Unofficial implementation of DocMAE: Document Image Rectification via Self-supervised Representation Learning

https://arxiv.org/abs/2304.10341

• Document background segmentation network using U2 net
• Synthetic data generation for self-supervised pre-training
• Pre-training
• Fine-tuning for document rectification (In progress)
• Evaluation
• Code clean up and documentation
• Model release

Find a jupyter notebook at demo/background_segmentation.ipynb

• 3411482 pages from ~1M documents from Docile dataset (https://github.com/rossumai/docile)
• Rendered with Doc3D https://github.com/Dawars/doc3D-renderer
• 558 HDR env lighting from https://hdri-haven.com/

Pretraining on 200k documents:

python pretrain.py -c config/config.json Visualize trained model using https://github.com/NielsRogge/Transformers-Tutorials/blob/master/ViTMAE/ViT_MAE_visualization_demo.ipynb

Test documents come from DIR300 dataset https://github.com/fh2019ustc/DocGeoNet