J., Jakeway, E., Ferriter, M., Adams, C., Yarasavage, N.,\nThomas, D., Zwaard, K., Weld, D.S.: The Newspaper Navigator Dataset: Extracting\nHeadlines and Visual Content from 16 Million Historic Newspaper Pages in\nChronicling America, p. 3055–3062. Association for Computing Machinery, New\nYork, NY, USA (2020), https://doi.org/10.1145/3340531.3412767\n\n[18] Li, M., Cui, L., Huang, S., Wei, F., Zhou, M., Li, Z.: Tablebank: Table benchmark\nfor image-based table detection and recognition. arXiv preprint arXiv:1903.01949\n(2019)\n\n[19] Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll´ar, P.,\nZitnick, C.L.: Microsoft coco: Common objects in context. In: European conference\non computer vision. pp. 740–755. Springer (2014)\n\n[20] Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic\nsegmentation. In: Proceedings of the IEEE conference on computer vision and\npattern recognition. pp. 3431–3440 (2015)\n\n[21] Neudecker, C., Schlarb, S., Dogan, Z.M., Missier, P., Sufi, S., Williams, A., Wolsten-\ncroft, K.: An experimental workflow development platform for historical document\ndigitisation and analysis. In: Proceedings of the 2011 workshop on historical\ndocument imaging and processing. pp. 161–168 (2011)\n\n[22] Oliveira, S.A., Seguin, B., Kaplan, F.: dhsegment: A generic deep-learning approach\nfor document segmentation. In: 2018 16th International Conference on Frontiers\nin Handwriting Recognition (ICFHR). pp. 7–12. IEEE (2018)\n\x0c16\n\nZ. Shen et al.\n\n[23] Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z.,\nDesmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)\n[24] Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen,\nT., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style,\nhigh-performance deep learning library. arXiv preprint arXiv:1912.01703 (2019)\n[25] Pletschacher, S., Antonacopoulos, A.: The page (page analysis and ground-truth\nelements) format framework. In: 2010 20th International Conference on Pattern\nRecognition. pp. 257–260. IEEE (2010)\n\n[26] Prasad, D., Gadpal, A., Kapadni, K., Visave, M., Sultanpure, K.: Cascadetabnet:\nAn approach for end to end table detection and structure recognition from image-\nbased documents. In: Proceedings of the IEEE/CVF Conference on Computer\nVision and Pattern Recognition Workshops. pp. 572–573 (2020)\n\n[27] Qasim, S.R., Mahmood, H., Shafait, F.: Rethinking table recognition using graph\nneural networks. In: 2019 International Conference on Document Analysis and\nRecognition (ICDAR). pp. 142–147. IEEE (2019)\n\n[28] Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object\ndetection with region proposal networks. In: Advances in neural information\nprocessing systems. pp. 91–99 (2015)\n\n[29] Scarselli, F., Gori, M., Tsoi, A.C., Hagenbuchner, M., Monfardini, G.: The graph\nneural network model. IEEE transactions on neural networks 20(1), 61–80 (2008)\n[30] Schreiber, S., Agne, S., Wolf, I., Dengel, A., Ahmed, S.: Deepdesrt: Deep learning\nfor detection and structure recognition of tables in document images. In: 2017 14th\nIAPR international conference on document analysis and recognition (ICDAR).\nvol. 1, pp. 1162–1167. IEEE (2017)\n\n[31] Shen, Z., Zhang, K., Dell, M.: A large dataset of historical japanese documents\nwith complex layouts. In: Proceedings of the IEEE/CVF Conference on Computer\nVision and Pattern Recognition Workshops. pp. 548–549 (2020)\n\n[32] Shen, Z., Zhao, J., Dell, M., Yu, Y., Li, W.: Olala: Object-level active learning\n\nbased layout annotation. arXiv preprint arXiv:2010.01762 (2020)\n\n[33] Studer, L., Alberti, M., Pondenkandath, V., Goktepe, P., Kolonko, T., Fischer,\nA., Liwicki, M., Ingold, R.: A comprehensive study of imagenet pre-training for\nhistorical document image analysis. In: 2019 International Conference on Document\nAnalysis and Recognition (ICDAR). pp. 720–725. IEEE (2019)\n\n[34] Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P.,\nRault, T., Louf, R., Funtowicz, M., et al.: Huggingface’s transformers: State-of-\nthe-art natural language processing. arXiv preprint arXiv:1910.03771 (2019)\n[35] Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detectron2. https://\n\ngithub.com/facebookresearch/detectron2 (2019)\n\n[36] Xu, Y., Xu, Y., Lv, T., Cui, L., Wei, F., Wang, G., Lu, Y., Florencio, D., Zhang, C.,\nChe, W., et al.: Layoutlmv2: Multi-modal pre-training for visually-rich document\nunderstanding. arXiv preprint arXiv:2012.14740 (2020)\n\n[37] Xu, Y., Li, M., Cui, L., Huang, S., Wei, F., Zhou, M.: Layoutlm: Pre-training of\n\ntext and layout for document image understanding (2019)\n\n[38] Zhong, X., Tang, J., Yepes, A.J.: Publaynet:\n\nlayout analysis.\n\nument\nAnalysis and Recognition (ICDAR). pp. 1015–1022.\nhttps://doi.org/10.1109/ICDAR.2019.00166\n\nlargest dataset ever for doc-\nIn: 2019 International Conference on Document\nIEEE (Sep 2019).')

import pprint

pprint.pp(docs[0].metadata)

{'author': '',
 'creationdate': '2021-06-22T01:27:10+00:00',
 'creator': 'LaTeX with hyperref',
 'keywords': '',
 'moddate': '2021-06-22T01:27:10+00:00',
 'ptex.fullbanner': 'This is pdfTeX, Version 3.14159265-2.6-1.40.21 (TeX Live '
                    '2020) kpathsea version 6.3.2',
 'producer': 'pdfTeX-1.40.21',
 'subject': '',
 'title': '',
 'trapped': 'False',
 'total_pages': 16,
 'source': './example_data/layout-parser-paper.pdf'}

Lazy Load

pages = []
for doc in loader.lazy_load():
    pages.append(doc)
    if len(pages) >= 10:
        # do some paged operation, e.g.
        # index.upsert(page)

        pages = []
len(pages)

1

Metadata attribute now contains several pieces of information about the file in addition to the total number of pages.
Why is this important? If we want to reference a document, we need to determine if it’s relevant. A reference is valid if it helps the user quickly locate the fragment within the document (using the page and/or a chunk excerpt). But if the URL points to a PDF file without a page number (for various reasons) and the file has a large number of pages, we want to remove the reference that doesn’t assist the user. There’s no point in referencing a 100-page document! The total_pages metadata can then be used.

pprint.pp(pages[0].page_content[:100])
pprint.pp(pages[0].metadata)

('1\n'
 '2\n'
 '0\n'
 '2\n'
 '\n'
 'n\n'
 'u\n'
 'J\n'
 '\n'
 '1\n'
 '2\n'
 '\n'
 ']\n'
 '\n'
 'V\n'
 'C\n'
 '.\n'
 's\n'
 'c\n'
 '[\n'
 '\n'
 '2\n'
 'v\n'
 '8\n'
 '4\n'
 '3\n'
 '5\n'
 '1\n'
 '.\n'
 '3\n'
 '0\n'
 '1\n'
 '2\n'
 ':\n'
 'v\n'
 'i\n'
 'X\n'
 'r\n'
 'a\n'
 '\n'
 'LayoutParser: A Unified Too')
{'author': '',
 'creationdate': '2021-06-22T01:27:10+00:00',
 'creator': 'LaTeX with hyperref',
 'keywords': '',
 'moddate': '2021-06-22T01:27:10+00:00',
 'ptex.fullbanner': 'This is pdfTeX, Version 3.14159265-2.6-1.40.21 (TeX Live '
                    '2020) kpathsea version 6.3.2',
 'producer': 'pdfTeX-1.40.21',
 'subject': '',
 'title': '',
 'trapped': 'False',
 'total_pages': 16,
 'source': './example_data/layout-parser-paper.pdf'}

Extract the PDF by page. Each page is extracted as a langchain Document object:

loader = PDFMinerLoader(
    "./example_data/layout-parser-paper.pdf",
    mode="page",
)
docs = loader.load()
print(len(docs))

16

Extract the whole PDF as a single langchain Document object:

loader = PDFMinerLoader(
    "./example_data/layout-parser-paper.pdf",
    mode="single",
)
docs = loader.load()
print(len(docs))

1

Add a custom pages_delimitor to identify where are ends of pages in single mode:

loader = PDFMinerLoader(
    "./example_data/layout-parser-paper.pdf",
    mode="single",
    pages_delimitor="\n-------THIS IS A CUSTOM END OF PAGE-------\n",
)
docs = loader.load()
print(docs[0].page_content[:5780])

1
2
0
2

n
u
J

1
2

]

V
C
.
s
c
[

2
v
8
4
3
5
1
.
3
0
1
2
:
v
i
X
r
a

LayoutParser: A Unified Toolkit for Deep
Learning Based Document Image Analysis

Zejiang Shen1 ((cid:0)), Ruochen Zhang2, Melissa Dell3, Benjamin Charles Germain
Lee4, Jacob Carlson3, and Weining Li5

1 Allen Institute for AI
shannons@allenai.org
2 Brown University
ruochen zhang@brown.edu
3 Harvard University
{melissadell,jacob carlson}@fas.harvard.edu
4 University of Washington
bcgl@cs.washington.edu
5 University of Waterloo
w422li@uwaterloo.ca

Abstract. Recent advances in document image analysis (DIA) have been
primarily driven by the application of neural networks. Ideally, research
outcomes could be easily deployed in production and extended for further
investigation. However, various factors like loosely organized codebases
and sophisticated model configurations complicate the easy reuse of im-
portant innovations by a wide audience. Though there have been on-going
efforts to improve reusability and simplify deep learning (DL) model
development in disciplines like natural language processing and computer
vision, none of them are optimized for challenges in the domain of DIA.
This represents a major gap in the existing toolkit, as DIA is central to
academic research across a wide range of disciplines in the social sciences
and humanities. This paper introduces LayoutParser, an open-source
library for streamlining the usage of DL in DIA research and applica-
tions. The core LayoutParser library comes with a set of simple and
intuitive interfaces for applying and customizing DL models for layout de-
tection, character recognition, and many other document processing tasks.
To promote extensibility, LayoutParser also incorporates a community
platform for sharing both pre-trained models and full document digiti-
zation pipelines. We demonstrate that LayoutParser is helpful for both
lightweight and large-scale digitization pipelines in real-word use cases.
The library is publicly available at https://layout-parser.github.io.

Keywords: Document Image Analysis · Deep Learning · Layout Analysis
· Character Recognition · Open Source library · Toolkit.

1

Introduction

Deep Learning(DL)-based approaches are the state-of-the-art for a wide range of
document image analysis (DIA) tasks including document image classification [11,
-------THIS IS A CUSTOM END OF PAGE-------
2

Z. Shen et al.

37], layout detection [38, 22], table detection [26], and scene text detection [4].
A generalized learning-based framework dramatically reduces the need for the
manual specification of complicated rules, which is the status quo with traditional
methods. DL has the potential to transform DIA pipelines and benefit a broad
spectrum of large-scale document digitization projects.

However, there are several practical difficulties for taking advantages of re-
cent advances in DL-based methods: 1) DL models are notoriously convoluted
for reuse and extension. Existing models are developed using distinct frame-
works like TensorFlow [1] or PyTorch [24], and the high-level parameters can
be obfuscated by implementation details [8]. It can be a time-consuming and
frustrating experience to debug, reproduce, and adapt existing models for DIA,
and many researchers who would benefit the most from using these methods lack
the technical background to implement them from scratch. 2) Document images
contain diverse and disparate patterns across domains, and customized training
is often required to achieve a desirable detection accuracy. Currently there is no
full-fledged infrastructure for easily curating the target document image datasets
and fine-tuning or re-training the models. 3) DIA usually requires a sequence of
models and other processing to obtain the final outputs. Often research teams use
DL models and then perform further document analyses in separate processes,
and these pipelines are not documented in any central location (and often not
documented at all). This makes it difficult for research teams to learn about how
full pipelines are implemented and leads them to invest significant resources in
reinventing the DIA wheel.

LayoutParser provides a unified toolkit to support DL-based document image
analysis and processing. To address the aforementioned challenges, LayoutParser
is built with the following components:

1. An off-the-shelf toolkit for applying DL models for layout detection, character

recognition, and other DIA tasks (Section 3)

2. A rich repository of pre-trained neural network models (Model Zoo) that

underlies the off-the-shelf usage

3. Comprehensive tools for efficient document image data annotation and model

tuning to support different levels of customization

4. A DL model hub and community platform for the easy sharing, distribu-
tion, and discussion of DIA models and pipelines, to promote reusability,
reproducibility, and extensibility (Section 4)

The library implements simple and intuitive Python APIs without sacrificing
generalizability and versatility, and can be easily installed via pip. Its convenient
functions for handling document image data can be seamlessly integrated with
existing DIA pipelines. With detailed documentations and carefully curated
tutorials, we hope this tool will benefit a variety of end-users, and will lead to
advances in applications in both industry and academic research.

LayoutParser is well aligned with recent efforts for improving DL model
reusability in other disciplines like natural language processing [8, 34] and com-
puter vision [35], but with a focus on unique challenges in DIA. We show
LayoutParser can be applied in sophisticated and large-scale digitization projects
-------THIS IS A CUSTOM END OF PAGE-------
LayoutParser: A Unified Toolkit for DL-Based DIA

3

that require precision,

This could simply be \n, or \f to clearly indicate a page change, or <!-- PAGE BREAK --> for seamless injection in a Markdown viewer without a visual effect.

Why is it important to identify page breaks when retrieving the full document flow? Because we generally want to provide a URL with the chunk’s location when the LLM answers. While it’s possible to reference the entire PDF, this isn’t practical if it’s more than two pages long. It’s better to indicate the specific page to display in the URL. Therefore, assistance is needed so that chunking algorithms can add the page metadata to each chunk. The choice of delimiter helps the algorithm prioritize this parameter.

Extract images from the PDF

In LangChain the OCR process used by the parsers involves asking the parser for the text on a page, then retrieving images to apply OCR.
In the previous implementation the text extracted from images was appended to the end of the page text. In a RAG context it means if in the original document a paragraph is spread across two pages it would have been cut in half by the OCR process putting the text (from the image) in between, worsening RAG model's performance.

To avoid this, we modified the strategy for injecting OCR results from images. Now, the result is inserted between the last and the second-to-last paragraphs of text (\n\n or \n) of the page.

Extract images from the PDF in html format (can also be markdown or text) with rapid OCR:

%pip install -q rapidocr-onnxruntime

Note: you may need to restart the kernel to use updated packages.

from langchain_community.document_loaders.parsers.pdf import (
    convert_images_to_text_with_rapidocr,
)

loader = PDFMinerLoader(
    "./example_data/layout-parser-paper.pdf",
    mode="page",
    extract_images=True,
    images_to_text=convert_images_to_text_with_rapidocr(format="html"),
)
docs = loader.load()

print(docs[5].page_content)

API Reference:convert_images_to_text_with_rapidocr

6

Z. Shen et al.

Fig. 2: The relationship between the three types of layout data structures.
Coordinate supports three kinds of variation; TextBlock consists of the co-
ordinate information and extra features like block text, types, and reading orders;
a Layout object is a list of all possible layout elements, including other Layout
objects. They all support the same set of transformation and operation APIs for
maximum flexibility.

Shown in Table 1, LayoutParser currently hosts 9 pre-trained models trained
on 5 different datasets. Description of the training dataset is provided alongside
with the trained models such that users can quickly identify the most suitable
models for their tasks. Additionally, when such a model is not readily available,
LayoutParser also supports training customized layout models and community
sharing of the models (detailed in Section 3.5).

