Important fixes for multibackend layout model support:

• Resolves the issues mentioned in #78 with other fixes to improve the multibackend layout model support #79
• Better tests for different backends #79 for preventing future related issues

• Fixes for automatically setting label_map in Detectron2LayoutModel #75
• Remove unnecessary class annotations (that might breaks Python 3.6 users) #75

We are excited to release LayoutParser v0.3.0, with a lot of exciting updates and functional improvements.

New Features

• The biggest change in this version is that LayoutParser now supports multiple deep learning backends: Detectron2, effdet, and paddledetection. This allows for more flexible usage of the layoutparser library, and makes it easier for implementing customized layout models in the future. #54 #67
• Additionally, the newly added AutoModel and improved model configuration parsing makes it easier load and use the layout detection models. #69
  • e.g, model = lp.AutoLayoutModel("lp://efficientdet/PubLayNet").
• To support this multi-backend framework, we implement the dynamic importing mechanism as well as better ways for installing layoutparser and the needed dependencies (see instructions). #65 #68
• And now layoutparser supports directly loading PDF files into as layout objects: #71

  import layoutparser as lp
  pdf_layout, pdf_images = lp.load_pdf("path/to/pdf", load_images=True)
  lp.draw_box(pdf_images[0], pdf_layout[0])
  

• To support more flexible processing of the layout objects, a set of new toolkits are available: #72

  import layout parser as lp
  page_layout = lp.load_pdf("tests/fixtures/io/example.pdf")[0]
  pdf_lines = lp.simple_line_detection(page_layout)
  

New Models

• Add MFD model that can detect (display) equation regions within scientific documents #59

Layout Parser v0.2.0 Release Notes

New Features

1. Support for loading and exporting the layout data in json and csv , see #6
2. Add support for union and intersect operations, see #20 and the detailed explanation

Improvements

1. Functional improvements:

   1. When loading Layout Parser official models, Detectron2LayoutModel can automatically detect the label_map, . For example,

      model = lp.Detectron2LayoutModel("lp://HJDataset/faster_rcnn_R_50_FPN_3x/config")
      model.label_map
      # {1: 'Page Frame', ... }
      

   2. Detectron2LayoutModel now supports the enforce_cpu flag that enforces using cpu even when CUDA devices are available.

   3. For visualization.draw_box, it now supports a show_element_type flag that shows the bbox category name on the top left corner of the layout objects.

2. Improve installation command and documentation, especially for installing Detectron2 on Windows platforms #25

New Models

1. Add the table bank detection models that can identify table regions

Fixes

1. Fix the incorrect layout issue mentioned in #9 - Thanks to @remidbs.
2. Fix the some of the dependency issues mentioned in #11 and #13 by using iopath instead of fvcore. See #18, Thanks to @edisongustavo.

Improvements:

• Supports lazy loading for the Detectron2 module. Now the dependency for Detectron2 will be requested only when you explicitly create a Detectron2LayoutModel object. This might be helpful for using the plain layoutparser library without installing the Detectron2 module.

New models:

• Incorporated a pre-trained model based on the NewspaperNavigator dataset: lp://NewspaperNavigator/faster_rcnn_R_50_FPN_3x/config

Fixes:

• Corrected a bug in visualization that might overwrite original the image

In this version, we released a new model for publaynet and made several improvements:

1. We released the mask_rcnn_X_101_32x8d_FPN_3x model trained on the publaynet dataset. Note: it's been trained on the full training set (while others are only trained on the validation set), and you could expect a 15% performance improvement based on this new model.
2. We improved the support for PIL images for both layout modeling and visualization
3. We improved the Default Language Settings for the Tesseract OCR model

Fixes

• Fixed a bug that could cause errors in loading Prima Models

Updates

• Update the prima MASK RCNN model with higher accuracy, and listed detailed evaluation reports.

layoutparser now supports the following functionalities:

• Coordinate system:

  • Supports the 3 basic coordinate system and their geometric relationships
  • Supports the TextBlook and Layout system for convenient coordinate and text processing

• OCR System:

  • Supports OCR based on Google Cloud Vision and Tesseract API.

• Layout Modeling:

  • Supports using pre-trained Deep Learning models for layout object detection using Detection2

• Visualization:

  • Supports highly-customizable presentation of the box coordinates and text in the detected layout