YOLO-Based Corner Extraction Methods for Document Image Rectification
Baojun Dai, Yuhang Luo, Lin Feng Wu, Ming-Tzau Lin, Jung-Kuei Yang · 2025
Geometric rectification of document images relies on precise corner localization. This study presents two YOLO-based corner extraction approaches: YOLO Pose (direct corner regression) and YOLO Segmentation (corner extraction via segmentation masks). To address structural distortions frequently observed in YOLO Pose predictions, we introduce a novel Geometric Structure Consistency Loss (GSCL) that constrains edge length ratios, angles, and relative corner positions. For validation, a publicly available dataset was constructed covering diverse shooting angles, distances, and backgrounds. Experimental results demonstrate YOLO Segmentation's superior structural robustness under complex conditions, achieving an average IoU above 0.97. Incorporating GSCL, significantly enhances the geometric stability and overall rectification accuracy of the YOLO Pose approach. The dataset is open-sourced to facilitate future research.