3.2 Layout Data Structures

A critical feature of LayoutParser is the implementation of a series of data
structures and operations that can be used to efficiently process and manipulate
the layout elements. In document image analysis pipelines, various post-processing
on the layout analysis model outputs is usually required to obtain the final
outputs. Traditionally, this requires exporting DL model outputs and then loading
the results into other pipelines. All model outputs from LayoutParser will be
stored in carefully engineered data types optimized for further processing, which
makes it possible to build an end-to-end document digitization pipeline within
LayoutParser. There are three key components in the data structure, namely
the Coordinate system, the TextBlock, and the Layout. They provide different
levels of abstraction for the layout data, and a set of APIs are supported for
transformations or operations on these classes.



<img alt="Coordinate
(x1, y1)
(X1, y1)
(x2,y2)
APIS
x-interval
tart
end
Quadrilateral
operation
Rectangle
y-interval
ena
(x2, y2)
(x4, y4)
(x3, y3)
and
textblock
Coordinate
transformation
+
Block
Block
Reading
Extra features
Text
Type
Order
coordinatel
textblockl
layout
 same
textblock2
layoutl
The
A list of the layout elements" />

RapidOCR is designed to work with Chinese and English, not other languages. Since the implementation uses a function rather than a method, it’s not possible to modify it. We have modified the various parsers to allow for selecting other algorithm to analyze images. Now, it’s possible to also use Tesseract, or invoke a multimodal LLM to get a description of the image.

Extract images from the PDF in text format (can also be html or markdown) with tesseract:

%pip install -qU pytesseract

Note: you may need to restart the kernel to use updated packages.

from langchain_community.document_loaders.parsers.pdf import (
    convert_images_to_text_with_tesseract,
)

loader = PDFMinerLoader(
    "./example_data/layout-parser-paper.pdf",
    mode="page",
    extract_images=True,
    images_to_text=convert_images_to_text_with_tesseract(format="text"),
)
docs = loader.load()
print(docs[5].page_content)

API Reference:convert_images_to_text_with_tesseract

6

Z. Shen et al.

Fig. 2: The relationship between the three types of layout data structures.
Coordinate supports three kinds of variation; TextBlock consists of the co-
ordinate information and extra features like block text, types, and reading orders;
a Layout object is a list of all possible layout elements, including other Layout
objects. They all support the same set of transformation and operation APIs for
maximum flexibility.

Shown in Table 1, LayoutParser currently hosts 9 pre-trained models trained
on 5 different datasets. Description of the training dataset is provided alongside
with the trained models such that users can quickly identify the most suitable
models for their tasks. Additionally, when such a model is not readily available,
LayoutParser also supports training customized layout models and community
sharing of the models (detailed in Section 3.5).

3.2 Layout Data Structures

A critical feature of LayoutParser is the implementation of a series of data
structures and operations that can be used to efficiently process and manipulate
the layout elements. In document image analysis pipelines, various post-processing
on the layout analysis model outputs is usually required to obtain the final
outputs. Traditionally, this requires exporting DL model outputs and then loading
the results into other pipelines. All model outputs from LayoutParser will be
stored in carefully engineered data types optimized for further processing, which
makes it possible to build an end-to-end document digitization pipeline within
LayoutParser. There are three key components in the data structure, namely
the Coordinate system, the TextBlock, and the Layout. They provide different
levels of abstraction for the layout data, and a set of APIs are supported for
transformations or operations on these classes.



Coordinate

textblock

x-interval

 

JeAsaqul-A

Coordinate
+

Extra features

Rectangle

Quadrilateral

 

Block
Text

 

 

 

Block
Type

 

 

Reading
Order

 

 

layout

[ coordinatel1 textblock1 |
'

“y textblock2 , layout1 ]

A list of the layout elements

The same transformation and operation APIs

Extract images from the PDF in markdown format (can also be html or text) with vision model:

%pip install -qU langchain_openai

Note: you may need to restart the kernel to use updated packages.

import os
from dotenv import load_dotenv

# Charger les variables d'environnement à partir du fichier .env
load_dotenv()

True

from getpass import getpass

if not os.environ.get("OPENAI_API_KEY"):
    os.environ["OPENAI_API_KEY"] = getpass("OpenAI API key =")

from langchain_openai import ChatOpenAI
from langchain_community.document_loaders.parsers.pdf import (
    convert_images_to_description,
)

loader = PDFMinerLoader(
    "./example_data/layout-parser-paper.pdf",
    mode="page",
    extract_images=True,
    images_to_text=convert_images_to_description(
        model=ChatOpenAI(model="gpt-4o", max_tokens=1024), format="markdown"
    ),
)
docs = loader.load()
print(docs[5].page_content)

API Reference:ChatOpenAI | convert_images_to_description

6

Z. Shen et al.

Fig. 2: The relationship between the three types of layout data structures.
Coordinate supports three kinds of variation; TextBlock consists of the co-
ordinate information and extra features like block text, types, and reading orders;
a Layout object is a list of all possible layout elements, including other Layout
objects. They all support the same set of transformation and operation APIs for
maximum flexibility.

Shown in Table 1, LayoutParser currently hosts 9 pre-trained models trained
on 5 different datasets. Description of the training dataset is provided alongside
with the trained models such that users can quickly identify the most suitable
models for their tasks. Additionally, when such a model is not readily available,
LayoutParser also supports training customized layout models and community
sharing of the models (detailed in Section 3.5).

3.2 Layout Data Structures

A critical feature of LayoutParser is the implementation of a series of data
structures and operations that can be used to efficiently process and manipulate
the layout elements. In document image analysis pipelines, various post-processing
on the layout analysis model outputs is usually required to obtain the final
outputs. Traditionally, this requires exporting DL model outputs and then loading
the results into other pipelines. All model outputs from LayoutParser will be
stored in carefully engineered data types optimized for further processing, which
makes it possible to build an end-to-end document digitization pipeline within
LayoutParser. There are three key components in the data structure, namely
the Coordinate system, the TextBlock, and the Layout. They provide different
levels of abstraction for the layout data, and a set of APIs are supported for
transformations or operations on these classes.



![**Summary:**
This image illustrates a hierarchical structure for organizing layout elements using coordinates and text blocks. It shows the use of coordinates in defining shapes like rectangles and quadrilaterals, and describes combining these with extra features in text blocks. The layout is depicted as a list of elements that includes coordinates and text blocks, all processed by the same transformation and operation APIs.

**Extracted Text:**
- Coordinate
- x-interval
- y-interval
- Rectangle
- Quadrilateral
- textblock
- Coordinate +
- Extra features
- Block Text
- Block Type
- Reading Order
- layout
- [ coordinate1, textblock1, ..., textblock2, layout1 \\]
- A list of the layout elements
- The same transformation and operation APIs](.)

Using PDFMiner to generate HTML text

This can be helpful for chunking texts semantically into sections as the output html content can be parsed via BeautifulSoup to get more structured and rich information about font size, page numbers, PDF headers / footers, etc.

from langchain_community.document_loaders import PDFMinerPDFasHTMLLoader

file_path = "./example_data/layout-parser-paper.pdf"
loader = PDFMinerPDFasHTMLLoader(file_path)
docs = loader.load()
docs[0]

