Rectification of curved document images based on single view three-dimensional reconstruction

Area of Science:

• Computer Vision
• Image Processing
• Document Analysis

Background:

• Distortions in camera-captured document images degrade Optical Character Recognition (OCR) accuracy.
• Existing methods often require multiple images or specific hardware, limiting practical application.
• A flexible, single-image solution is needed for effective document digitalization.

Purpose of the Study:

• To propose a novel framework for 3D reconstruction and rectification of camera-captured document images.
• To develop a practical method that handles general, locally smooth document surfaces using only a single input image.
• To improve the accuracy of OCR systems by removing image distortions.

Main Methods:

• Iterative refinement scheme for baseline fitting from text line components.
• Efficient discrete vertical text direction estimation using convex hull projection profile analysis.
• 2D distortion grid construction via text direction function estimation with 3D regularization.

Main Results:

• The proposed framework successfully performs 3D reconstruction and rectification on single camera-captured document images.
• Experimental evaluations show superior performance compared to recent methods in visual distortion removal.
• Significant improvements in OCR accuracy were achieved post-rectification.

Conclusions:

• The novel framework offers a flexible and practical solution for camera-captured document image rectification.
• The method effectively removes distortions, leading to enhanced document digitalization and OCR performance.
• This approach advances the field of document image analysis and processing.