A Wealth of Pre-trained Models on Different Datasets

To accommodate heterogenous document layout structures, Layout Parser a collection of DL models trained on different datasets. Currently the are 9 models trained on 5 diverse datasets, and they can be loaded via a unified interface.

▲ Exemplar images in the 5 used datasets (screenshots are taken from their papers or open-sourced datasets).

See Also:

• Layout Parser Model Zoo

Easy Layout Data Manipulation and Processing

Layout Parser supports different levels of abstraction of layout data, and provide three classes of representation for layout data, namely, Coordinates, TextBlock, and Layout. The same operations and transformations are supported inter and intra these classes to maximize the efficiency when processing the layout data.

See Also:

• Layout Data Structure Documentation

Highly Customizable Layout Data Visualization

Layout Parser visualizes the layout data using a simple syntax: lp.draw_box or lp.draw_text. It provides two modes for displaying the layout data: Mode I directly overlays the layout region bounding boxes and categories over the original image. Mode II recreates the original document via drawing the OCR’d texts at their corresponding positions on the image canvas.

See Also:

• Layout Parser Visualization Documentation

Export Layout Data in Your Favorite Format

Layout Parser supports loading and exporting layout data to different formats, including general formats like csv, json, or domain-specific formats like PAGE, COCO, or METS/ALTO format (Full support for them will be released soon). It provides the flexibility for integrating Layout Parser with other document image analysis pipelines, and makes it easy to share your outputs with the community.

Easy to Train Your Own Layout Models with Best Accuracy

Layout Parser also comes with full support for customized layout model training on your own dataset. This enables you to achieve optimal prediction accuracy on your own dataset and can simplify your pipeline. Different training mode are supported, i.e., training from scratch or fine-tuning from an existing model. Learn more details at Layout-Parser/layout-model-training.

See Also:

• The layout-model-training Repository

Efficiently Creating Your Training Dataset

No labeled training data available? Don’t worry! Layout Parser also incorporates a data annotation toolkit that enables creating the training dataset much more efficiently. Shown in the illustration below, the tool loads layout predictions from pre-trained models (a), and users only need to select and check a small percentage of model predictions to correct or relabel (b). False-Negative Highlighter (c) helps recognize mis-identified objects from the model predictions. After these steps, the full image annotation is created with less effort.

▲ Illustration of the annotation interface with Object-Level Active Learning features. Learn more in this paper.

See Also:

• Layout Annotation Service
• The PAWLS Annotation Tool

Layout Models & Pipelines Sharing Platform

Layout Parser also aims to create a community platform for document image analysis (DIA) research and application. One key challenge in current DIA is the reusability of both layout models and pipelines. Layout Parser maintainers are currently working on implementing the platform for practitioners to share their models and pipelines easily. By streamline the distribution of layout models, as well as promoting the reusability of layout pipelines, it can potentially transform the DIA workflow and pipelines broadly.

Note: This is Layout Parser’s presentation video at ICDAR 2021, a top venue for document recognition and understanding. In this talk, we discussed the design of the Layout Parser and demoed its key features.