API Reference:PDFMinerPDFasHTMLLoader

Document(metadata={'source': './example_data/layout-parser-paper.pdf'}, page_content='<html><head>\n<meta http-equiv="Content-Type" content="text/html">\n</head><body>\n<span style="position:absolute; border: gray 1px solid; left:0px; top:50px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:50px;"><a name="1">Page 1</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:16px; top:263px; width:20px; height:40px;"><span style="font-family: Times-Roman; font-size:10px">1\n<br>2\n<br>0\n<br>2\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:16px; top:308px; width:20px; height:27px;"><span style="font-family: Times-Roman; font-size:10px">n\n<br>u\n<br></span><span style="font-family: Times-Roman; font-size:7px">J\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:16px; top:341px; width:20px; height:20px;"><span style="font-family: Times-Roman; font-size:10px">1\n<br>2\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:16px; top:371px; width:20px; height:6px;"><span style="font-family: Times-Roman; font-size:6px">]\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:16px; top:377px; width:20px; height:56px;"><span style="font-family: Times-Roman; font-size:14px">V\n<br></span><span style="font-family: Times-Roman; font-size:13px">C\n<br></span><span style="font-family: Times-Roman; font-size:5px">.\n<br></span><span style="font-family: Times-Roman; font-size:7px">s\n<br></span><span style="font-family: Times-Roman; font-size:8px">c\n<br></span><span style="font-family: Times-Roman; font-size:6px">[\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:16px; top:443px; width:20px; height:166px;"><span style="font-family: Times-Roman; font-size:10px">2\n<br>v\n<br>8\n<br>4\n<br>3\n<br>5\n<br>1\n<br></span><span style="font-family: Times-Roman; font-size:5px">.\n<br></span><span style="font-family: Times-Roman; font-size:10px">3\n<br>0\n<br>1\n<br>2\n<br></span><span style="font-family: Times-Roman; font-size:5px">:\n<br></span><span style="font-family: Times-Roman; font-size:10px">v\n<br></span><span style="font-family: Times-Roman; font-size:5px">i\n<br></span><span style="font-family: Times-Roman; font-size:14px">X\n<br></span><span style="font-family: Times-Roman; font-size:6px">r\n<br></span><span style="font-family: Times-Roman; font-size:8px">a\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:157px; top:164px; width:300px; height:32px;"><span style="font-family: CMTT12; font-size:14px">LayoutParser</span><span style="font-family: CMBX12; font-size:14px">: A Unified Toolkit for Deep\n<br>Learning Based Document Image Analysis\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:218px; width:345px; height:23px;"><span style="font-family: CMR10; font-size:9px">Zejiang Shen</span><span style="font-family: CMR7; font-size:6px">1 </span><span style="font-family: CMR10; font-size:9px">(</span><span style="font-family: unknown; font-size:9px">(cid:0)</span><span style="font-family: CMR10; font-size:9px">), Ruochen Zhang</span><span style="font-family: CMR7; font-size:6px">2</span><span style="font-family: CMR10; font-size:9px">, Melissa Dell</span><span style="font-family: CMR7; font-size:6px">3</span><span style="font-family: CMR10; font-size:9px">, Benjamin Charles Germain\n<br>Lee</span><span style="font-family: CMR7; font-size:6px">4</span><span style="font-family: CMR10; font-size:9px">, Jacob Carlson</span><span style="font-family: CMR7; font-size:6px">3</span><span style="font-family: CMR10; font-size:9px">, and Weining Li</span><span style="font-family: CMR7; font-size:6px">5\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:207px; top:252px; width:200px; height:109px;"><span style="font-family: CMR6; font-size:5px">1 </span><span style="font-family: CMR9; font-size:8px">Allen Institute for AI\n<br></span><span style="font-family: CMTT9; font-size:8px">shannons@allenai.org\n<br></span><span style="font-family: CMR6; font-size:5px">2 </span><span style="font-family: CMR9; font-size:8px">Brown University\n<br></span><span style="font-family: CMTT9; font-size:8px">ruochen zhang@brown.edu\n<br></span><span style="font-family: CMR6; font-size:5px">3 </span><span style="font-family: CMR9; font-size:8px">Harvard University\n<br></span><span style="font-family: CMSY9; font-size:8px">{</span><span style="font-family: CMTT9; font-size:8px">melissadell,jacob carlson</span><span style="font-family: CMSY9; font-size:8px">}</span><span style="font-family: CMTT9; font-size:8px">@fas.harvard.edu\n<br></span><span style="font-family: CMR6; font-size:5px">4 </span><span style="font-family: CMR9; font-size:8px">University of Washington\n<br></span><span style="font-family: CMTT9; font-size:8px">bcgl@cs.washington.edu\n<br></span><span style="font-family: CMR6; font-size:5px">5 </span><span style="font-family: CMR9; font-size:8px">University of Waterloo\n<br></span><span style="font-family: CMTT9; font-size:8px">w422li@uwaterloo.ca\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:162px; top:388px; width:291px; height:228px;"><span style="font-family: CMBX9; font-size:8px">Abstract. </span><span style="font-family: CMR9; font-size:8px">Recent advances in document image analysis (DIA) have been\n<br>primarily driven by the application of neural networks. Ideally, research\n<br>outcomes could be easily deployed in production and extended for further\n<br>investigation. However, various factors like loosely organized codebases\n<br>and sophisticated model configurations complicate the easy reuse of im-\n<br>portant innovations by a wide audience. Though there have been on-going\n<br>efforts to improve reusability and simplify deep learning (DL) model\n<br>development in disciplines like natural language processing and computer\n<br>vision, none of them are optimized for challenges in the domain of DIA.\n<br>This represents a major gap in the existing toolkit, as DIA is central to\n<br>academic research across a wide range of disciplines in the social sciences\n<br>and humanities. This paper introduces </span><span style="font-family: CMTT9; font-size:8px">LayoutParser</span><span style="font-family: CMR9; font-size:8px">, an open-source\n<br>library for streamlining the usage of DL in DIA research and applica-\n<br>tions. The core </span><span style="font-family: CMTT9; font-size:8px">LayoutParser </span><span style="font-family: CMR9; font-size:8px">library comes with a set of simple and\n<br>intuitive interfaces for applying and customizing DL models for layout de-\n<br>tection, character recognition, and many other document processing tasks.\n<br>To promote extensibility, </span><span style="font-family: CMTT9; font-size:8px">LayoutParser </span><span style="font-family: CMR9; font-size:8px">also incorporates a community\n<br>platform for sharing both pre-trained models and full document digiti-\n<br></span><span style="font-family: CMR9; font-size:8px">zation pipelines. We demonstrate that </span><span style="font-family: CMTT9; font-size:8px">LayoutParser </span><span style="font-family: CMR9; font-size:8px">is helpful for both\n<br>lightweight and large-scale digitization pipelines in real-word use cases.\n<br>The library is publicly available at </span><span style="font-family: CMTT9; font-size:8px">https://layout-parser.github.io</span><span style="font-family: CMR9; font-size:8px">.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:162px; top:627px; width:289px; height:21px;"><span style="font-family: CMBX9; font-size:8px">Keywords: </span><span style="font-family: CMR9; font-size:8px">Document Image Analysis </span><span style="font-family: SFRM0900; font-size:8px">· </span><span style="font-family: CMR9; font-size:8px">Deep Learning </span><span style="font-family: SFRM0900; font-size:8px">· </span><span style="font-family: CMR9; font-size:8px">Layout Analysis\n<br></span><span style="font-family: SFRM0900; font-size:8px">· </span><span style="font-family: CMR9; font-size:8px">Character Recognition </span><span style="font-family: SFRM0900; font-size:8px">· </span><span style="font-family: CMR9; font-size:8px">Open Source library </span><span style="font-family: SFRM0900; font-size:8px">· </span><span style="font-family: CMR9; font-size:8px">Toolkit.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:669px; width:6px; height:11px;"><span style="font-family: CMBX12; font-size:11px">1\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:154px; top:669px; width:74px; height:11px;"><span style="font-family: CMBX12; font-size:11px">Introduction\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:692px; width:347px; height:21px;"><span style="font-family: CMR10; font-size:9px">Deep Learning(DL)-based approaches are the state-of-the-art for a wide range of\n<br></span><span style="font-family: CMR10; font-size:9px">document image analysis (DIA) tasks including document image classification [11,\n<br></span></div><span style="position:absolute; border: black 1px solid; left:287px; top:293px; width:2px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:292px; top:315px; width:2px; height:0px;"></span>\n<span style="font-family: Times-Roman; font-size:5px"> \n<br> \n<br> \n<br> \n<br> \n<br> \n<br><span style="position:absolute; border: gray 1px solid; left:0px; top:892px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:892px;"><a name="2">Page 2</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:984px; width:4px; height:8px;"><span style="font-family: CMR9; font-size:8px">2\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:167px; top:984px; width:54px; height:8px;"><span style="font-family: CMR9; font-size:8px">Z. Shen et al.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:1009px; width:348px; height:57px;"><span style="font-family: CMR10; font-size:9px">37], layout detection [38, 22], table detection [26], and scene text detection [4].\n<br>A generalized learning-based framework dramatically reduces the need for the\n<br>manual specification of complicated rules, which is the status quo with traditional\n<br>methods. DL has the potential to transform DIA pipelines and benefit a broad\n<br>spectrum of large-scale document digitization projects.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:1069px; width:348px; height:213px;"><span style="font-family: CMR10; font-size:9px">However, there are several practical difficulties for taking advantages of re-\n<br>cent advances in DL-based methods: 1) DL models are notoriously convoluted\n<br>for reuse and extension. Existing models are developed using distinct frame-\n<br>works like TensorFlow [1] or PyTorch [24], and the high-level parameters can\n<br>be obfuscated by implementation details [8]. It can be a time-consuming and\n<br>frustrating experience to debug, reproduce, and adapt existing models for DIA,\n<br>and </span><span style="font-family: CMTI10; font-size:9px">many researchers who would benefit the most from using these methods lack\n<br>the technical background to implement them from scratch. </span><span style="font-family: CMR10; font-size:9px">2) Document images\n<br></span><span style="font-family: CMR10; font-size:9px">contain diverse and disparate patterns across domains, and customized training\n<br></span><span style="font-family: CMR10; font-size:9px">is often required to achieve a desirable detection accuracy. Currently </span><span style="font-family: CMTI10; font-size:9px">there is no\n<br>full-fledged infrastructure for easily curating the target document image datasets\n<br>and fine-tuning or re-training the models. </span><span style="font-family: CMR10; font-size:9px">3) DIA usually requires a sequence of\n<br>models and other processing to obtain the final outputs. Often research teams use\n<br>DL models and then perform further document analyses in separate processes,\n<br>and these pipelines are not documented in any central location (and often not\n<br>documented at all). This makes it </span><span style="font-family: CMTI10; font-size:9px">difficult for research teams to learn about how\n<br>full pipelines are implemented </span><span style="font-family: CMR10; font-size:9px">and </span><span style="font-family: CMTI10; font-size:9px">leads them to invest significant resources in\n<br>reinventing the DIA wheel</span><span style="font-family: CMR10; font-size:9px">.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:1284px; width:346px; height:33px;"><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">provides a unified toolkit to support DL-based document image\n<br>analysis and processing. To address the aforementioned challenges, </span><span style="font-family: CMTT10; font-size:9px">LayoutParser\n<br></span><span style="font-family: CMR10; font-size:9px">is built with the following components:\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:138px; top:1326px; width:341px; height:9px;"><span style="font-family: CMR10; font-size:9px">1. An off-the-shelf toolkit for applying DL models for layout detection, character\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:151px; top:1338px; width:194px; height:9px;"><span style="font-family: CMR10; font-size:9px">recognition, and other DIA tasks (Section 3)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:138px; top:1350px; width:341px; height:9px;"><span style="font-family: CMR10; font-size:9px">2. A rich repository of pre-trained neural network models (Model Zoo) that\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:151px; top:1362px; width:137px; height:9px;"><span style="font-family: CMR10; font-size:9px">underlies the off-the-shelf usage\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:138px; top:1373px; width:341px; height:9px;"><span style="font-family: CMR10; font-size:9px">3. Comprehensive tools for efficient document image data annotation and model\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:151px; top:1385px; width:218px; height:9px;"><span style="font-family: CMR10; font-size:9px">tuning to support different levels of customization\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:138px; top:1397px; width:343px; height:33px;"><span style="font-family: CMR10; font-size:9px">4. A DL model hub and community platform for the easy sharing, distribu-\n<br>tion, and discussion of DIA models and pipelines, to promote reusability,\n<br>reproducibility, and extensibility (Section 4)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:1439px; width:346px; height:69px;"><span style="font-family: CMR10; font-size:9px">The library implements simple and intuitive Python APIs without sacrificing\n<br>generalizability and versatility, and can be easily installed via pip. Its convenient\n<br>functions for handling document image data can be seamlessly integrated with\n<br>existing DIA pipelines. With detailed documentations and carefully curated\n<br>tutorials, we hope this tool will benefit a variety of end-users, and will lead to\n<br>advances in applications in both industry and academic research.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:1511px; width:347px; height:46px;"><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is well aligned with recent efforts for improving DL model\n<br>reusability in other disciplines like natural language processing [8, 34] and com-\n<br>puter vision [35], but with a focus on unique challenges in DIA. We show\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">can be applied in sophisticated and large-scale digitization projects\n<br></span></div><span style="position:absolute; border: gray 1px solid; left:0px; top:1734px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:1734px;"><a name="3">Page 3</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:237px; top:1826px; width:210px; height:9px;"><span style="font-family: CMTT9; font-size:8px">LayoutParser</span><span style="font-family: CMR9; font-size:8px">: A Unified Toolkit for DL-Based DIA\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:475px; top:1826px; width:4px; height:8px;"><span style="font-family: CMR9; font-size:8px">3\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:1851px; width:348px; height:57px;"><span style="font-family: CMR10; font-size:9px">that require precision, efficiency, and robustness, as well as simple and light-\n<br>weight document processing tasks focusing on efficacy and flexibility (Section 5).\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is being actively maintained, and support for more deep learning\n<br>models and novel methods in text-based layout analysis methods [37, 34] is\n<br>planned.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:1911px; width:347px; height:58px;"><span style="font-family: CMR10; font-size:9px">The rest of the paper is organized as follows. Section 2 provides an overview\n<br>of related work. The core </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">library, DL Model Zoo, and customized\n<br>model training are described in Section 3, and the DL model hub and commu-\n<br>nity platform are detailed in Section 4. Section 5 shows two examples of how\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">can be used in practical DIA projects, and Section 6 concludes.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:1990px; width:102px; height:11px;"><span style="font-family: CMBX12; font-size:11px">2 Related Work\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:2016px; width:347px; height:81px;"><span style="font-family: CMR10; font-size:9px">Recently, various DL models and datasets have been developed for layout analysis\n<br>tasks. The dhSegment [22] utilizes fully convolutional networks [20] for segmen-\n<br>tation tasks on historical documents. Object detection-based methods like Faster\n<br>R-CNN [28] and Mask R-CNN [12] are used for identifying document elements [38]\n<br>and detecting tables [30, 26]. Most recently, Graph Neural Networks [29] have also\n<br>been used in table detection [27]. However, these models are usually implemented\n<br>individually and there is no unified framework to load and use such models.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:2100px; width:348px; height:189px;"><span style="font-family: CMR10; font-size:9px">There has been a surge of interest in creating open-source tools for document\n<br>image processing: a search of </span><span style="font-family: CMTT10; font-size:9px">document image analysis </span><span style="font-family: CMR10; font-size:9px">in Github leads to 5M\n<br>relevant code pieces </span><span style="font-family: CMR7; font-size:6px">6</span><span style="font-family: CMR10; font-size:9px">; yet most of them rely on traditional rule-based methods\n<br>or provide limited functionalities. The closest prior research to our work is the\n<br>OCR-D project</span><span style="font-family: CMR7; font-size:6px">7</span><span style="font-family: CMR10; font-size:9px">, which also tries to build a complete toolkit for DIA. However,\n<br>similar to the platform developed by Neudecker et al. [21], it is designed for\n<br>analyzing historical documents, and provides no supports for recent DL models.\n<br>The </span><span style="font-family: CMTT10; font-size:9px">DocumentLayoutAnalysis </span><span style="font-family: CMR10; font-size:9px">project</span><span style="font-family: CMR7; font-size:6px">8 </span><span style="font-family: CMR10; font-size:9px">focuses on processing born-digital PDF\n<br>documents via analyzing the stored PDF data. Repositories like </span><span style="font-family: CMTT10; font-size:9px">DeepLayout</span><span style="font-family: CMR7; font-size:6px">9\n<br></span><span style="font-family: CMR10; font-size:9px">and </span><span style="font-family: CMTT10; font-size:9px">Detectron2-PubLayNet</span><span style="font-family: CMR7; font-size:6px">10 </span><span style="font-family: CMR10; font-size:9px">are individual deep learning models trained on\n<br>layout analysis datasets without support for the full DIA pipeline. The Document\n<br>Analysis and Exploitation (DAE) platform [15] and the DeepDIVA project [2]\n<br>aim to improve the reproducibility of DIA methods (or DL models), yet they\n<br>are not actively maintained. OCR engines like </span><span style="font-family: CMTT10; font-size:9px">Tesseract </span><span style="font-family: CMR10; font-size:9px">[14], </span><span style="font-family: CMTT10; font-size:9px">easyOCR</span><span style="font-family: CMR7; font-size:6px">11 </span><span style="font-family: CMR10; font-size:9px">and\n<br></span><span style="font-family: CMTT10; font-size:9px">paddleOCR</span><span style="font-family: CMR7; font-size:6px">12 </span><span style="font-family: CMR10; font-size:9px">usually do not come with comprehensive functionalities for other\n<br>DIA tasks like layout analysis.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:2291px; width:347px; height:21px;"><span style="font-family: CMR10; font-size:9px">Recent years have also seen numerous efforts to create libraries for promoting\n<br>reproducibility and reusability in the field of DL. Libraries like Dectectron2 [35],\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:133px; top:2322px; width:295px; height:76px;"><span style="font-family: CMR6; font-size:5px">6 </span><span style="font-family: CMR9; font-size:8px">The number shown is obtained by specifying the search type as ‘code’.\n<br></span><span style="font-family: CMR6; font-size:5px">7 </span><span style="font-family: CMR9; font-size:8px">https://ocr-d.de/en/about\n<br></span><span style="font-family: CMR6; font-size:5px">8 </span><span style="font-family: CMR9; font-size:8px">https://github.com/BobLd/DocumentLayoutAnalysis\n<br></span><span style="font-family: CMR6; font-size:5px">9 </span><span style="font-family: CMR9; font-size:8px">https://github.com/leonlulu/DeepLayout\n<br></span><span style="font-family: CMR6; font-size:5px">10 </span><span style="font-family: CMR9; font-size:8px">https://github.com/hpanwar08/detectron2\n<br></span><span style="font-family: CMR6; font-size:5px">11 </span><span style="font-family: CMR9; font-size:8px">https://github.com/JaidedAI/EasyOCR\n<br></span><span style="font-family: CMR6; font-size:5px">12 </span><span style="font-family: CMR9; font-size:8px">https://github.com/PaddlePaddle/PaddleOCR\n<br></span></div><span style="position:absolute; border: black 1px solid; left:134px; top:2320px; width:56px; height:0px;"></span>\n<span style="position:absolute; border: gray 1px solid; left:0px; top:2576px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:2576px;"><a name="4">Page 4</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:2668px; width:4px; height:8px;"><span style="font-family: CMR9; font-size:8px">4\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:167px; top:2668px; width:54px; height:8px;"><span style="font-family: CMR9; font-size:8px">Z. Shen et al.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:2838px; width:348px; height:105px;"><span style="font-family: CMR10; font-size:9px">Fig. 1: The overall architecture of </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">. For an input document image,\n<br></span><span style="font-family: CMR10; font-size:9px">the core </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">library provides a set of off-the-shelf tools for layout\n<br>detection, OCR, visualization, and storage, backed by a carefully designed layout\n<br>data structure. </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">also supports high level customization via efficient\n<br>layout annotation and model training functions. These improve model accuracy\n<br>on the target samples. The community platform enables the easy sharing of DIA\n<br>models and whole digitization pipelines to promote reusability and reproducibility.\n<br>A collection of detailed documentation, tutorials and exemplar projects make\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">easy to learn and use.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:2970px; width:346px; height:81px;"><span style="font-family: CMR10; font-size:9px">AllenNLP [8] and transformers [34] have provided the community with complete\n<br>DL-based support for developing and deploying models for general computer\n<br>vision and natural language processing problems. </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">, on the other\n<br>hand, specializes specifically in DIA tasks. </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is also equipped with a\n<br>community platform inspired by established model hubs such as </span><span style="font-family: CMTT10; font-size:9px">Torch Hub </span><span style="font-family: CMR10; font-size:9px">[23]\n<br>and </span><span style="font-family: CMTT10; font-size:9px">TensorFlow Hub </span><span style="font-family: CMR10; font-size:9px">[1]. It enables the sharing of pretrained models as well as\n<br>full document processing pipelines that are unique to DIA tasks.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:3054px; width:347px; height:81px;"><span style="font-family: CMR10; font-size:9px">There have been a variety of document data collections to facilitate the\n<br>development of DL models. Some examples include PRImA [3](magazine layouts),\n<br>PubLayNet [38](academic paper layouts), Table Bank [18](tables in academic\n<br>papers), Newspaper Navigator Dataset [16, 17](newspaper figure layouts) and\n<br></span><span style="font-family: CMTT10; font-size:9px">HJDataset </span><span style="font-family: CMR10; font-size:9px">[31](historical Japanese document layouts). A spectrum of models\n<br></span><span style="font-family: CMR10; font-size:9px">trained on these datasets are currently available in the </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">model zoo\n<br>to support different use cases.