基于深度学习的多尺寸汽车轮辋焊缝检测与定位系统研究  

Research on multi size automobile rim weld detection and positioning system based on depth learning

作　　者：潘睿志 林涛[1] 李超[1] 胡波[1] PAN Ruizhi;LIN Tao;LI Chao;HU Bo(College of Mechanical and Electrical Engineering,Chengdu University of Technology,Chengdu 610059,China)

机构地区：[1]成都理工大学机电工程学院,四川成都610059

出　　处：《光学精密工程》2023年第8期1174-1187,共14页Optics and Precision Engineering

基　　金：成都市重点研发支撑计划资助项目(No.2018-YFYF-00002-GX);成都理工大学科研启动基金资助项目(No.10912-KYQD2019_07733)。

摘　　要：为了实现汽车轮辋生产装备自动化与智能化,提升汽车轮辋的生产效率,降低人工成本,本文提出了一种基于YOLOv5s算法的多尺寸汽车轮辋焊缝检测与定位系统。首先,由图像采集装置拍摄实际生产中的多尺寸轮辋焊缝图像,构建轮辋焊缝数据集,使用K-means算法重新生成数据集锚定框,提升网络的收敛速度和特征提取能力;其次,引入CBAM(Convolutional Block Attention Module)混合域注意力机制,提高模型对于轮辋焊缝关注度,减少背景干扰;然后,采用EIOU(Efficient Intersection Over Union Loss)边框位置回归损失函数,提高轮辋焊缝识别框的准确率;最后,增加了ASFF(Adaptively Spatial Feature Fusion)自适应特征融合网络,使目标检测模型对多个级别的特征进行空间滤波。实验结果表明,改进后的算法准确率和mAP0.5分别达到了98.4%和99.2%,相比于原YOLOv5s算法分别提高了4.5%和3.7%。训练好的模型采用推理加速框架TensorRT进行加速部署在工控机上,搭配视觉检测软件与工业触摸屏形成交互及显示平台,经过在实际生产环境对多批次不同尺寸的3000个轮辋焊缝验证,其漏检率在0.5%左右,满足汽车企业对于多尺寸轮辋焊缝检测精度要求。In order to realize the automation and intelligence of automobile rim production equipment,im⁃prove the production efficiency of automobile rims,and reduce labor costs,this paper proposes a multi size automobile rim weld detection and positioning system based on YOLOv5s algorithm.First,the image ac⁃quisition device captures the image of the multi size rim weld seam in actual production,builds the rim weld seam data set,and uses K-means algorithm to regenerate the anchor frame of the data set to improve the convergence speed and feature extraction ability of the network;Secondly,CBAM(Convolutional Block Attention Module,CBAM)mixed domain attention mechanism is introduced to improve the model's attention to the rim weld and reduce background interference;Then,EIOU(Efficient Intersection Over Union Loss,EIOU)frame position regression loss function is used to improve the accuracy of rim weld identification frame;Finally,ASFF(Adaptive Spatial Feature Fusion,ASFF)adaptive feature fu⁃sion network is added to enable the target detection model to perform spatial filtering on multiple levels of features.The experimental results show that the accuracy and mAP0.5 of the improved algorithm are 98.4%and 99.2%respectively,which are 4.5%and 3.7%higher than the original YOLOv5s algo⁃rithm.The trained model is accelerated and deployed on the industrial personal computer using the reason⁃ing acceleration framework TensorRT,and forms an interactive and display platform with the visual in⁃spection software and the industrial touch screen.Through the verification of 3000 wheel rim welds of dif⁃ferent sizes in multiple batches in the actual production environment,the leakage rate is about 0.5%,which meets the requirements of automobile enterprises for the detection accuracy of multi size wheel rim welds.

关 键 词：焊缝 损失函数 YOLOv5s 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术]