软件导刊2025，Vol.24Issue(8)：65-71,7.DOI:10.11907/rjdk.241525
基于改进的Yolov5s钢板表面缺陷检测模型研究
Research on Steel Plate Surface Defect Detection Model Based on Improved Yolov5s
摘要
Abstract
For the steel plate surface characterization in steel production,the steel plate surface color distribution is shown as a continuous grayscale with subtle variations thus leading to redundant feature maps.Traditional target detection algorithms use conventional convolution for feature extraction,which generates a large number of unnecessary model parameters and computational resource consumption when dealing with redundant features.To solve this problem,this paper proposes an improved lightweight steel plate surface defect detection model using Yolov5s as the base model.First,Ghost convolution is used to replace the traditional convolution in Yolov5s,which effectively solves the gray-scale map feature redundancy problem and significantly reduces the number of model parameters and floating-point computation.Meanwhile,in order to solve the problem of accuracy degradation of the model that may be caused by the reduction of parameters and computation,this study introduces the SE attention mechanism in the neck module of the base model to improve the sensitivity of the model to key features.In this study,on a steel production line,steel plate surface defect data are collected on site and the dataset is labeled and preprocessed for model training and validation.The experimental results show that the improved Yolov5s model achieves a mean accuracy(mAP)of 88.7%,a model volume of 7.54 MB,and a floating-point computation of 8.1 GFLOPs.compared with the pre-improved YOLOv5s,the model volume is com-pressed by 45.4%and the computation volume is reduced by 48.7%while keeping the mAP constant.This improved model is easier to be de-ployed on resource-limited devices and provides an efficient and practical solution in the field of steel plate surface defect detection.关键词
Yolov5s/Ghost/SE注意力机制/表面缺陷检测Key words
Yolov5s/Ghost/SE attention mechanism/surface defect detection分类
信息技术与安全科学引用本文复制引用
宁安安,胡安明..基于改进的Yolov5s钢板表面缺陷检测模型研究[J].软件导刊,2025,24(8):65-71,7.基金项目
广州理工学院校级科研项目(2023KYQ016) （2023KYQ016）