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:3156px; width:202px; height:12px;"><span style="font-family: CMBX12; font-size:11px">3 The Core </span><span style="font-family: CMTT12; font-size:11px">LayoutParser </span><span style="font-family: CMBX12; font-size:11px">Library\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:3183px; width:347px; height:57px;"><span style="font-family: CMR10; font-size:9px">At the core of </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is an off-the-shelf toolkit that streamlines DL-\n<br>based document image analysis. Five components support a simple interface\n<br>with comprehensive functionalities: 1) The </span><span style="font-family: CMTI10; font-size:9px">layout detection models </span><span style="font-family: CMR10; font-size:9px">enable using\n<br>pre-trained or self-trained DL models for layout detection with just four lines\n<br></span><span style="font-family: CMR10; font-size:9px">of code. 2) The detected layout information is stored in carefully engineered\n<br></span></div><div style="position:absolute; border: figure 1px solid; writing-mode:False; left:169px; top:2691px; width:276px; height:136px;"><span style="position:absolute; border: black 1px solid; left:169px; top:2691px; width:276px; height:136px;"></span>\n<span style="position:absolute; border: black 1px solid; left:179px; top:2701px; width:77px; height:66px;"></span>\n<span style="position:absolute; border: black 1px solid; left:185px; top:2719px; width:66px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:190px; top:2719px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:250px; top:2724px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:190px; top:2736px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:184px; top:2724px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:184px; top:2732px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:246px; top:2732px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:246px; top:2719px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:184px; top:2719px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:188px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:188px; top:2726px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:191px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:191px; top:2726px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:194px; top:2727px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:195px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:195px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:198px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:198px; top:2727px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:199px; top:2726px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:199px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:200px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:200px; top:2727px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:202px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:202px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:205px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:206px; top:2726px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:209px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:209px; top:2726px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:209px; top:2726px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:212px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:212px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:213px; top:2729px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:215px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:215px; top:2726px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:217px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:217px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:218px; top:2729px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:221px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:221px; top:2726px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:222px; top:2727px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:225px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:225px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:228px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:228px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:231px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:231px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:232px; top:2727px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:234px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:234px; top:2726px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:236px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:236px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:236px; top:2729px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:2726px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:240px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:241px; top:2727px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:241px; top:2726px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:242px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:242px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:243px; top:2727px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:245px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:245px; top:2727px; width:2px; height:3px;"></span>\n<span style="font-family: unknown; font-size:6px">Efficient Data Annotation<span style="position:absolute; border: black 1px solid; left:185px; top:2744px; width:66px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:190px; top:2744px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:250px; top:2749px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:190px; top:2761px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:184px; top:2749px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:184px; top:2757px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:246px; top:2757px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:246px; top:2744px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:184px; top:2744px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:187px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:187px; top:2751px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:190px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:190px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:193px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:193px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:195px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:196px; top:2751px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:197px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:197px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:198px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:200px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:200px; top:2752px; width:4px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:204px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:205px; top:2752px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:205px; top:2751px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:206px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:206px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:208px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:208px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:209px; top:2752px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:211px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:211px; top:2751px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:212px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:215px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:215px; top:2751px; width:4px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:220px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:220px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:220px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:222px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:223px; top:2751px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:223px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:225px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:226px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:226px; top:2752px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:228px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:228px; top:2751px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:231px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:231px; top:2751px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:233px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:233px; top:2752px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:235px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:235px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:236px; top:2754px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:238px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:238px; top:2752px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:2751px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:242px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:242px; top:2752px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:243px; top:2751px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:243px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:243px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:246px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:246px; top:2752px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:247px; top:2752px; width:1px; height:2px;"></span>\nCustomized Model Training<span style="position:absolute; border: black 1px solid; left:179px; top:2701px; width:77px; height:12px;"></span>\n<span style="position:absolute; border: black 1px solid; left:179px; top:2707px; width:77px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:188px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:188px; top:2705px; width:4px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:193px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:194px; top:2706px; width:3px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:195px; top:2707px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:197px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:197px; top:2705px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:198px; top:2707px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:201px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:201px; top:2706px; width:3px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:202px; top:2707px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:204px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:204px; top:2705px; width:0px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:207px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:207px; top:2705px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:211px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:212px; top:2706px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:215px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:215px; top:2706px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:218px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:218px; top:2706px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:220px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:220px; top:2706px; width:3px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:221px; top:2707px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:224px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:224px; top:2706px; width:4px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:229px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:229px; top:2706px; width:0px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:229px; top:2705px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:231px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:231px; top:2706px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:234px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:234px; top:2706px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:235px; top:2708px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:237px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:237px; top:2706px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:240px; top:2706px; width:0px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:240px; top:2705px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:241px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:241px; top:2706px; width:3px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:242px; top:2707px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:245px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:245px; top:2706px; width:2px; height:3px;"></span>\nModel Customization<span style="position:absolute; border: black 1px solid; left:358px; top:2701px; width:77px; height:66px;"></span>\n<span style="position:absolute; border: black 1px solid; left:363px; top:2719px; width:66px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:368px; top:2719px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:429px; top:2724px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:368px; top:2736px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:362px; top:2724px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:362px; top:2732px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:424px; top:2732px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:424px; top:2719px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:362px; top:2719px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:378px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:378px; top:2726px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:379px; top:2726px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:382px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:382px; top:2726px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:383px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:383px; top:2726px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:384px; top:2727px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:388px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:388px; top:2726px; width:4px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:393px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:393px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:394px; top:2727px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:396px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:396px; top:2726px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:397px; top:2727px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:399px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:399px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:400px; top:2727px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:402px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:402px; top:2726px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:405px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:405px; top:2726px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:408px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:409px; top:2727px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:411px; top:2730px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:412px; top:2726px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:412px; top:2727px; width:1px; height:2px;"></span>\nDIA Model Hub<span style="position:absolute; border: black 1px solid; left:363px; top:2744px; width:66px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:368px; top:2744px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:429px; top:2749px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:368px; top:2761px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:362px; top:2749px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:362px; top:2757px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:424px; top:2757px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:424px; top:2744px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:362px; top:2744px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:372px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:372px; top:2751px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:372px; top:2751px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:375px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:375px; top:2751px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:377px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:376px; top:2751px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:378px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:382px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:382px; top:2751px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:383px; top:2751px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:385px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:385px; top:2752px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:386px; top:2751px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:386px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:386px; top:2752px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:387px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:390px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:390px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:390px; top:2752px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:392px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:393px; top:2751px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:394px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:394px; top:2752px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:395px; top:2751px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:395px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:395px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:398px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:399px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:399px; top:2752px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:403px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:403px; top:2751px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:406px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:406px; top:2751px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:409px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:409px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:409px; top:2754px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:412px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:412px; top:2752px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:414px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:414px; top:2752px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:414px; top:2751px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:415px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:415px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:418px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:418px; top:2752px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:419px; top:2752px; width:1px; height:2px;"></span>\nDIA Pipeline Sharing<span style="position:absolute; border: black 1px solid; left:358px; top:2701px; width:77px; height:12px;"></span>\n<span style="position:absolute; border: black 1px solid; left:358px; top:2707px; width:77px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:367px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:367px; top:2705px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:371px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:371px; top:2706px; width:3px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:372px; top:2707px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:375px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:375px; top:2706px; width:4px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:380px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:381px; top:2706px; width:4px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:386px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:386px; top:2706px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:389px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:389px; top:2706px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:393px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:393px; top:2706px; width:0px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:393px; top:2705px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:394px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:394px; top:2706px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:396px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:396px; top:2706px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:401px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:402px; top:2705px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:402px; top:2706px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:405px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:406px; top:2705px; width:0px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:407px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:407px; top:2706px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:408px; top:2708px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:410px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:410px; top:2706px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:412px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:412px; top:2705px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:415px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:415px; top:2706px; width:3px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:416px; top:2707px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:418px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:419px; top:2706px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:420px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:421px; top:2706px; width:4px; height:3px;"></span>\nCommunity Platform<span style="position:absolute; border: black 1px solid; left:179px; top:2774px; width:256px; height:43px;"></span>\n<span style="position:absolute; border: black 1px solid; left:265px; top:2737px; width:83px; height:36px;"></span>\n<span style="position:absolute; border: black 1px solid; left:274px; top:2744px; width:66px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:279px; top:2744px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:340px; top:2749px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:279px; top:2761px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:273px; top:2749px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:273px; top:2757px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:335px; top:2757px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:335px; top:2744px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:273px; top:2744px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:277px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:277px; top:2751px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:280px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:280px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:281px; top:2754px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:282px; top:2752px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:285px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:285px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:286px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:288px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:289px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:291px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:292px; top:2751px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:295px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:295px; top:2751px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:295px; top:2751px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:298px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:298px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:299px; top:2752px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:301px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:301px; top:2751px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:303px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:303px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:304px; top:2752px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:306px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:306px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:308px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:309px; top:2751px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:310px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:310px; top:2752px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:311px; top:2751px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:312px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:312px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:313px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:315px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:315px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:319px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:319px; top:2751px; width:4px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:324px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:324px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:325px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:327px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:327px; top:2751px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:328px; top:2752px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:330px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:330px; top:2752px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:331px; top:2752px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:333px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:333px; top:2751px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:334px; top:2755px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:335px; top:2752px; width:2px; height:3px;"></span>\nLayout Detection Models<span style="position:absolute; border: black 1px solid; left:281px; top:2701px; width:52px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:280px; top:2701px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:333px; top:2701px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:333px; top:2719px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:280px; top:2719px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:281px; top:2701px; width:52px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:333px; top:2701px; width:1px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:281px; top:2719px; width:52px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:280px; top:2701px; width:1px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:285px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:285px; top:2708px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:286px; top:2709px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:289px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:289px; top:2709px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:290px; top:2710px; width:0px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:292px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:292px; top:2709px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:294px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:295px; top:2709px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:297px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:297px; top:2709px; width:4px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:302px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:302px; top:2709px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:303px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:305px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:305px; top:2709px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:308px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:308px; top:2708px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:311px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:311px; top:2708px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:313px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:313px; top:2709px; width:4px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:318px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:318px; top:2709px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:318px; top:2711px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:320px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:320px; top:2709px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:321px; top:2710px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:324px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:324px; top:2709px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:325px; top:2710px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:327px; top:2712px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:327px; top:2709px; width:2px; height:3px;"></span>\nDocument Images <span style="position:absolute; border: black 1px solid; left:179px; top:2805px; width:256px; height:12px;"></span>\n<span style="position:absolute; border: black 1px solid; left:179px; top:2805px; width:256px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:266px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:266px; top:2809px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:270px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:270px; top:2809px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:273px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:273px; top:2810px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:274px; top:2811px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:278px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:278px; top:2809px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:282px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:282px; top:2810px; width:3px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:2811px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:285px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:286px; top:2810px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:287px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:288px; top:2810px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:288px; top:2811px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:292px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:292px; top:2809px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:295px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:295px; top:2810px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:296px; top:2812px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:298px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:298px; top:2810px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:301px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:302px; top:2810px; width:3px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:302px; top:2811px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:305px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:305px; top:2810px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:308px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:308px; top:2809px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:310px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:311px; top:2809px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:311px; top:2810px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:314px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:314px; top:2810px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:315px; top:2812px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:317px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:318px; top:2810px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:319px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:320px; top:2810px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:322px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:323px; top:2810px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:324px; top:2811px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:326px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:326px; top:2810px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:329px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:330px; top:2809px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:332px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:333px; top:2810px; width:0px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:333px; top:2809px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:334px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:334px; top:2809px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:335px; top:2811px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:337px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:338px; top:2810px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:339px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:340px; top:2810px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:341px; top:2812px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:343px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:343px; top:2810px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:345px; top:2813px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:345px; top:2810px; width:3px; height:4px;"></span>\n</span><span