基于YOLOv4算法的冲压件缺陷检测  

Defect detection of stamping parts based on YOLOv4 algorithm

作　　者：孙永鹏 钟佩思[1,2] 刘梅[2] 曹爱霞[3] 李梁 Sun Yongpeng;Zhong Peisi;Liu Mei;Cao Aixia;Li Liang(Advanced Manufacturing Technology Center,Shandong University of Science and Technology,Qingdao 266590,China;College of Mechanical and Electronic Engineering,Shandong University of Science and Technology,Qingdao 266590,China;College of Intelligent Manufacturing,Qingdao Huanghai University,Qingdao 266427,China)

机构地区：[1]山东科技大学先进制造技术研究中心,山东青岛266590 [2]山东科技大学机械电子工程学院,山东青岛266590 [3]青岛黄海学院智能制造学院,山东青岛266427

出　　处：《锻压技术》2022年第1期222-228,共7页Forging & Stamping Technology

基　　金：山东省重点研发计划资助项目(2019GGX104102);山东省自然科学基金资助项目(ZR2017MEE066)。

摘　　要：针对冲压件缺陷检测目前存在的人工检测强度大、效率低等问题,提出了一种基于改进YOLOv4(You Only Look Once)模型的快速检测算法(YOLOv4-Mobile)。该方法使用改进的MobileNetV3网络代替YOLOv4结构中的CSPDarknet53网络,改进的MobileNetV3网络结合了深度可分离卷积、具有线性瓶颈的倒残差结构以及SE结构(轻量级注意力结构)。利用车间采集的冲压件图像,建立缺陷数据集并进行数据增强,使用K均值(K-means)聚类算法得到一组对应冲压件缺陷数据集的先验框参数,提高了先验框与特征图层的匹配度。实验结果表明:基于改进YOLOv4模型的快速检测算法的平均精度达到89%,高于SSD算法;同时,单张检测时间达到0.15 s,优于原有的YOLOv4算法。For the problems of high manual detection intensity and low efficiency in defect detection of stamping parts at present, a fast detection algorithm(YOLOv4-Mobile) based on the improved YOLOv4(You Only Look Once) model was proposed, which used the improved MobileNetV3 network to replace the CSPDarknet53 network in YOLOv4 structure, and the improved MobileNetV3 network combined a depthwise separable convolution, an inverted residual structure with a linear bottleneck and SE(Squeeze and Excitation) structure. Then, the image of stamping parts collected in the workshop was used to establish the defect data set and enhance the data set, and a set of prior frame parameters corresponding to the defect data set of stamping parts was obtained by K-means clustering algorithm to improve the matching degree of prior frame and feature layer. The test results show that based on the improved YOLOv4 model, the mAP(mean Average Precision) of the fast detection algorithm reaches 89%, which is higher than that of SSD algorithm. Meanwhile, the detection speed reaches 0.15 s per sheet, which is better than the original YOLOv4 algorithm.

关 键 词：冲压件 缺陷检测 YOLOv4 K-MEANS MobileNetV3 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TG38[自动化与计算机技术—计算机科学与技术]