强化特征提取的橡胶衬套表面缺陷检测方法  

A Method for Detecting Surface Defects of Rubber Bushings Based on Enhanced Feature Extraction

作　　者：华珀玺 王宸[1,2,3] 杨帅 周林 Hua Poxi;Wang Chen;Yang Shuai;Zhou Lin(School of Mechanical Engineering,Hubei University of Automotive Technology,Shiyan 442002,China;Shanghai University,Shanghai 200072,China;Shiyan Industrial Technology Research Institute of China Engineering Science and Technology,Shiyan 442000,China)

机构地区：[1]湖北汽车工业学院机械工程学院,湖北十堰442002 [2]上海大学,上海200072 [3]中国工程科技十堰产业技术研究院,湖北十堰442000

出　　处：《湖北汽车工业学院学报》2025年第1期62-67,共6页Journal of Hubei University Of Automotive Technology

基　　金：国家自然科学基金(51475150);湖北省重点研发计划(2021BAA056);湖北省高等学校中青年科技创新团队计划(NOT20200018);湖北省社科基金(21Q174);湖北省教育厅青年项目(Q20191801);湖北汽车工业学院博士科研启动基金(BK201905)。

摘　　要：针对橡胶衬套表面微小气泡和中片压伤缺陷检测中存在漏检率高、深度学习模型特征融合不足等问题,提出了基于多注意力特征融合的YOLOX-A模型。搭建橡胶衬套表面缺陷检测试验台并制作数据集,在骨干网络中添加Res2NetBlock和CoT注意力模块,在特征融合层内置坐标注意力模块。试验结果显示:YOLOX-A模型在测试集上的平均精度达到86.68%,较YOLOX提高了11.91%,其中气泡缺陷的F1值为0.99,中片压伤缺陷F1值为0.75,FPS为42.5。To address the problems of high missed detection rates and insufficient feature fusion in deep learning models for detecting surface tiny bubbles and mid-layer compression damage of rubber bushings,a YOLOX-A model based on multi-attention feature fusion was proposed.A test bench for surface defect detection of rubber bushings was constructed,and a corresponding dataset was created.The Res2NetBlock module and the CoT attention module were incorporated into the backbone network.A coordinate attention module was integrated into the feature fusion layer.The experimental results indicate that the YOLOX-A model achieves a mean average precision of 86.68% on the test dataset,which represents an 11.91% improvement over the original YOLOX model,achieves F1 scores of 0.99 for bubble defects and 0.75 for mid-layer compression damage defects,and a FPS of 42.5.

关 键 词：计算机视觉 橡胶衬套 缺陷检测 YOLOX 多注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]