一种面向钢材表面缺陷检测的改进型YOLOv5算法  

An Improved YOLOv5 Algorithm for Steel Surface Defect Detection

作　　者：李少雄 史再峰 孔凡宁 王若琪 罗韬[2] Li Shaoxiong;Shi Zaifeng;Kong Fanning;Wang Ruoqi;Luo Tao(School of Microelectronics,Tianjin University,Tianjin 300072,China;College of Intelligence and Computing,Tianjin University,Tianjin 300072,China;Tianjin Key Laboratory of Imaging and Sensing Microelectronic Technology,Tianjin 300072,China)

机构地区：[1]天津大学微电子学院,天津300072 [2]天津大学智能与计算学部,天津300072 [3]天津市成像与感知微电子技术重点实验室,天津300072

出　　处：《激光与光电子学进展》2023年第24期184-192,共9页Laser & Optoelectronics Progress

基　　金：天津市科技计划项目(22JCYBJC00140)。

摘　　要：针对钢材表面缺陷尺度不一,现有检测算法多尺度特征处理能力较差、精度有待提高的问题,提出了一种面向钢材表面缺陷检测的改进型YOLOv5算法。首先,在骨干网络的特征输出层后添加感受野模块以增强特征的判别性与鲁棒性,可以更好地感知不同尺度的特征信息;然后,利用对齐的特征聚合模块替换传统的特征融合结构,解决了高低分辨率特征图在融合过程中存在的特征错位问题;最后,采用带有高效通道注意力机制的解耦头输出检测结果,注意力机制可以自适应地校准通道响应,解耦头使得分类与回归任务可以独立执行。在NEU-DET数据集上的实验结果显示,所提出方法的平均精度均值为80.51%,相比基准模型提升了4.48%,检测速度为31.96 frame/s。相比其他主流的目标检测算法,在保持一定检测速度的前提下,所提算法具有更高的精度,能够实现高效的钢材表面缺陷检测。Scale of steel surface defects is different,but existing detection algorithms have poor multi-scale feature processing ability and low accuracy.Therefore,an improved YOLOv5 algorithm for steel surface defect detection is proposed.First,receptive field modules are added after the feature output layer of the backbone to enhance the discrimination and robustness of the features which can better perceive the feature information of different scales.Then,aligned feature aggregation modules are used to replace the traditional feature fusion structure to solve the feature misalignment problem in the fusion process of high and low resolution feature maps.Finally,decoupled heads with efficient channel attention mechanisms are used to output the detection results.The attention mechanism can adaptively calibrate the channel response,and the decoupled heads enable classification and regression tasks to be performed independently.The experimental results on NEU-DET dataset show that the mean average precision of the proposed method is 80.51%,which is 4.48%higher than that of the benchmark model,and the detection speed is 31.96 frame/s.Compared with other mainstream object detection algorithms,the proposed algorithm has higher accuracy while maintaining certain detection speed,enabling efficient steel surface defect detection.

关 键 词：表面缺陷检测 感受野 特征对齐 解耦头 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]