基于RCSI-YOLOv5的轴承表面缺陷检测方法  

Bearing surface defect detection method based on RCSI-YOLOv5

作　　者：卜扬 屈霞[1] 陈涛[1] 武伟宁 BU Yang;QU Xia;CHEN Tao;WU Wei-ning(School of Mechanical Engineering and Rail Transit,Changzhou University,Changzhou 213164,China;Wang Zheng School of Microelectronics,Changzhou University,Changzhou 213164,China)

机构地区：[1]常州大学机械与轨道交通学院,江苏常州213164 [2]常州大学王诤微电子学院,江苏常州213164

出　　处：《陕西科技大学学报》2025年第2期203-214,共12页Journal of Shaanxi University of Science & Technology

基　　金：教育部高等教育司产学合作协同育人项目(230804973313225,230801913313416)。

摘　　要：针对轴承表面缺陷检测中的小目标漏检、相似特征目标误检、高低质量样本不平衡等问题,提出一种基于RCSI-YOLOv5的轴承表面缺陷检测模型.在主干网络中构建Res2ConvModC3特征提取模块,来提高模型对浅层小目标的特征提取能力和对相似特征的辨别能力;在颈部网络前端设计CGCA注意力机制,增强网络对目标特征的定位能力;在检测头中加入了SimAM注意力机制,提高模型对微小缺陷目标的关注度;设计ISInner-CIoU计算边界框回归损失,缓解高低质量样本不平衡问题.实验结果表明,与原YOLOv5算法相比,RCSI-YOLOv5的mAP@0.5提升1.5%,F1-Score提升1%,凹槽、擦伤、划痕的AP分别提升2.1%、0.5%、1.7%,FNR分别降低1.3%、0.4%、2.1%.有效提升了模型的检测精度,减少了目标的漏检、误检.To address the issues of missed detection of small targets,false detection of targets with similar features,and imbalance between high-quality and low-quality samples in bearing surface defect detection,a bearing surface defect detection model based on RCSI-YOLOv5 is proposed.A Res2ConvModC3 feature extraction module is constructed in the backbone network to enhance the model′s capability in extracting features of shallow small targets and discriminating similar features;a CGCA attention mechanism is designed at the front end of the neck network to enhance the network′s ability to locate target features;a SimAM attention mechanism is added in the detection head to increase the model′s focus on minute defect targets;ISInner-CIoU is designed to compute the bounding box regression loss,mitigating the issue of imbalance between high-quality and low-quality samples.Experimental results show that compared to the original YOLOv5 algorithm,the mAP@0.5 of RCSI-YOLOv5 increases by 1.5%,and the F1-Score improves by 1%,with the AP for grooves,abrasions,and scratches increasing by 2.1%,0.5%,and 1.7%respectively,while the FNR decreases by 1.3%,0.4%,and 2.1%respectively.This effectively enhances the model′s detection accuracy and reduces both missed and false detections.

关 键 词：轴承表面缺陷检测 YOLOv5 卷积调制 CGCA SimAM ISInner-CIoU 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]