Metric rectification of curved document images

Area of Science:

• Computer Vision
• Image Processing
• Document Analysis

Background:

• Document images captured by cameras often suffer from geometric distortions due to page curvature.
• Existing methods struggle with variations in camera position, viewing angle, and document shape.
• Accurate metric rectification is crucial for subsequent analysis, such as Optical Character Recognition (OCR).

Purpose of the Study:

• To propose a novel metric rectification method for restoring geometrically distorted document images from single camera captures.
• To develop a robust framework that is insensitive to camera parameters and document page shapes.
• To enhance the accuracy of document image analysis tasks like OCR.

Main Methods:

• Constructing an isometric image mesh by exploiting page surface geometry and camera parameters.
• Modeling the curved page shape using a general cylindrical surface (GCS).
• Utilizing the line convergent symmetric property of horizontal text lines under perspective projection for parameter estimation.
• Introducing a paraperspective projection to approximate nonlinear perspective projection and deriving close-form formulas for GCS directrix and document aspect ratio estimation.

Main Results:

• The proposed method successfully restores images from single camera-captured document images.
• Comprehensive experiments on synthetic and real images demonstrate the method's efficiency.
• Comparative experiments on the CBDAR2007 dataset show superior performance over state-of-the-art methods in OCR accuracy and rectification errors.

Conclusions:

• The developed metric rectification framework offers a straightforward and robust solution for distorted document images.
• The method's insensitivity to camera and page variations makes it highly practical.
• The approach significantly improves OCR accuracy and reduces rectification errors, advancing document image analysis.