style="font-family: unknown; font-size:6px">The Core LayoutParser Library<span style="position:absolute; border: black 1px solid; left:185px; top:2782px; width:66px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:190px; top:2781px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:250px; top:2787px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:190px; top:2799px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:184px; top:2787px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:184px; top:2794px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:246px; top:2794px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:246px; top:2781px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:184px; top:2781px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:203px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:203px; top:2788px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:204px; top:2789px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:207px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:207px; top:2788px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:210px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:210px; top:2788px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:211px; top:2789px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:215px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:215px; top:2788px; width:4px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:219px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:220px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:220px; top:2790px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:223px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:223px; top:2788px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:223px; top:2790px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:226px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:226px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:229px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:229px; top:2788px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:230px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:230px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:231px; top:2790px; width:1px; height:0px;"></span>\n</span><span style="font-family: unknown; font-size:6px">OCR Module<span style="position:absolute; border: black 1px solid; left:363px; top:2782px; width:66px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:368px; top:2781px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:429px; top:2787px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:368px; top:2799px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:363px; top:2787px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:363px; top:2794px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:425px; top:2794px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:425px; top:2781px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:363px; top:2781px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:368px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:368px; top:2788px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:371px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:372px; top:2789px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:373px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:373px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:374px; top:2790px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:376px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:376px; top:2789px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:378px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:378px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:379px; top:2791px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:381px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:381px; top:2789px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:382px; top:2790px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:384px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:384px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:385px; top:2790px; width:1px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:388px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:389px; top:2788px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:389px; top:2791px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:390px; top:2789px; width:0px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:393px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:394px; top:2788px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:397px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:397px; top:2789px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:397px; top:2788px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:398px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:398px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:401px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:401px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:404px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:404px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:405px; top:2791px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:407px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:407px; top:2788px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:408px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:408px; top:2789px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:409px; top:2788px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:410px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:410px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:412px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:412px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:413px; top:2791px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:415px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:416px; top:2789px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:417px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:417px; top:2789px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:418px; top:2788px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:419px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:419px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:419px; top:2790px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:422px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:422px; top:2789px; width:2px; height:3px;"></span>\nStorage &amp; Visualization<span style="position:absolute; border: black 1px solid; left:274px; top:2782px; width:66px; height:17px;"></span>\n<span style="position:absolute; border: black 1px solid; left:279px; top:2781px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:340px; top:2787px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:279px; top:2799px; width:56px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:273px; top:2787px; width:1px; height:7px;"></span>\n<span style="position:absolute; border: black 1px solid; left:273px; top:2794px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:335px; top:2794px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:335px; top:2781px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:273px; top:2781px; width:5px; height:5px;"></span>\n<span style="position:absolute; border: black 1px solid; left:281px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:281px; top:2788px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:284px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:284px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:284px; top:2791px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:287px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:286px; top:2789px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:289px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:289px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:290px; top:2790px; width:1px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:292px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:292px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:295px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:295px; top:2789px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:298px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:299px; top:2788px; width:3px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:299px; top:2789px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:302px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:302px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:303px; top:2791px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:305px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:305px; top:2789px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:307px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:307px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:307px; top:2791px; width:1px; height:1px;"></span>\n<span style="position:absolute; border: black 1px solid; left:311px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:311px; top:2788px; width:2px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:314px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:315px; top:2789px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:316px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:316px; top:2789px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:318px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:318px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:321px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:321px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:324px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:324px; top:2789px; width:1px; height:4px;"></span>\n<span style="position:absolute; border: black 1px solid; left:326px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:326px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:329px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:329px; top:2789px; width:1px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:330px; top:2793px; width:0px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:331px; top:2789px; width:2px; height:3px;"></span>\n<span style="position:absolute; border: black 1px solid; left:331px; top:2790px; width:1px; height:0px;"></span>\nLayout Data Structure<span style="position:absolute; border: black 1px solid; left:304px; top:2767px; width:7px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:304px; top:2723px; width:7px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:255px; top:2789px; width:12px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:265px; top:2788px; width:3px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:256px; top:2791px; width:12px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:255px; top:2791px; width:3px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:255px; top:2752px; width:12px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:265px; top:2750px; width:3px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:256px; top:2754px; width:12px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:255px; top:2754px; width:3px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:345px; top:2789px; width:12px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:355px; top:2788px; width:3px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:346px; top:2791px; width:12px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:345px; top:2791px; width:3px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:345px; top:2752px; width:12px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:355px; top:2750px; width:3px; height:2px;"></span>\n<span style="position:absolute; border: black 1px solid; left:346px; top:2754px; width:12px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:345px; top:2754px; width:3px; height:2px;"></span>\n</span></div><span style="position:absolute; border: gray 1px solid; left:0px; top:3418px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:3418px;"><a name="5">Page 5</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:237px; top:3510px; width:210px; height:9px;"><span style="font-family: CMTT9; font-size:8px">LayoutParser</span><span style="font-family: CMR9; font-size:8px">: A Unified Toolkit for DL-Based DIA\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:475px; top:3510px; width:4px; height:8px;"><span style="font-family: CMR9; font-size:8px">5\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:146px; top:3544px; width:321px; height:10px;"><span style="font-family: CMR10; font-size:9px">Table 1: Current layout detection models in the </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">model zoo\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:150px; top:3559px; width:26px; height:6px;"><span style="font-family: CMBX9; font-size:6px">Dataset\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:194px; top:3558px; width:118px; height:7px;"><span style="font-family: CMBX9; font-size:6px">Base Model</span><span style="font-family: CMR6; font-size:4px">1 </span><span style="font-family: CMBX9; font-size:6px">Large Model Notes\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:3573px; width:47px; height:46px;"><span style="font-family: CMR9; font-size:6px">PubLayNet [38]\n<br>PRImA [3]\n<br>Newspaper [17]\n<br>TableBank [18]\n<br>HJDataset [31]\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:205px; top:3573px; width:18px; height:46px;"><span style="font-family: CMR9; font-size:6px">F / M\n<br>M\n<br>F\n<br>F\n<br>F / M\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:261px; top:3573px; width:6px; height:46px;"><span style="font-family: CMR9; font-size:6px">M\n<br>-\n<br>-\n<br>F\n<br>-\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:293px; top:3573px; width:181px; height:46px;"><span style="font-family: CMR9; font-size:6px">Layouts of modern scientific documents\n<br>Layouts of scanned modern magazines and scientific reports\n<br>Layouts of scanned US newspapers from the 20th century\n<br>Table region on modern scientific and business document\n<br>Layouts of history Japanese documents\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:136px; top:3624px; width:342px; height:48px;"><span style="font-family: CMR6; font-size:4px">1 </span><span style="font-family: CMR9; font-size:6px">For each dataset, we train several models of different sizes for different needs (the trade-off between accuracy\n<br>vs. computational cost). For “base model” and “large model”, we refer to using the ResNet 50 or ResNet 101\n<br>backbones [13], respectively. One can train models of different architectures, like Faster R-CNN [28] (F) and Mask\n<br>R-CNN [12] (M). For example, an F in the Large Model column indicates it has a Faster R-CNN model trained\n<br>using the ResNet 101 backbone. The platform is maintained and a number of additions will be made to the model\n<br>zoo in coming months.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:3697px; width:346px; height:81px;"><span style="font-family: CMTI10; font-size:9px">layout data structures</span><span style="font-family: CMR10; font-size:9px">, which are optimized for efficiency and versatility. 3) When\n<br></span><span style="font-family: CMR10; font-size:9px">necessary, users can employ existing or customized OCR models via the unified\n<br>API provided in the </span><span style="font-family: CMTI10; font-size:9px">OCR module</span><span style="font-family: CMR10; font-size:9px">. 4) </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">comes with a set of utility\n<br>functions for the </span><span style="font-family: CMTI10; font-size:9px">visualization and storage </span><span style="font-family: CMR10; font-size:9px">of the layout data. 5) </span><span style="font-family: CMTT10; font-size:9px">LayoutParser\n<br></span><span style="font-family: CMR10; font-size:9px">is also highly customizable, via its integration with functions for </span><span style="font-family: CMTI10; font-size:9px">layout data\n<br>annotation and model training</span><span style="font-family: CMR10; font-size:9px">. We now provide detailed descriptions for each\n<br>component.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:3797px; width:152px; height:9px;"><span style="font-family: CMBX10; font-size:9px">3.1 Layout Detection Models\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:3816px; width:347px; height:105px;"><span style="font-family: CMR10; font-size:9px">In </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">, a layout model takes a document image as an input and\n<br>generates a list of rectangular boxes for the target content regions. Different\n<br>from traditional methods, it relies on deep convolutional neural networks rather\n<br>than manually curated rules to identify content regions. It is formulated as an\n<br>object detection problem and state-of-the-art models like Faster R-CNN [28] and\n<br>Mask R-CNN [12] are used. This yields prediction results of high accuracy and\n<br>makes it possible to build a concise, generalized interface for layout detection.\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">, built upon Detectron2 [35], provides a minimal API that can\n<br>perform layout detection with only four lines of code in Python:\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:126px; top:3929px; width:267px; height:30px;"><span style="font-family: CMR5; font-size:4px">1 </span><span style="font-family: CMTT9; font-size:8px">import layoutparser as lp\n<br></span><span style="font-family: CMR5; font-size:4px">2 </span><span style="font-family: CMTT9; font-size:8px">image = cv2 . imread ( &quot; image_file &quot; ) # load images\n<br></span><span style="font-family: CMR5; font-size:4px">3 </span><span style="font-family: CMTT9; font-size:8px">model = lp . De t e c tro n2 Lay outM odel (\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:157px; top:3961px; width:270px; height:8px;"><span style="font-family: CMTT9; font-size:8px">&quot; lp :// PubLayNet / f as t er _ r c nn _ R _ 50 _ F P N_ 3 x / config &quot; )\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:126px; top:3964px; width:166px; height:16px;"><span style="font-family: CMR5; font-size:4px">4\n<br>5 </span><span style="font-family: CMTT9; font-size:8px">layout = model . detect ( image )\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:3989px; width:347px; height:93px;"><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">provides a wealth of pre-trained model weights using various\n<br>datasets covering different languages, time periods, and document types. Due to\n<br>domain shift [7], the prediction performance can notably drop when models are ap-\n<br>plied to target samples that are significantly different from the training dataset. As\n<br>document structures and layouts vary greatly in different domains, it is important\n<br>to select models trained on a dataset similar to the test samples. A semantic syntax\n<br>is used for initializing the model weights in </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">, using both the dataset\n<br></span><span style="font-family: CMR10; font-size:9px">name and model name </span><span style="font-family: CMTT10; font-size:9px">lp://&lt;dataset-name&gt;/&lt;model-architecture-name&gt;</span><span style="font-family: CMR10; font-size:9px">.\n<br></span></div><span style="position:absolute; border: black 1px solid; left:137px; top:3556px; width:341px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:191px; top:3558px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:3558px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:290px; top:3558px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:3569px; width:341px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:191px; top:3572px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:3572px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:290px; top:3572px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:191px; top:3581px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:3581px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:290px; top:3581px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:191px; top:3591px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:3591px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:290px; top:3591px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:191px; top:3601px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:3601px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:290px; top:3601px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:191px; top:3611px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:239px; top:3611px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:290px; top:3611px; width:0px; height:9px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:3622px; width:341px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:134px; top:3928px; width:345px; height:10px;"></span>\n<span style="position:absolute; border: black 1px solid; left:134px; top:3939px; width:345px; height:10px;"></span>\n<span style="position:absolute; border: black 1px solid; left:134px; top:3950px; width:345px; height:10px;"></span>\n<span style="position:absolute; border: black 1px solid; left:134px; top:3961px; width:345px; height:10px;"></span>\n<span style="position:absolute; border: black 1px solid; left:134px; top:3972px; width:345px; height:10px;"></span>\n<span style="position:absolute; border: gray 1px solid; left:0px; top:4260px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:4260px;"><a name="6">Page 6</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:4352px; width:4px; height:8px;"><span style="font-family: CMR9; font-size:8px">6\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:167px; top:4352px; width:54px; height:8px;"><span style="font-family: CMR9; font-size:8px">Z. Shen et al.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:4567px; width:347px; height:69px;"><span style="font-family: CMR10; font-size:9px">Fig. 2: The relationship between the three types of layout data structures.\n<br></span><span style="font-family: CMTT10; font-size:9px">Coordinate </span><span style="font-family: CMR10; font-size:9px">supports three kinds of variation; </span><span style="font-family: CMTT10; font-size:9px">TextBlock </span><span style="font-family: CMR10; font-size:9px">consists of the co-\n<br>ordinate information and extra features like block text, types, and reading orders;\n<br>a </span><span style="font-family: CMTT10; font-size:9px">Layout </span><span style="font-family: CMR10; font-size:9px">object is a list of all possible layout elements, including other </span><span style="font-family: CMTT10; font-size:9px">Layout\n<br></span><span style="font-family: CMR10; font-size:9px">objects. They all support the same set of transformation and operation APIs for\n<br>maximum flexibility.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:4665px; width:347px; height:69px;"><span style="font-family: CMR10; font-size:9px">Shown in Table 1, </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">currently hosts 9 pre-trained models trained\n<br>on 5 different datasets. Description of the training dataset is provided alongside\n<br>with the trained models such that users can quickly identify the most suitable\n<br>models for their tasks. Additionally, when such a model is not readily available,\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">also supports training customized layout models and community\n<br>sharing of the models (detailed in Section 3.5).\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:4759px; width:144px; height:9px;"><span style="font-family: CMBX10; font-size:9px">3.2 Layout Data Structures\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:4783px; width:346px; height:141px;"><span style="font-family: CMR10; font-size:9px">A critical feature of </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is the implementation of a series of data\n<br></span><span style="font-family: CMR10; font-size:9px">structures and operations that can be used to efficiently process and manipulate\n<br>the layout elements. In document image analysis pipelines, various post-processing\n<br>on the layout analysis model outputs is usually required to obtain the final\n<br>outputs. Traditionally, this requires exporting DL model outputs and then loading\n<br>the results into other pipelines. All model outputs from </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">will be\n<br>stored in carefully engineered data types optimized for further processing, which\n<br>makes it possible to build an end-to-end document digitization pipeline within\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">. There are three key components in the data structure, namely\n<br>the </span><span style="font-family: CMTT10; font-size:9px">Coordinate </span><span style="font-family: CMR10; font-size:9px">system, the </span><span style="font-family: CMTT10; font-size:9px">TextBlock</span><span style="font-family: CMR10; font-size:9px">, and the </span><span style="font-family: CMTT10; font-size:9px">Layout</span><span style="font-family: CMR10; font-size:9px">. They provide different\n<br>levels of abstraction for the layout data, and a set of APIs are supported for\n<br></span><span style="font-family: CMR10; font-size:9px">transformations or operations on these classes.\n<br></span></div><div style="position:absolute; border: figure 1px solid; writing-mode:False; left:195px; top:4375px; width:224px; height:181px;"></div><span style="position:absolute; border: gray 1px solid; left:0px; top:5102px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:5102px;"><a name="7">Page 7</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:237px; top:5194px; width:210px; height:9px;"><span style="font-family: CMTT9; font-size:8px">LayoutParser</span><span style="font-family: CMR9; font-size:8px">: A Unified Toolkit for DL-Based DIA\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:475px; top:5194px; width:4px; height:8px;"><span style="font-family: CMR9; font-size:8px">7\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:5219px; width:347px; height:177px;"><span style="font-family: CMR10; font-size:9px">Coordinates are the cornerstones for storing layout information. Currently,\n<br>three types of </span><span style="font-family: CMTT10; font-size:9px">Coordinate </span><span style="font-family: CMR10; font-size:9px">data structures are provided in </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">, shown\n<br>in Figure 2. </span><span style="font-family: CMTT10; font-size:9px">Interval </span><span style="font-family: CMR10; font-size:9px">and </span><span style="font-family: CMTT10; font-size:9px">Rectangle </span><span style="font-family: CMR10; font-size:9px">are the most common data types and\n<br>support specifying 1D or 2D regions within a document. They are parameterized\n<br>with 2 and 4 parameters. A </span><span style="font-family: CMTT10; font-size:9px">Quadrilateral </span><span style="font-family: CMR10; font-size:9px">class is also implemented to support\n<br>a more generalized representation of rectangular regions when the document\n<br>is skewed or distorted, where the 4 corner points can be specified and a total\n<br>of 8 degrees of freedom are supported. A wide collection of transformations\n<br>like </span><span style="font-family: CMTT10; font-size:9px">shift</span><span style="font-family: CMR10; font-size:9px">, </span><span style="font-family: CMTT10; font-size:9px">pad</span><span style="font-family: CMR10; font-size:9px">, and </span><span style="font-family: CMTT10; font-size:9px">scale</span><span style="font-family: CMR10; font-size:9px">, and operations like </span><span style="font-family: CMTT10; font-size:9px">intersect</span><span style="font-family: CMR10; font-size:9px">, </span><span style="font-family: CMTT10; font-size:9px">union</span><span style="font-family: CMR10; font-size:9px">, and </span><span style="font-family: CMTT10; font-size:9px">is_in</span><span style="font-family: CMR10; font-size:9px">,\n<br>are supported for these classes. Notably, it is common to separate a segment\n<br>of the image and analyze it individually. </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">provides full support\n<br>for this scenario via image cropping operations </span><span style="font-family: CMTT10; font-size:9px">crop_image </span><span style="font-family: CMR10; font-size:9px">and coordinate\n<br>transformations like </span><span style="font-family: CMTT10; font-size:9px">relative_to </span><span style="font-family: CMR10; font-size:9px">and </span><span style="font-family: CMTT10; font-size:9px">condition_on </span><span style="font-family: CMR10; font-size:9px">that transform coordinates\n<br></span><span style="font-family: CMR10; font-size:9px">to and from their relative representations. We refer readers to Table 2 for a more\n<br></span><span style="font-family: CMR10; font-size:9px">detailed description of these operations</span><span style="font-family: CMR7; font-size:6px">13</span><span style="font-family: CMR10; font-size:9px">.