南京理工大学学报（自然科学版）2026，Vol.50Issue(2)：161-171,11.DOI:10.14177/j.cnki.32-1397n.2026.50.02.006
基于深度学习的流程图线段检测方法
Deep learning-based method for detecting line segments in flowcharts
摘要
Abstract
Flowchart line segment detection has been underexplored due to the scarcity of high-quality datasets and the weak interference resistance of endpoint labels.To address these challenges,this article proposes a flowchart line segment detection method based on label reshaping and dual-metric combined with the YOLOv5 model.Firstly,the FlowchartLine dataset containing 19 647 line segments is constructed to provide abundant training samples.Secondly,an endpoint-guided label reshaping module is introduced to transform the task from traditional endpoint detection to object detection,effectively mitigating the false detection issues caused by adjacent endpoints in traditional endpoint detection methods that lack holistic line segment representation.Finally,a geometry-aware enhanced joint loss is developed by coupling normalized Wasserstein distance(NWD)with original complete intersection over union(CIoU)loss to improve precise correction capability for line segment detection.Experimental results demonstrate that the proposed method achieves structural average precision scores of 98.2%,98.3%and 98.4%at sAP5,sAP10 and sAP15 thresholds respectively,outperforming current line segment detection method by 1.6%,1.0%and 0.9%,demonstrating superior accuracy in flowchart line segment detection.关键词
流程图/线段检测/标签重塑/目标检测/YOLOv5/归一化瓦斯坦距离Key words
flowchart/line segment detection/label reshaping/object detection/YOLOv5/normalized Wasserstein distance分类
信息技术与安全科学引用本文复制引用
姚瑶,陈涛,孙泽人..基于深度学习的流程图线段检测方法[J].南京理工大学学报（自然科学版）,2026,50(2):161-171,11.基金项目
国家自然科学基金(62506169) （62506169）
中央高校基本科研业务费专项资金(30923010303) （30923010303）
