电子科技2025，Vol.38Issue(10)：10-18,9.DOI:10.16180/j.cnki.issn1007-7820.2025.10.002
改进YOLOv5s的小目标钢材表面缺陷检测算法
Improved YOLOv5s Algorithm for Small Target Surface Defect Detection on Steel
摘要
Abstract
Aiming at the difficulty of feature extraction,lack of precision,error detection and missing detection of small and medium targets in steel surface defect detection,a small target detection algorithm based on improved YOLOv5s(You Only Look Once version5s)model is proposed.The focus on small target features is enhanced by in-troducing SE(Squeeze-and-Excitation)attention mechanisms into the backbone network.DSConv(Dynamic Snake Convolution)is used to replace some C3 modules in the backbone network,which effectively improves the ability of weak feature extraction.By using NWD(Normalized Wasserstein Distance)optimized EIoU(Efficient Intersection o-ver Union)loss function,the sensitivity to position deviation of small targets is reduced,and the detection perform-ance of small targets is improved.The decoupling head is introduced to optimize the model head,which solves the conflict between classification and regression tasks,reduces the occurrence of false detection and missing detection,and improves the classification and positioning accuracy of small targets.Experiments on NEU-DET(Northeastern U-niversity Detection)data set verify the effectiveness of the proposed algorithm.The mAP(mean Average Precision)of the proposed algorithm is 80.4%,which is 5%higher than that of the original algorithm,and the detection speed is maintained at 61.72 frame·s-1.The results show that the improved algorithm is superior to other comparison algo-rithms in detecting speed and precision,which proves its superiority in detecting small target steel surface defects effi-ciently.关键词
深度学习/YOLOv5s/钢材/卷积神经网络/小目标缺陷检测/注意力机制/损失函数/分类定位Key words
deep learning/YOLOv5s/steel/convolution neural network/small target defect detection/attention mechanism/loss function/classification localization分类
信息技术与安全科学引用本文复制引用
冒浩杰,巩永旺..改进YOLOv5s的小目标钢材表面缺陷检测算法[J].电子科技,2025,38(10):10-18,9.基金项目
国家自然科学基金(62301473) （62301473）
教育部人文社会科学研究基金(21YJAZH025) （21YJAZH025）
江苏省研究生科研与实践创新计划(SJCX23_XY064)National Natural Science Foundation of China(62301473) （SJCX23_XY064）
Humanities and Social Science Fund of the Ministry of Education(21YJAZH025) （21YJAZH025）
Postgraduate Research&Practice Innovation Program of Jiangsu(SJCX23_XY064) （SJCX23_XY064）
