改进YOLOv5模型的带钢表面缺陷检测方法
Surface defect detection method of strip steel by improved YOLOv5 model
陈万志 1张春光1
作者信息
- 1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105
- 折叠
摘要
针对带钢表面缺陷检测中的漏检和精度较低问题,提出一种融合 swin-transformer 和坐标注意力(coordinate attention,CA)模块的改进 YOLOv5 模型检测方法.在 YOLOv5 模型的主干网络中引入swin-transformer 特征提取模块,使主干网络更聚焦于图像全局特征信息的提取;在特征融合网络输出分支末端嵌入CA模块,进一步增强目标缺陷方向和位置信息的敏感度.研究结果表明:改进模型在NEU-DET数据集上的平均精度值(mAP)达到了 77.6%,较原YOLOv5 模型提高了 3 个百分点.改进模型提升了带钢表面缺陷检测精度,具有更好的缺陷检测能力.
Abstract
Aiming at the problem of missing detection and low precision in strip surface defect detection,an improved YOLOv5 network model detection method combining swin-transformer and coordinate attention module(CA)is proposed.Firstly the swin-transformer module of feature extraction is introduced into the feature extraction backbone network of YOLOv5 to make it the backbone network more focused on extracting global feature information.Secondly,embed CA module at the end of the output branch of feature fusion network to further enhance the sensitivity of target defect direction and position information.The results show that the average precision value(mAP)of the improved network model on the NEU-DET data set reaches 77.6%,which is 3.0%higher than the original YOLOv5 network model.The improved network model improves the detection accuracy of strip surface defects and has better defect detection ability.
关键词
带钢表面缺陷检测/swin-transformer模块/坐标注意力模块/YOLOv5网络Key words
surface defect detection of strip steel/swin-transformer module/coordinate attention module/YOLOv5 network引用本文复制引用
基金项目
国家重点研发计划项目(2018YFB1403303)
辽宁省教育厅高等学校基本科研项目(LJKZ0327)
出版年
2024
国家科技期刊平台
NSTL
万方数据