DocCPLNet: Document Image Rectification via Control Point and Illumination Correction

Area of Science:

• Computer Vision
• Image Processing
• Document Analysis

Background:

• Mobile device adoption necessitates efficient document digitization.
• Image distortions and shadows degrade document quality, impacting OCR and readability.
• Existing methods struggle with complex distortion and shadow removal.

Purpose of the Study:

• To propose a novel method for rectifying geometric distortions and eliminating shadows in mobile-captured document images.
• To enhance image quality for improved optical character recognition (OCR) accuracy and human readability.
• To develop a robust solution applicable across diverse document capture scenarios.

Main Methods:

• Utilizing control points and illumination prediction for distortion rectification and shadow elimination.
• Employing spatial attention for interpolation between control and reference points to correct geometric distortions.
• Applying an illumination correction model post-geometric unwarping to remove shadows and enhance clarity.

Main Results:

• Successfully eliminated geometric distortions in captured document images.
• Effectively removed shadows and enhanced surface clarity.
• Demonstrated robust performance across diverse scenarios, achieving competitive results on the DocUNet benchmark dataset.

Conclusions:

• The proposed method effectively rectifies document distortions and removes shadows from mobile-captured images.
• Enhanced image quality leads to significant improvements in both human readability and OCR accuracy.
• The approach offers a competitive solution for real-world document digitization challenges.