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:5399px; width:346px; height:81px;"><span style="font-family: CMR10; font-size:9px">Based on </span><span style="font-family: CMTT10; font-size:9px">Coordinate</span><span style="font-family: CMR10; font-size:9px">s, we implement the </span><span style="font-family: CMTT10; font-size:9px">TextBlock </span><span style="font-family: CMR10; font-size:9px">class that stores both\n<br>the positional and extra features of individual layout elements. It also supports\n<br>specifying the reading orders via setting the </span><span style="font-family: CMTT10; font-size:9px">parent </span><span style="font-family: CMR10; font-size:9px">field to the index of the parent\n<br>object. A </span><span style="font-family: CMTT10; font-size:9px">Layout </span><span style="font-family: CMR10; font-size:9px">class is built that takes in a list of </span><span style="font-family: CMTT10; font-size:9px">TextBlock</span><span style="font-family: CMR10; font-size:9px">s and supports\n<br>processing the elements in batch. </span><span style="font-family: CMTT10; font-size:9px">Layout </span><span style="font-family: CMR10; font-size:9px">can also be nested to support hierarchical\n<br>layout structures. They support the same operations and transformations as the\n<br></span><span style="font-family: CMTT10; font-size:9px">Coordinate </span><span style="font-family: CMR10; font-size:9px">classes, minimizing both learning and deployment effort.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:5498px; width:51px; height:9px;"><span style="font-family: CMBX10; font-size:9px">3.3 OCR\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:5516px; width:346px; height:105px;"><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">provides a unified interface for existing OCR tools. Though there\n<br>are many OCR tools available, they are usually configured differently with distinct\n<br>APIs or protocols for using them. It can be inefficient to add new OCR tools into\n<br>an existing pipeline, and difficult to make direct comparisons among the available\n<br>tools to find the best option for a particular project. To this end, </span><span style="font-family: CMTT10; font-size:9px">LayoutParser\n<br></span><span style="font-family: CMR10; font-size:9px">builds a series of wrappers among existing OCR engines, and provides nearly\n<br>the same syntax for using them. It supports a plug-and-play style of using OCR\n<br>engines, making it effortless to switch, evaluate, and compare different OCR\n<br>modules:\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:126px; top:5629px; width:267px; height:30px;"><span style="font-family: CMR5; font-size:4px">1 </span><span style="font-family: CMTT9; font-size:8px">ocr_agent = lp . TesseractAgent ()\n<br></span><span style="font-family: CMR5; font-size:4px">2 </span><span style="font-family: CMTT9; font-size:8px"># Can be easily switched to other OCR software\n<br></span><span style="font-family: CMR5; font-size:4px">3 </span><span style="font-family: CMTT9; font-size:8px">tokens = ocr_agent . detect ( image )\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:5667px; width:347px; height:45px;"><span style="font-family: CMR10; font-size:9px">The OCR outputs will also be stored in the aforementioned layout data\n<br>structures and can be seamlessly incorporated into the digitization pipeline.\n<br>Currently </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">supports the Tesseract and Google Cloud Vision OCR\n<br>engines.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:5715px; width:346px; height:33px;"><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">also comes with a DL-based CNN-RNN OCR model [6] trained\n<br>with the Connectionist Temporal Classification (CTC) loss [10]. It can be used\n<br>like the other OCR modules, and can be easily trained on customized datasets.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:133px; top:5756px; width:271px; height:10px;"><span style="font-family: CMR6; font-size:5px">13 </span><span style="font-family: CMR9; font-size:8px">This is also available in the </span><span style="font-family: CMTT9; font-size:8px">LayoutParser </span><span style="font-family: CMR9; font-size:8px">documentation pages.\n<br></span></div><span style="position:absolute; border: black 1px solid; left:134px; top:5628px; width:345px; height:10px;"></span>\n<span style="position:absolute; border: black 1px solid; left:134px; top:5639px; width:345px; height:10px;"></span>\n<span style="position:absolute; border: black 1px solid; left:134px; top:5650px; width:345px; height:10px;"></span>\n<span style="position:absolute; border: black 1px solid; left:134px; top:5754px; width:56px; height:0px;"></span>\n<span style="position:absolute; border: gray 1px solid; left:0px; top:5944px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:5944px;"><a name="8">Page 8</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:6036px; width:4px; height:8px;"><span style="font-family: CMR9; font-size:8px">8\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:167px; top:6036px; width:54px; height:8px;"><span style="font-family: CMR9; font-size:8px">Z. Shen et al.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:6070px; width:347px; height:34px;"><span style="font-family: CMR10; font-size:9px">Table 2: All operations supported by the layout elements. The same APIs are\n<br>supported across different layout element classes including </span><span style="font-family: CMTT10; font-size:9px">Coordinate </span><span style="font-family: CMR10; font-size:9px">types,\n<br></span><span style="font-family: CMTT10; font-size:9px">TextBlock </span><span style="font-family: CMR10; font-size:9px">and </span><span style="font-family: CMTT10; font-size:9px">Layout</span><span style="font-family: CMR10; font-size:9px">.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6110px; width:65px; height:7px;"><span style="font-family: CMBX9; font-size:7px">Operation Name\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:287px; top:6110px; width:44px; height:7px;"><span style="font-family: CMBX9; font-size:7px">Description\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6125px; width:310px; height:7px;"><span style="font-family: CMTT9; font-size:7px">block.pad(top, bottom, right, left) </span><span style="font-family: CMR9; font-size:7px">Enlarge the current block according to the input\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6146px; width:75px; height:7px;"><span style="font-family: CMTT9; font-size:7px">block.scale(fx, fy)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6172px; width:75px; height:7px;"><span style="font-family: CMTT9; font-size:7px">block.shift(dx, dy)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:287px; top:6140px; width:129px; height:18px;"><span style="font-family: CMR9; font-size:7px">Scale the current block given the ratio\n<br>in x and y direction\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:287px; top:6167px; width:127px; height:18px;"><span style="font-family: CMR9; font-size:7px">Move the current block with the shift\n<br>distances in x and y direction\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6193px; width:77px; height:7px;"><span style="font-family: CMTT9; font-size:7px">block1.is in(block2)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:287px; top:6193px; width:116px; height:7px;"><span style="font-family: CMR9; font-size:7px">Whether block1 is inside of block2\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6214px; width:94px; height:7px;"><span style="font-family: CMTT9; font-size:7px">block1.intersect(block2)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6240px; width:79px; height:7px;"><span style="font-family: CMTT9; font-size:7px">block1.union(block2)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6266px; width:101px; height:7px;"><span style="font-family: CMTT9; font-size:7px">block1.relative to(block2)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6292px; width:105px; height:7px;"><span style="font-family: CMTT9; font-size:7px">block1.condition on(block2)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:287px; top:6208px; width:186px; height:18px;"><span style="font-family: CMR9; font-size:7px">Return the intersection region of block1 and block2.\n<br></span><span style="font-family: CMR9; font-size:7px">Coordinate type to be determined based on the inputs.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:287px; top:6234px; width:186px; height:18px;"><span style="font-family: CMR9; font-size:7px">Return the union region of block1 and block2.\n<br>Coordinate type to be determined based on the inputs.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:287px; top:6261px; width:154px; height:18px;"><span style="font-family: CMR9; font-size:7px">Convert the absolute coordinates of block1 to\n<br>relative coordinates to block2\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:287px; top:6287px; width:170px; height:18px;"><span style="font-family: CMR9; font-size:7px">Calculate the absolute coordinates of block1 given\n<br>the canvas block2’s absolute coordinates\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:141px; top:6313px; width:89px; height:7px;"><span style="font-family: CMTT9; font-size:7px">block.crop image(image)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:287px; top:6313px; width:158px; height:7px;"><span style="font-family: CMR9; font-size:7px">Obtain the image segments in the block region\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:6350px; width:152px; height:9px;"><span style="font-family: CMBX10; font-size:9px">3.4 Storage and visualization\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:6371px; width:346px; height:129px;"><span style="font-family: CMR10; font-size:9px">The end goal of DIA is to transform the image-based document data into a\n<br>structured database. </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">supports exporting layout data into different\n<br>formats like </span><span style="font-family: CMTT10; font-size:9px">JSON</span><span style="font-family: CMR10; font-size:9px">, </span><span style="font-family: CMTT10; font-size:9px">csv</span><span style="font-family: CMR10; font-size:9px">, and will add the support for the METS/ALTO XML\n<br>format </span><span style="font-family: CMR7; font-size:6px">14 </span><span style="font-family: CMR10; font-size:9px">. It can also load datasets from layout analysis-specific formats like\n<br>COCO [38] and the Page Format [25] for training layout models (Section 3.5).\n<br>Visualization of the layout detection results is critical for both presentation\n<br>and debugging. </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is built with an integrated API for displaying the\n<br>layout information along with the original document image. Shown in Figure 3, it\n<br>enables presenting layout data with rich meta information and features in different\n<br>modes. More detailed information can be found in the online </span><span style="font-family: CMTT10; font-size:9px">LayoutParser\n<br></span><span style="font-family: CMR10; font-size:9px">documentation page.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:6522px; width:166px; height:9px;"><span style="font-family: CMBX10; font-size:9px">3.5 Customized Model Training\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:6543px; width:347px; height:45px;"><span style="font-family: CMR10; font-size:9px">Besides the off-the-shelf library, </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is also highly customizable with\n<br>supports for highly unique and challenging document analysis tasks. Target\n<br>document images can be vastly different from the existing datasets for train-\n<br>ing layout models, which leads to low layout detection accuracy. Training data\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:133px; top:6598px; width:111px; height:10px;"><span style="font-family: CMR6; font-size:5px">14 </span><span style="font-family: CMR9; font-size:8px">https://altoxml.github.io\n<br></span></div><span style="position:absolute; border: black 1px solid; left:137px; top:6105px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6108px; width:0px; height:11px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6121px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6123px; width:0px; height:11px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6136px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6139px; width:0px; height:22px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6162px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6165px; width:0px; height:22px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6189px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:177px; top:6199px; width:2px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6191px; width:0px; height:11px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6204px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6206px; width:0px; height:22px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6230px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6233px; width:0px; height:22px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6256px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:200px; top:6272px; width:2px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6259px; width:0px; height:22px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6283px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:204px; top:6298px; width:2px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6285px; width:0px; height:22px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6309px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:181px; top:6319px; width:2px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:283px; top:6312px; width:0px; height:11px;"></span>\n<span style="position:absolute; border: black 1px solid; left:137px; top:6324px; width:340px; height:0px;"></span>\n<span style="position:absolute; border: black 1px solid; left:134px; top:6596px; width:56px; height:0px;"></span>\n<span style="position:absolute; border: gray 1px solid; left:0px; top:6786px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:6786px;"><a name="9">Page 9</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:237px; top:6878px; width:210px; height:9px;"><span style="font-family: CMTT9; font-size:8px">LayoutParser</span><span style="font-family: CMR9; font-size:8px">: A Unified Toolkit for DL-Based DIA\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:475px; top:6878px; width:4px; height:8px;"><span style="font-family: CMR9; font-size:8px">9\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:7103px; width:346px; height:69px;"><span style="font-family: CMR10; font-size:9px">Fig. 3: Layout detection and OCR results visualization generated by the\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">APIs. Mode I directly overlays the layout region bounding boxes\n<br>and categories over the original image. Mode II recreates the original document\n<br>via drawing the OCR’d texts at their corresponding positions on the image\n<br>canvas. In this figure, tokens in textual regions are filtered using the API and\n<br>then displayed.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:7212px; width:347px; height:33px;"><span style="font-family: CMR10; font-size:9px">can also be highly sensitive and not sharable publicly. To overcome these chal-\n<br>lenges, </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is built with rich features for efficient data annotation and\n<br>customized model training.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:7255px; width:347px; height:81px;"><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">incorporates a toolkit optimized for annotating document lay-\n<br>outs using object-level active learning [32]. With the help from a layout detection\n<br>model trained along with labeling, only the most important layout objects within\n<br>each image, rather than the whole image, are required for labeling. The rest of\n<br>the regions are automatically annotated with high confidence predictions from\n<br>the layout detection model. This allows a layout dataset to be created more\n<br></span><span style="font-family: CMR10; font-size:9px">efficiently with only around 60% of the labeling budget.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:7345px; width:347px; height:105px;"><span style="font-family: CMR10; font-size:9px">After the training dataset is curated, </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">supports different modes\n<br>for training the layout models. </span><span style="font-family: CMTI10; font-size:9px">Fine-tuning </span><span style="font-family: CMR10; font-size:9px">can be used for training models on a\n<br></span><span style="font-family: CMTI10; font-size:9px">small </span><span style="font-family: CMR10; font-size:9px">newly-labeled dataset by initializing the model with existing pre-trained\n<br>weights. </span><span style="font-family: CMTI10; font-size:9px">Training from scratch </span><span style="font-family: CMR10; font-size:9px">can be helpful when the source dataset and\n<br>target are significantly different and a large training set is available. However, as\n<br>suggested in Studer et al.’s work[33], loading pre-trained weights on large-scale\n<br>datasets like ImageNet [5], even from totally different domains, can still boost\n<br>model performance. Through the integrated API provided by </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">,\n<br></span><span style="font-family: CMR10; font-size:9px">users can easily compare model performances on the benchmark datasets.\n<br></span></div><div style="position:absolute; border: figure 1px solid; writing-mode:False; left:169px; top:6901px; width:276px; height:191px;"></div><span style="position:absolute; border: gray 1px solid; left:0px; top:7628px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:7628px;"><a name="10">Page 10</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:7720px; width:9px; height:8px;"><span style="font-family: CMR9; font-size:8px">10\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:167px; top:7720px; width:54px; height:8px;"><span style="font-family: CMR9; font-size:8px">Z. Shen et al.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:7912px; width:346px; height:57px;"><span style="font-family: CMR10; font-size:9px">Fig. 4: Illustration of (a) the original historical Japanese document with layout\n<br>detection results and (b) a recreated version of the document image that achieves\n<br>much better character recognition recall. The reorganization algorithm rearranges\n<br>the tokens based on the their detected bounding boxes given a maximum allowed\n<br>height.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:7992px; width:224px; height:12px;"><span style="font-family: CMBX12; font-size:11px">4 </span><span style="font-family: CMTT12; font-size:11px">LayoutParser </span><span style="font-family: CMBX12; font-size:11px">Community Platform\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:8016px; width:347px; height:81px;"><span style="font-family: CMR10; font-size:9px">Another focus of </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is promoting the reusability of layout detection\n<br>models and full digitization pipelines. Similar to many existing deep learning\n<br>libraries, </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">comes with a community model hub for distributing\n<br>layout models. End-users can upload their self-trained models to the model hub,\n<br>and these models can be loaded into a similar interface as the currently available\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">pre-trained models. For example, the model trained on the News\n<br>Navigator dataset [17] has been incorporated in the model hub.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:8099px; width:347px; height:129px;"><span style="font-family: CMR10; font-size:9px">Beyond DL models, </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">also promotes the sharing of entire doc-\n<br>ument digitization pipelines. For example, sometimes the pipeline requires the\n<br>combination of multiple DL models to achieve better accuracy. Currently, pipelines\n<br>are mainly described in academic papers and implementations are often not pub-\n<br>licly available. To this end, the </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">community platform also enables\n<br></span><span style="font-family: CMR10; font-size:9px">the sharing of layout pipelines to promote the discussion and reuse of techniques.\n<br>For each shared pipeline, it has a dedicated project page, with links to the source\n<br>code, documentation, and an outline of the approaches. A discussion panel is\n<br>provided for exchanging ideas. Combined with the core </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">library,\n<br>users can easily build reusable components based on the shared pipelines and\n<br>apply them to solve their unique problems.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:8247px; width:79px; height:11px;"><span style="font-family: CMBX12; font-size:11px">5 Use Cases\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:8270px; width:346px; height:21px;"><span style="font-family: CMR10; font-size:9px">The core objective of </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is to make it easier to create both large-scale\n<br></span><span style="font-family: CMR10; font-size:9px">and light-weight document digitization pipelines. Large-scale document processing\n<br></span></div><div style="position:absolute; border: figure 1px solid; writing-mode:False; left:134px; top:7743px; width:345px; height:158px;"></div><span style="position:absolute; border: gray 1px solid; left:0px; top:8470px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:8470px;"><a name="11">Page 11</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:237px; top:8562px; width:210px; height:9px;"><span style="font-family: CMTT9; font-size:8px">LayoutParser</span><span style="font-family: CMR9; font-size:8px">: A Unified Toolkit for DL-Based DIA\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:471px; top:8562px; width:9px; height:8px;"><span style="font-family: CMR9; font-size:8px">11\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:8587px; width:347px; height:117px;"><span style="font-family: CMR10; font-size:9px">focuses on precision, efficiency, and robustness. The target documents may have\n<br>complicated structures, and may require training multiple layout detection models\n<br>to achieve the optimal accuracy. Light-weight pipelines are built for relatively\n<br>simple documents, with an emphasis on development ease, speed and flexibility.\n<br>Ideally one only needs to use existing resources, and model training should be\n<br>avoided. Through two exemplar projects, we show how practitioners in both\n<br>academia and industry can easily build such pipelines using </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">and\n<br>extract high-quality structured document data for their downstream tasks. The\n<br>source code for these projects will be publicly available in the </span><span style="font-family: CMTT10; font-size:9px">LayoutParser\n<br></span><span style="font-family: CMR10; font-size:9px">community hub.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:8730px; width:330px; height:9px;"><span style="font-family: CMBX10; font-size:9px">5.1 A Comprehensive Historical Document Digitization Pipeline\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:8755px; width:347px; height:189px;"><span style="font-family: CMR10; font-size:9px">The digitization of historical documents can unlock valuable data that can shed\n<br>light on many important social, economic, and historical questions. Yet due to\n<br>scan noises, page wearing, and the prevalence of complicated layout structures, ob-\n<br>taining a structured representation of historical document scans is often extremely\n<br>complicated.\n<br>In this example, </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">was\n<br>used to develop a comprehensive\n<br>pipeline, shown in Figure 5, to gener-\n<br>ate high-quality structured data from\n<br>historical Japanese firm financial ta-\n<br>bles with complicated layouts. The\n<br>pipeline applies two layout models to\n<br>identify different levels of document\n<br>structures and two customized OCR\n<br>engines for optimized character recog-\n<br>nition accuracy.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:8947px; width:165px; height:153px;"><span style="font-family: CMR10; font-size:9px">As shown in Figure 4 (a), the\n<br>document contains columns of text\n<br>written vertically </span><span style="font-family: CMR7; font-size:6px">15</span><span style="font-family: CMR10; font-size:9px">, a common style\n<br>in Japanese. Due to scanning noise\n<br>and archaic printing technology, the\n<br></span><span style="font-family: CMR10; font-size:9px">columns can be skewed or have vari-\n<br>able widths, and hence cannot be eas-\n<br>ily identified via rule-based methods.\n<br>Within each column, words are sepa-\n<br>rated by white spaces of variable size,\n<br>and the vertical positions of objects\n<br>can be an indicator of their layout\n<br>type.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:307px; top:9052px; width:174px; height:33px;"><span style="font-family: CMR10; font-size:9px">Fig. 5: Illustration of how </span><span style="font-family: CMTT10; font-size:9px">LayoutParser\n<br></span><span style="font-family: CMR10; font-size:9px">helps with the historical document digi-\n<br>tization pipeline.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:133px; top:9113px; width:347px; height:10px;"><span style="font-family: CMR6; font-size:5px">15 </span><span style="font-family: CMR9; font-size:8px">A document page consists of eight rows like this. For simplicity we skip the row\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:144px; top:9125px; width:308px; height:8px;"><span style="font-family: CMR9; font-size:8px">segmentation discussion and refer readers to the source code when available.\n<br></span></div><div style="position:absolute; border: figure 1px solid; writing-mode:False; left:307px; top:8849px; width:172px; height:182px;"></div><span style="position:absolute; border: black 1px solid; left:134px; top:9111px; width:56px; height:0px;"></span>\n<span style="position:absolute; border: gray 1px solid; left:0px; top:9312px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:9312px;"><a name="12">Page 12</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:9404px; width:9px; height:8px;"><span style="font-family: CMR9; font-size:8px">12\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:167px; top:9404px; width:54px; height:8px;"><span style="font-family: CMR9; font-size:8px">Z. Shen et al.