Synthetic Document Images with Diverse Shadows for Deep Shadow Removal Networks

Area of Science:

• Computer Vision
• Digital Image Processing
• Machine Learning

Background:

• Document shadow removal is crucial for digitized documents.
• Existing methods struggle with limited diverse datasets and synthetic data limitations.
• Synthesized datasets often lack document diversity and varied lighting conditions.

Purpose of the Study:

• Introduce a large-scale, diverse Synthetic Document with Diverse Shadows (SynDocDS) dataset.
• Propose a Dual Shadow Fusion Network (DSFN) for robust document shadow removal.
• Enhance deep shadow removal network performance using diverse synthetic data.

Main Methods:

• Developed the SynDocDS dataset using a physics-based illumination model for diverse shadow rendering.
• Proposed the Dual Shadow Fusion Network (DSFN) with high global color comprehension.
• Trained and evaluated models on SynDocDS and public datasets (OSR, Kligler's, Jung's).

Main Results:

• Training on SynDocDS improved average PSNR from 23.00 dB to 25.70 dB and SSIM from 0.959 to 0.971.
• The proposed DSFN outperformed existing networks on multiple metrics (PSNR, SSIM).
• Demonstrated significant improvements in Optical Character Recognition (OCR) performance post-shadow removal.

Conclusions:

• The SynDocDS dataset enables training of more robust and high-performance shadow removal networks.
• DSFN effectively handles document-specific color features for superior shadow removal.
• The proposed approach significantly advances document image digitization quality and usability.