汽车漆面缺陷高精度检测系统  

High precision detection system for automotive paint defects

作　　者：陆玉凯 袁帅科 熊树生[1,4] 朱绍鹏 张宁[3] LU Yu-kai;YUAN Shuai-ke;XIONG Shu-sheng;ZHU Shao-peng;ZHANG Ning(Power Machinery&Vehicular Engineering Institute,Zhejiang University,Hangzhou 310014,China;Zhejiang Geely Automobile Co.,Ltd.,Hangzhou 310051,China;College of Mechanical Engineering,Yanshan University,Qinhuangdao 066004,China;Longquan Industrial Innovation Research Institute,Longquan 323700,China)

机构地区：[1]浙江大学动力机械及车辆工程研究所,杭州310014 [2]浙江吉利汽车有限公司,杭州310051 [3]燕山大学机械工程学院,河北秦皇岛066004 [4]龙泉产业创新研究院,浙江龙泉323700

出　　处：《吉林大学学报（工学版）》2024年第5期1205-1213,共9页Journal of Jilin University:Engineering and Technology Edition

基　　金：工信部重点领域及特定场景工业互联网平台应用项目(TC200802D);工信部“5G+工业互联网”高质量网络和公共服务平台-离散行业高质量网络项目(TC200A00N)。

摘　　要：汽车涂装过程中产生的漆面缺陷影响着整车外观质量,针对人工检测存在漏检、低效以及传统检测方案的高实施成本等问题,提出了一种基于改进YOLOv7算法的汽车漆面缺陷检测系统。构建了汽车漆面缺陷数据集,共有4023张图像,其中包含5种常见汽车漆面缺陷;针对YOLOv7算法在微小缺陷上检测精度不足的问题,在原网络中引入了GAM注意力机制和SPPFCSPC模块,用于提高算法对微小缺陷特征的提取能力,同时采用改进的ELAN模块对网络结构进行改进,减少网络过深造成的小目标信息丢失问题,保证在减轻网络模型的同时提高网络对微小特征的识别精度;实验结果表明:本文方法大幅提升了对微小漆面缺陷的检测性能,缺陷的平均检测精度达到了88.9%,与多种算法相比检测精度最高。The paint defects that exist during the automotive painting process affect the overall appearance quality of the car. In response to the problems of missed inspection,low efficiency,and high implementation cost of traditional inspection schemes in manual inspection,a paint defect detection method based on the improved YOLOv7 algorithm is proposed. A dataset of automotive paint defects was constructed,consisting of 4023 images,including 5 types of automotive paint defects;In response to the problem of insufficient detection accuracy of YOLOv7 algorithm on small defects,GAM attention mechanism and SPPFCSPC module were introduced into the original network to improve the algorithm′s ability to extract small defect features. At the same time,an improved ELAN module was used to improve the network structure to reduce the problem of small target information loss caused by deep network,ensuring that the network model is reduced while improving the recognition accuracy of small features;Based on the constructed dataset,the defect detection performance of different algorithms was tested and the effectiveness of the module was verified. The experimental results show that this method significantly improves the detection ability of small defects on paint surfaces,with an average detection accuracy of 88.9%,which is the highest detection accuracy compared to various algorithms.

关 键 词：车辆工程 汽车漆面 缺陷检测 深度学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]