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:149px; top:9429px; width:148px; height:9px;"><span style="font-family: CMR10; font-size:9px">To decipher the complicated layout\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:9441px; width:347px; height:141px;"><span style="font-family: CMR10; font-size:9px">structure, two object detection models have been trained to recognize individual\n<br>columns and tokens, respectively. A small training set (400 images with approxi-\n<br>mately 100 annotations each) is curated via the active learning based annotation\n<br>tool [32] in </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">. The models learn to identify both the categories and\n<br>regions for each token or column via their distinct visual features. The layout\n<br>data structure enables easy grouping of the tokens within each column, and\n<br>rearranging columns to achieve the correct reading orders based on the horizontal\n<br>position. Errors are identified and rectified via checking the consistency of the\n<br>model predictions. Therefore, though trained on a small dataset, the pipeline\n<br>achieves a high level of layout detection accuracy: it achieves a 96.97 AP [19]\n<br>score across 5 categories for the column detection model, and a 89.23 AP across\n<br>4 categories for the token detection model.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:9587px; width:346px; height:117px;"><span style="font-family: CMR10; font-size:9px">A combination of character recognition methods is developed to tackle the\n<br>unique challenges in this document. In our experiments, we found that irregular\n<br>spacing between the tokens led to a low character recognition recall rate, whereas\n<br>existing OCR models tend to perform better on densely-arranged texts. To\n<br>overcome this challenge, we create a document reorganization algorithm that\n<br>rearranges the text based on the token bounding boxes detected in the layout\n<br>analysis step. Figure 4 (b) illustrates the generated image of dense text, which is\n<br>sent to the OCR APIs as a whole to reduce the transaction costs. The flexible\n<br>coordinate system in </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">is used to transform the OCR results relative\n<br>to their original positions on the page.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:9710px; width:347px; height:141px;"><span style="font-family: CMR10; font-size:9px">Additionally, it is common for historical documents to use unique fonts\n<br>with different glyphs, which significantly degrades the accuracy of OCR models\n<br>trained on modern texts. In this document, a special flat font is used for printing\n<br>numbers and could not be detected by off-the-shelf OCR engines. Using the highly\n<br>flexible functionalities from </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">, a pipeline approach is constructed\n<br>that achieves a high recognition accuracy with minimal effort. As the characters\n<br>have unique visual structures and are usually clustered together, we train the\n<br>layout model to identify number regions with a dedicated category. Subsequently,\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">crops images within these regions, and identifies characters within\n<br>them using a self-trained OCR model based on a CNN-RNN [6]. The model\n<br>detects a total of 15 possible categories, and achieves a 0.98 Jaccard score</span><span style="font-family: CMR7; font-size:6px">16 </span><span style="font-family: CMR10; font-size:9px">and\n<br></span><span style="font-family: CMR10; font-size:9px">a 0.17 average Levinstein distances</span><span style="font-family: CMR7; font-size:6px">17 </span><span style="font-family: CMR10; font-size:9px">for token prediction on the test set.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:9856px; width:345px; height:57px;"><span style="font-family: CMR10; font-size:9px">Overall, it is possible to create an intricate and highly accurate digitization\n<br>pipeline for large-scale digitization using </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">. The pipeline avoids\n<br>specifying the complicated rules used in traditional methods, is straightforward\n<br>to develop, and is robust to outliers. The DL models also generate fine-grained\n<br>results that enable creative approaches like page reorganization for OCR.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:133px; top:9933px; width:346px; height:10px;"><span style="font-family: CMR6; font-size:5px">16 </span><span style="font-family: CMR9; font-size:8px">This measures the overlap between the detected and ground-truth characters, and\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:144px; top:9945px; width:75px; height:8px;"><span style="font-family: CMR9; font-size:8px">the maximum is 1.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:133px; top:9955px; width:348px; height:10px;"><span style="font-family: CMR6; font-size:5px">17 </span><span style="font-family: CMR9; font-size:8px">This measures the number of edits from the ground-truth text to the predicted text,\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:144px; top:9967px; width:78px; height:8px;"><span style="font-family: CMR9; font-size:8px">and lower is better.\n<br></span></div><span style="position:absolute; border: black 1px solid; left:134px; top:9931px; width:56px; height:0px;"></span>\n<span style="position:absolute; border: gray 1px solid; left:0px; top:10154px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:10154px;"><a name="13">Page 13</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:237px; top:10246px; width:210px; height:9px;"><span style="font-family: CMTT9; font-size:8px">LayoutParser</span><span style="font-family: CMR9; font-size:8px">: A Unified Toolkit for DL-Based DIA\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:471px; top:10246px; width:9px; height:8px;"><span style="font-family: CMR9; font-size:8px">13\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:10398px; width:346px; height:45px;"><span style="font-family: CMR10; font-size:9px">Fig. 6: This lightweight table detector can identify tables (outlined in red) and\n<br>cells (shaded in blue) in different locations on a page. In very few cases (d), it\n<br>might generate minor error predictions, e.g, failing to capture the top text line of\n<br></span><span style="font-family: CMR10; font-size:9px">a table.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:10478px; width:216px; height:9px;"><span style="font-family: CMBX10; font-size:9px">5.2 A light-weight Visual Table Extractor\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:10518px; width:347px; height:81px;"><span style="font-family: CMR10; font-size:9px">Detecting tables and parsing their structures (table extraction) are of central im-\n<br>portance for many document digitization tasks. Many previous works [26, 30, 27]\n<br>and tools </span><span style="font-family: CMR7; font-size:6px">18 </span><span style="font-family: CMR10; font-size:9px">have been developed to identify and parse table structures. Yet they\n<br>might require training complicated models from scratch, or are only applicable\n<br>for born-digital PDF documents. In this section, we show how </span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">can\n<br>help build a light-weight accurate visual table extractor for legal docket tables\n<br>using the existing resources with minimal effort.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:10606px; width:347px; height:177px;"><span style="font-family: CMR10; font-size:9px">The extractor uses a pre-trained layout detection model for identifying the\n<br>table regions and some simple rules for pairing the rows and the columns in the\n<br>PDF image. Mask R-CNN [12] trained on the PubLayNet dataset [38] from the\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">Model Zoo can be used for detecting table regions. By filtering\n<br>out model predictions of low confidence and removing overlapping predictions,\n<br></span><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">can identify the tabular regions on each page, which significantly\n<br>simplifies the subsequent steps. By applying the line detection functions within\n<br></span><span style="font-family: CMR10; font-size:9px">the tabular segments, provided in the utility module from </span><span style="font-family: CMTT10; font-size:9px">LayoutParser</span><span style="font-family: CMR10; font-size:9px">, the\n<br>pipeline can identify the three distinct columns in the tables. A row clustering\n<br>method is then applied via analyzing the y coordinates of token bounding boxes in\n<br>the left-most column, which are obtained from the OCR engines. A non-maximal\n<br>suppression algorithm is used to remove duplicated rows with extremely small\n<br>gaps. Shown in Figure 6, the built pipeline can detect tables at different positions\n<br>on a page accurately. Continued tables from different pages are concatenated,\n<br>and a structured table representation has been easily created.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:133px; top:10808px; width:315px; height:10px;"><span style="font-family: CMR6; font-size:5px">18 </span><span style="font-family: CMR9; font-size:8px">https://github.com/atlanhq/camelot, https://github.com/tabulapdf/tabula\n<br></span></div><div style="position:absolute; border: figure 1px solid; writing-mode:False; left:134px; top:10269px; width:345px; height:118px;"></div><span style="position:absolute; border: black 1px solid; left:134px; top:10806px; width:56px; height:0px;"></span>\n<span style="position:absolute; border: gray 1px solid; left:0px; top:10996px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:10996px;"><a name="14">Page 14</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:11088px; width:9px; height:8px;"><span style="font-family: CMR9; font-size:8px">14\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:167px; top:11088px; width:54px; height:8px;"><span style="font-family: CMR9; font-size:8px">Z. Shen et al.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:11112px; width:84px; height:11px;"><span style="font-family: CMBX12; font-size:11px">6 Conclusion\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:11139px; width:346px; height:117px;"><span style="font-family: CMTT10; font-size:9px">LayoutParser </span><span style="font-family: CMR10; font-size:9px">provides a comprehensive toolkit for deep learning-based document\n<br>image analysis. The off-the-shelf library is easy to install, and can be used to\n<br>build flexible and accurate pipelines for processing documents with complicated\n<br>structures. It also supports high-level customization and enables easy labeling and\n<br>training of DL models on unique document image datasets. The </span><span style="font-family: CMTT10; font-size:9px">LayoutParser\n<br></span><span style="font-family: CMR10; font-size:9px">community platform facilitates sharing DL models and DIA pipelines, inviting\n<br>discussion and promoting code reproducibility and reusability. The </span><span style="font-family: CMTT10; font-size:9px">LayoutParser\n<br></span><span style="font-family: CMR10; font-size:9px">team is committed to keeping the library updated continuously and bringing\n<br>the most recent advances in DL-based DIA, such as multi-modal document\n<br>modeling [37, 36, 9] (an upcoming priority), to a diverse audience of end-users.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:11279px; width:347px; height:45px;"><span style="font-family: CMBX10; font-size:9px">Acknowledgements </span><span style="font-family: CMR10; font-size:9px">We thank the anonymous reviewers for their comments\n<br>and suggestions. This project is supported in part by NSF Grant OIA-2033558\n<br>and funding from the Harvard Data Science Initiative and Harvard Catalyst.\n<br>Zejiang Shen thanks Doug Downey for suggestions.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:11347px; width:62px; height:11px;"><span style="font-family: CMBX12; font-size:11px">References\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:11375px; width:341px; height:85px;"><span style="font-family: CMR9; font-size:8px">[1] Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado,\n<br>G.S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A.,\n<br>Irving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg,\n<br>J., Man´e, D., Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J.,\n<br>Steiner, B., Sutskever, I., Talwar, K., Tucker, P., Vanhoucke, V., Vasudevan, V.,\n<br>Vi´egas, F., Vinyals, O., Warden, P., Wattenberg, M., Wicke, M., Yu, Y., Zheng,\n<br>X.: TensorFlow: Large-scale machine learning on heterogeneous systems (2015),\n<br></span><span style="font-family: CMTT9; font-size:8px">https://www.tensorflow.org/</span><span style="font-family: CMR9; font-size:8px">, software available from tensorflow.org\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:11463px; width:341px; height:41px;"><span style="font-family: CMR9; font-size:8px">[2] Alberti, M., Pondenkandath, V., W¨ursch, M., Ingold, R., Liwicki, M.: Deepdiva: a\n<br>highly-functional python framework for reproducible experiments. In: 2018 16th\n<br>International Conference on Frontiers in Handwriting Recognition (ICFHR). pp.\n<br>423–428. IEEE (2018)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:11507px; width:341px; height:41px;"><span style="font-family: CMR9; font-size:8px">[3] Antonacopoulos, A., Bridson, D., Papadopoulos, C., Pletschacher, S.: A realistic\n<br>dataset for performance evaluation of document layout analysis. In: 2009 10th\n<br></span><span style="font-family: CMR9; font-size:8px">International Conference on Document Analysis and Recognition. pp. 296–300.\n<br></span><span style="font-family: CMR9; font-size:8px">IEEE (2009)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:11551px; width:340px; height:30px;"><span style="font-family: CMR9; font-size:8px">[4] Baek, Y., Lee, B., Han, D., Yun, S., Lee, H.: Character region awareness for text\n<br>detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and\n<br>Pattern Recognition. pp. 9365–9374 (2019)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:11585px; width:340px; height:8px;"><span style="font-family: CMR9; font-size:8px">[5] Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-Scale\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:155px; top:11596px; width:200px; height:8px;"><span style="font-family: CMR9; font-size:8px">Hierarchical Image Database. In: CVPR09 (2009)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:11607px; width:341px; height:30px;"><span style="font-family: CMR9; font-size:8px">[6] Deng, Y., Kanervisto, A., Ling, J., Rush, A.M.: Image-to-markup generation with\n<br>coarse-to-fine attention. In: International Conference on Machine Learning. pp.\n<br>980–989. PMLR (2017)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:11640px; width:341px; height:19px;"><span style="font-family: CMR9; font-size:8px">[7] Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation.\n<br>In: International conference on machine learning. pp. 1180–1189. PMLR (2015)\n<br></span></div><span style="position:absolute; border: gray 1px solid; left:0px; top:11838px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:11838px;"><a name="15">Page 15</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:237px; top:11930px; width:210px; height:9px;"><span style="font-family: CMTT9; font-size:8px">LayoutParser</span><span style="font-family: CMR9; font-size:8px">: A Unified Toolkit for DL-Based DIA\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:471px; top:11930px; width:9px; height:8px;"><span style="font-family: CMR9; font-size:8px">15\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:11956px; width:341px; height:63px;"><span style="font-family: CMR9; font-size:8px">[8] Gardner, M., Grus, J., Neumann, M., Tafjord, O., Dasigi, P., Liu, N., Peters,\n<br>M., Schmitz, M., Zettlemoyer, L.: Allennlp: A deep semantic natural language\n<br>processing platform. arXiv preprint arXiv:1803.07640 (2018)\n<br>(cid:32)Lukasz Garncarek, Powalski, R., Stanis(cid:32)lawek, T., Topolski, B., Halama, P.,\n<br>Grali´nski, F.: Lambert: Layout-aware (language) modeling using bert for in-\n<br>formation extraction (2020)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:140px; top:11989px; width:9px; height:8px;"><span style="font-family: CMR9; font-size:8px">[9]\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12022px; width:346px; height:41px;"><span style="font-family: CMR9; font-size:8px">[10] Graves, A., Fern´andez, S., Gomez, F., Schmidhuber, J.: Connectionist temporal\n<br>classification: labelling unsegmented sequence data with recurrent neural networks.\n<br>In: Proceedings of the 23rd international conference on Machine learning. pp.\n<br>369–376 (2006)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12066px; width:344px; height:41px;"><span style="font-family: CMR9; font-size:8px">[11] Harley, A.W., Ufkes, A., Derpanis, K.G.: Evaluation of deep convolutional nets for\n<br>document image classification and retrieval. In: 2015 13th International Conference\n<br>on Document Analysis and Recognition (ICDAR). pp. 991–995. IEEE (2015)\n<br>[12] He, K., Gkioxari, G., Doll´ar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:155px; top:12110px; width:293px; height:8px;"><span style="font-family: CMR9; font-size:8px">IEEE international conference on computer vision. pp. 2961–2969 (2017)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12121px; width:346px; height:30px;"><span style="font-family: CMR9; font-size:8px">[13] He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.\n<br>In: Proceedings of the IEEE conference on computer vision and pattern recognition.\n<br>pp. 770–778 (2016)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12153px; width:346px; height:8px;"><span style="font-family: CMR9; font-size:8px">[14] Kay, A.: Tesseract: An open-source optical character recognition engine. Linux J.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:154px; top:12164px; width:99px; height:8px;"><span style="font-family: CMBX9; font-size:8px">2007</span><span style="font-family: CMR9; font-size:8px">(159), 2 (Jul 2007)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12175px; width:344px; height:30px;"><span style="font-family: CMR9; font-size:8px">[15] Lamiroy, B., Lopresti, D.: An open architecture for end-to-end document analysis\n<br>benchmarking. In: 2011 International Conference on Document Analysis and\n<br>Recognition. pp. 42–47. IEEE (2011)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12208px; width:347px; height:64px;"><span style="font-family: CMR9; font-size:8px">[16] Lee, B.C., Weld, D.S.: Newspaper navigator: Open faceted search for 1.5\n<br>million images. In: Adjunct Publication of the 33rd Annual ACM Sym-\n<br>posium on User\n<br>Interface Software and Technology. p. 120–122. UIST\n<br>’20 Adjunct, Association for Computing Machinery, New York, NY, USA\n<br>(2020). https://doi.org/10.1145/3379350.3416143, </span><span style="font-family: CMTT9; font-size:8px">https://doi-org.offcampus.\n<br>lib.washington.edu/10.1145/3379350.3416143\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12274px; width:345px; height:53px;"><span style="font-family: CMR9; font-size:8px">[17] Lee, B.C.G., Mears, J., Jakeway, E., Ferriter, M., Adams, C., Yarasavage, N.,\n<br>Thomas, D., Zwaard, K., Weld, D.S.: The Newspaper Navigator Dataset: Extracting\n<br>Headlines and Visual Content from 16 Million Historic Newspaper Pages in\n<br>Chronicling America, p. 3055–3062. Association for Computing Machinery, New\n<br>York, NY, USA (2020), </span><span style="font-family: CMTT9; font-size:8px">https://doi.org/10.1145/3340531.3412767\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12329px; width:344px; height:30px;"><span style="font-family: CMR9; font-size:8px">[18] Li, M., Cui, L., Huang, S., Wei, F., Zhou, M., Li, Z.: Tablebank: Table benchmark\n<br>for image-based table detection and recognition. arXiv preprint arXiv:1903.01949\n<br>(2019)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12362px; width:345px; height:30px;"><span style="font-family: CMR9; font-size:8px">[19] Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll´ar, P.,\n<br></span><span style="font-family: CMR9; font-size:8px">Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference\n<br></span><span style="font-family: CMR9; font-size:8px">on computer vision. pp. 740–755. Springer (2014)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12395px; width:344px; height:30px;"><span style="font-family: CMR9; font-size:8px">[20] Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic\n<br>segmentation. In: Proceedings of the IEEE conference on computer vision and\n<br>pattern recognition. pp. 3431–3440 (2015)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12427px; width:346px; height:41px;"><span style="font-family: CMR9; font-size:8px">[21] Neudecker, C., Schlarb, S., Dogan, Z.M., Missier, P., Sufi, S., Williams, A., Wolsten-\n<br>croft, K.: An experimental workflow development platform for historical document\n<br>digitisation and analysis. In: Proceedings of the 2011 workshop on historical\n<br>document imaging and processing. pp. 161–168 (2011)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12471px; width:344px; height:30px;"><span style="font-family: CMR9; font-size:8px">[22] Oliveira, S.A., Seguin, B., Kaplan, F.: dhsegment: A generic deep-learning approach\n<br>for document segmentation. In: 2018 16th International Conference on Frontiers\n<br>in Handwriting Recognition (ICFHR). pp. 7–12. IEEE (2018)\n<br></span></div><span style="position:absolute; border: gray 1px solid; left:0px; top:12680px; width:612px; height:792px;"></span>\n<div style="position:absolute; top:12680px;"><a name="16">Page 16</a></div>\n<div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:134px; top:12772px; width:9px; height:8px;"><span style="font-family: CMR9; font-size:8px">16\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:167px; top:12772px; width:54px; height:8px;"><span style="font-family: CMR9; font-size:8px">Z. Shen et al.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12798px; width:345px; height:85px;"><span style="font-family: CMR9; font-size:8px">[23] Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z.,\n<br>Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)\n<br>[24] Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen,\n<br>T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style,\n<br>high-performance deep learning library. arXiv preprint arXiv:1912.01703 (2019)\n<br>[25] Pletschacher, S., Antonacopoulos, A.: The page (page analysis and ground-truth\n<br>elements) format framework. In: 2010 20th International Conference on Pattern\n<br>Recognition. pp. 257–260. IEEE (2010)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12886px; width:346px; height:41px;"><span style="font-family: CMR9; font-size:8px">[26] Prasad, D., Gadpal, A., Kapadni, K., Visave, M., Sultanpure, K.: Cascadetabnet:\n<br>An approach for end to end table detection and structure recognition from image-\n<br>based documents. In: Proceedings of the IEEE/CVF Conference on Computer\n<br>Vision and Pattern Recognition Workshops. pp. 572–573 (2020)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12930px; width:344px; height:30px;"><span style="font-family: CMR9; font-size:8px">[27] Qasim, S.R., Mahmood, H., Shafait, F.: Rethinking table recognition using graph\n<br>neural networks. In: 2019 International Conference on Document Analysis and\n<br>Recognition (ICDAR). pp. 142–147. IEEE (2019)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12962px; width:344px; height:30px;"><span style="font-family: CMR9; font-size:8px">[28] Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object\n<br>detection with region proposal networks. In: Advances in neural information\n<br>processing systems. pp. 91–99 (2015)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:12995px; width:346px; height:63px;"><span style="font-family: CMR9; font-size:8px">[29] Scarselli, F., Gori, M., Tsoi, A.C., Hagenbuchner, M., Monfardini, G.: The graph\n<br>neural network model. IEEE transactions on neural networks </span><span style="font-family: CMBX9; font-size:8px">20</span><span style="font-family: CMR9; font-size:8px">(1), 61–80 (2008)\n<br>[30] Schreiber, S., Agne, S., Wolf, I., Dengel, A., Ahmed, S.: Deepdesrt: Deep learning\n<br>for detection and structure recognition of tables in document images. In: 2017 14th\n<br>IAPR international conference on document analysis and recognition (ICDAR).\n<br>vol. 1, pp. 1162–1167. IEEE (2017)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:13061px; width:344px; height:30px;"><span style="font-family: CMR9; font-size:8px">[31] Shen, Z., Zhang, K., Dell, M.: A large dataset of historical japanese documents\n<br>with complex layouts. In: Proceedings of the IEEE/CVF Conference on Computer\n<br>Vision and Pattern Recognition Workshops. pp. 548–549 (2020)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:13094px; width:344px; height:8px;"><span style="font-family: CMR9; font-size:8px">[32] Shen, Z., Zhao, J., Dell, M., Yu, Y., Li, W.: Olala: Object-level active learning\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:155px; top:13105px; width:261px; height:8px;"><span style="font-family: CMR9; font-size:8px">based layout annotation. arXiv preprint arXiv:2010.01762 (2020)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:13116px; width:345px; height:41px;"><span style="font-family: CMR9; font-size:8px">[33] Studer, L., Alberti, M., Pondenkandath, V., Goktepe, P., Kolonko, T., Fischer,\n<br>A., Liwicki, M., Ingold, R.: A comprehensive study of imagenet pre-training for\n<br>historical document image analysis. In: 2019 International Conference on Document\n<br>Analysis and Recognition (ICDAR). pp. 720–725. IEEE (2019)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:13160px; width:346px; height:42px;"><span style="font-family: CMR9; font-size:8px">[34] Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P.,\n<br>Rault, T., Louf, R., Funtowicz, M., et al.: Huggingface’s transformers: State-of-\n<br>the-art natural language processing. arXiv preprint arXiv:1910.03771 (2019)\n<br>[35] Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detectron2. </span><span style="font-family: CMTT9; font-size:8px">https://\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:155px; top:13204px; width:207px; height:9px;"><span style="font-family: CMTT9; font-size:8px">github.com/facebookresearch/detectron2 </span><span style="font-family: CMR9; font-size:8px">(2019)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:13215px; width:345px; height:30px;"><span style="font-family: CMR9; font-size:8px">[36] </span><span style="font-family: CMR9; font-size:8px">Xu, Y., Xu, Y., Lv, T., Cui, L., Wei, F., Wang, G., Lu, Y., Florencio, D., Zhang, C.,\n<br></span><span style="font-family: CMR9; font-size:8px">Che, W., et al.: Layoutlmv2: Multi-modal pre-training for visually-rich document\n<br>understanding. arXiv preprint arXiv:2012.14740 (2020)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:13247px; width:344px; height:8px;"><span style="font-family: CMR9; font-size:8px">[37] Xu, Y., Li, M., Cui, L., Huang, S., Wei, F., Zhou, M.: Layoutlm: Pre-training of\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:155px; top:13258px; width:234px; height:8px;"><span style="font-family: CMR9; font-size:8px">text and layout for document image understanding (2019)\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:135px; top:13269px; width:216px; height:8px;"><span style="font-family: CMR9; font-size:8px">[38] Zhong, X., Tang, J., Yepes, A.J.: Publaynet:\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:188px; top:13280px; width:67px; height:8px;"><span style="font-family: CMR9; font-size:8px">layout analysis.\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:154px; top:13280px; width:237px; height:30px;"><span style="font-family: CMR9; font-size:8px">ument\n<br>Analysis and Recognition (ICDAR). pp. 1015–1022.\n<br>https://doi.org/10.1109/ICDAR.2019.00166\n<br></span></div><div style="position:absolute; border: textbox 1px solid; writing-mode:lr-tb; left:263px; top:13269px; width:219px; height:30px;"><span style="font-family: CMR9; font-size:8px">largest dataset ever for doc-\n<br>In: 2019 International Conference on Document\n<br>IEEE (Sep 2019).\n<br></span></div><div style="position:absolute; top:0px;">Page: <a href="#1">1</a>, <a href="#2">2</a>, <a href="#3">3</a>, <a href="#4">4</a>, <a href="#5">5</a>, <a href="#6">6</a>, <a href="#7">7</a>, <a href="#8">8</a>, <a href="#9">9</a>, <a href="#10">10</a>, <a href="#11">11</a>, <a href="#12">12</a>, <a href="#13">13</a>, <a href="#14">14</a>, <a href="#15">15</a>, <a href="#16">16</a></div>\n</body></html>\n')

