改进轻量化VTG-YOLOv7-tiny的钢材表面缺陷检测
Improving the lightweight VTG-YOLOv7-tiny for steel surface defect detection
梁礼明 1龙鹏威 1冯耀 1卢宝贺1
作者信息
- 1. 江西理工大学 电气工程与自动化学院,江西 赣州 341000
- 折叠
摘要
针对钢材表面缺陷形态多样、结构复杂且存在检测目标漏检现象和算法参数量过大等问题,提出一种轻量化VTG-YOLOv7-tiny的钢材缺陷检测算法.该方法一是设计VoVGA-FPN网络,以减少信息传递过程中的丢失,增强网络特征融合能力;二是构建三重坐标注意力机制,提升模型对空间和通道信息的特征提取能力;三是引入鬼影混洗卷积,在提高精度的同时降低模型参数量和计算量;四是增加大目标检测层,改善特征图中部分缺陷占比较大,导致检测精度低的问题.在NEU-DET和Severstal钢材缺陷数据集进行实验验证,改进后算法与原模型相比,mAP分别提升5.7%和8.5%;参数量和计算量分别降低0.61 M和4.2 G;精确度和召回率分别提升7.1%,1.8%和8.9%,7.0%.实验结果表明,改进后的算法更好地平衡了检测精度和轻量化,为边缘终端设备提供了参考.
Abstract
To address the problems of diverse and complex shapes of steel surface defects,detection target missing,and large number of algorithm parameters,a lightweight VTG-YOLOv7-tiny steel defect detec-tion algorithm was proposed.The method first designed VoVGA-FPN network to reduce the loss of infor-mation during information transmission and enhance the network feature fusion ability;second,it con-structed a triple coordinate attention mechanism to improve the model's feature extraction ability of spatial and channel information;third,it introduceed ghost shuffle convolution to reduce the model parameters and computation while improving the accuracy;fourth,it added a large target detection layer to improve the problem that some defects in the feature map occupy a large proportion,resulting in low detection accu-racy.The improved algorithm was verified on the NEU-DET and Severstal steel defect datasets.Com-pared with the original model,the mAP of the improved algorithm is increased by 5.7%and 8.5%,re-spectively;the parameters and computation are reduced by 0.61 M and 4.2 G,respectively;the accuracy and recall are increased by 7.1%,1.8%and 8.9%,7.0%,respectively.The experimental results show that the improved algorithm better balances the detection accuracy and lightweight,and provides a refer-ence for edge terminal devices.
关键词
缺陷检测/轻量化YOLOv7-tiny/VoVGA-FPN网络/三重坐标注意力/鬼影混洗卷积Key words
defect detection/Lightweight YOLOv7-tiny/VoVGA-FPN network/Triplet Coordinate Attention(TCA)/Ghost Shuffle Convolution(GSConv)引用本文复制引用
基金项目
国家自然科学基金资助项目(51365017)
国家自然科学基金资助项目(6146301)
江西省自然科学基金资助项目(20192BAB205084)
江西省教育厅科学技术研究重点项目(GJJ170491)
出版年
2024
国家科技期刊平台
NSTL
万方数据