Free-form document detection and ratio-preserving rectification
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Document detection and rectification are essential preprocessing steps for transforming mobile camera images of document into images suitable for optical character recognition (OCR). Conventional approaches rely on geometric features such as corners and edges to detect documents and assume standardized formats (e.g., A4) for rectification. Although these methods do not require a flatbed scanner, they are limited to specific paper sizes. In this paper, we propose a framework for detecting and rectifying arbitrary rectangular documents without assuming fixed aspect ratios. Our approach introduces CornerNet, which estimates the probability distributions of the four document corners, thereby enabling robust detection at arbitrary viewing angles. Based on these detected corners, we also designed a ratio-preserving rectification module that estimates the intrinsic aspect ratio of the document and normalizes it to a reference size. Finally, a sub-pixel refinement step corrects the rectified corners to increase precision. The experimental results demonstrate that the proposed method achieves accurate corner localization and rectification that preserves the aspect ratio, enabling reliable document scanning from unconstrained mobile phone images.
Abstract
Document detection and rectification are essential preprocessing steps for transforming mobile camera images of document into images suitable for optical character recognition (OCR). Conventional approaches rely on geometric features such as corners and edges to detect documents and assume standardized formats (e.g., A4) for rectification. Although these methods do not require a flatbed scanner, they are limited to specific paper sizes. In this paper, we propose a framework for detecting and rectifying arbitrary rectangular documents without assuming fixed aspect ratios. Our approach introduces CornerNet, which estimates the probability distributions of the four document corners, thereby enabling robust detection at arbitrary viewing angles. Based on these detected corners, we also designed a ratio-preserving rectification module that estimates the intrinsic aspect ratio of the document and normalizes it to a reference size. Finally, a sub-pixel refinement step corrects the rectified corners to increase precision. The experimental results demonstrate that the proposed method achieves accurate corner localization and rectification that preserves the aspect ratio, enabling reliable document scanning from unconstrained mobile phone images.