Adaptive dewarping of severely warped camera-captured document images based on document map generation

Area of Science:

• Computer Vision
• Pattern Recognition
• Document Image Analysis

Background:

• Automated dewarping of camera-captured documents is challenging.
• Existing methods often require user interaction and struggle with complex distortions.
• Current systems typically assume simple boundary shapes (trapezoidal to octahedral) with linear distortions.

Purpose of the Study:

• To propose a fully automated technique for detecting control points in camera-captured document images.
• To address simple-to-complex geometrical distortions without user intervention.
• To enhance the readability and OCR accuracy of distorted document images.

Main Methods:

• Image preprocessing
• Corner point detection
• Document map generation
• Rendering of the de-warped document image

Main Results:

• Achieved Intersection Over Union (IoU) scores of 0.92, 0.88, and 0.80 for low-, medium-, and high-complexity datasets, respectively.
• Demonstrated improved text recognition accuracy using a leading OCR engine on enhanced images.
• Qualitative analysis confirmed improved readability for severely distorted samples.

Conclusions:

• The proposed automated system effectively dewarps camera-captured handwritten documents with varying geometric distortions.
• The technique significantly enhances image quality and subsequent OCR performance.
• This method offers a reliable solution for dewarping challenging document images.