from bs4 import BeautifulSoup

soup = BeautifulSoup(docs[0].page_content, "html.parser")
content = soup.find_all("div")

import re

cur_fs = None
cur_text = ""
snippets = []  # first collect all snippets that have the same font size
for c in content:
    sp = c.find("span")
    if not sp:
        continue
    st = sp.get("style")
    if not st:
        continue
    fs = re.findall("font-size:(\d+)px", st)
    if not fs:
        continue
    fs = int(fs[0])
    if not cur_fs:
        cur_fs = fs
    if fs == cur_fs:
        cur_text += c.text
    else:
        snippets.append((cur_text, cur_fs))
        cur_fs = fs
        cur_text = c.text
snippets.append((cur_text, cur_fs))
# Note: The above logic is very straightforward. One can also add more strategies such as removing duplicate snippets (as
# headers/footers in a PDF appear on multiple pages so if we find duplicates it's safe to assume that it is redundant info)

from langchain_core.documents import Document

cur_idx = -1
semantic_snippets = []
# Assumption: headings have higher font size than their respective content
for s in snippets:
    # if current snippet's font size > previous section's heading => it is a new heading
    if (
        not semantic_snippets
        or s[1] > semantic_snippets[cur_idx].metadata["heading_font"]
    ):
        metadata = {"heading": s[0], "content_font": 0, "heading_font": s[1]}
        metadata.update(docs[0].metadata)
        semantic_snippets.append(Document(page_content="", metadata=metadata))
        cur_idx += 1
        continue

    # if current snippet's font size <= previous section's content => content belongs to the same section (one can also create
    # a tree like structure for sub sections if needed but that may require some more thinking and may be data specific)
    if (
        not semantic_snippets[cur_idx].metadata["content_font"]
        or s[1] <= semantic_snippets[cur_idx].metadata["content_font"]
    ):
        semantic_snippets[cur_idx].page_content += s[0]
        semantic_snippets[cur_idx].metadata["content_font"] = max(
            s[1], semantic_snippets[cur_idx].metadata["content_font"]
        )
        continue

    # if current snippet's font size > previous section's content but less than previous section's heading than also make a new
    # section (e.g. title of a PDF will have the highest font size but we don't want it to subsume all sections)
    metadata = {"heading": s[0], "content_font": 0, "heading_font": s[1]}
    metadata.update(docs[0].metadata)
    semantic_snippets.append(Document(page_content="", metadata=metadata))
    cur_idx += 1

print(semantic_snippets[4])

API Reference:Document

page_content='Recently, various DL models and datasets have been developed for layout analysis
tasks. The dhSegment [22] utilizes fully convolutional networks [20] for segmen-
tation tasks on historical documents. Object detection-based methods like Faster
R-CNN [28] and Mask R-CNN [12] are used for identifying document elements [38]
and detecting tables [30, 26]. Most recently, Graph Neural Networks [29] have also
been used in table detection [27]. However, these models are usually implemented
individually and there is no unified framework to load and use such models.
There has been a surge of interest in creating open-source tools for document
image processing: a search of document image analysis in Github leads to 5M
relevant code pieces 6; yet most of them rely on traditional rule-based methods
or provide limited functionalities. The closest prior research to our work is the
OCR-D project7, which also tries to build a complete toolkit for DIA. However,
similar to the platform developed by Neudecker et al. [21], it is designed for
analyzing historical documents, and provides no supports for recent DL models.
The DocumentLayoutAnalysis project8 focuses on processing born-digital PDF
documents via analyzing the stored PDF data. Repositories like DeepLayout9
and Detectron2-PubLayNet10 are individual deep learning models trained on
layout analysis datasets without support for the full DIA pipeline. The Document
Analysis and Exploitation (DAE) platform [15] and the DeepDIVA project [2]
aim to improve the reproducibility of DIA methods (or DL models), yet they
are not actively maintained. OCR engines like Tesseract [14], easyOCR11 and
paddleOCR12 usually do not come with comprehensive functionalities for other
DIA tasks like layout analysis.
Recent years have also seen numerous efforts to create libraries for promoting
reproducibility and reusability in the field of DL. Libraries like Dectectron2 [35],
6 The number shown is obtained by specifying the search type as ‘code’.
7 https://ocr-d.de/en/about
8 https://github.com/BobLd/DocumentLayoutAnalysis
9 https://github.com/leonlulu/DeepLayout
10 https://github.com/hpanwar08/detectron2
11 https://github.com/JaidedAI/EasyOCR
12 https://github.com/PaddlePaddle/PaddleOCR
4
Z. Shen et al.
Fig. 1: The overall architecture of LayoutParser. For an input document image,
the core LayoutParser library provides a set of off-the-shelf tools for layout
detection, OCR, visualization, and storage, backed by a carefully designed layout
data structure. LayoutParser also supports high level customization via efficient
layout annotation and model training functions. These improve model accuracy
on the target samples. The community platform enables the easy sharing of DIA
models and whole digitization pipelines to promote reusability and reproducibility.
A collection of detailed documentation, tutorials and exemplar projects make
LayoutParser easy to learn and use.
AllenNLP [8] and transformers [34] have provided the community with complete
DL-based support for developing and deploying models for general computer
vision and natural language processing problems. LayoutParser, on the other
hand, specializes specifically in DIA tasks. LayoutParser is also equipped with a
community platform inspired by established model hubs such as Torch Hub [23]
and TensorFlow Hub [1]. It enables the sharing of pretrained models as well as
full document processing pipelines that are unique to DIA tasks.
There have been a variety of document data collections to facilitate the
development of DL models. Some examples include PRImA [3](magazine layouts),
PubLayNet [38](academic paper layouts), Table Bank [18](tables in academic
papers), Newspaper Navigator Dataset [16, 17](newspaper figure layouts) and
HJDataset [31](historical Japanese document layouts). A spectrum of models
trained on these datasets are currently available in the LayoutParser model zoo
to support different use cases.
' metadata={'heading': '2 Related Work\n', 'content_font': 9, 'heading_font': 11, 'source': './example_data/layout-parser-paper.pdf'}

API reference

For detailed documentation of all PDFMinerLoader features and configurations head to the API reference: https://python.langchain.com/api_reference/community/document_loaders/langchain_community.document_loaders.pdf.PDFMinerLoader.html

Related

• Document loader conceptual guide
• Document loader how-to guides