Perspective rectification of camera-based document images using local linear structure

This paper addresses the problem of perspective distortion rectification of camera-based document images. A perspective rectification method using local linear structure is proposed. In this work, only local structure information such as characters, text lines, and strokes are involved in perspective parallel line detection. Horizontal lines are detected from the baselines of words or text lines, and vertical lines are detected from vertical strokes. An iterative method is used to locate vanishing points from perspective line bundles. Finally, the perspective image is rectified by a quadrilateral-rectangle pair constructed by horizontal and vertical vanishing point. Not using global information such as document boundary, paragraph format, this method has good generalization and is able to rectify partial documents or sparse documents.