基于改进GANomaly网络的旋开盖缺陷检测方法  

A defect detection method for screw-on caps based on improved GANomaly network

作　　者：舒军[1] 王祥 李灵 雷建军[2] 何俊成 杨莉[2] SHU Jun;WANG Xiang;LI Ling;LEI Jianjun;HE Juncheng;YANG Li(College of electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China;College of Computer,Hubei University of Education,Wuhan 430025,China)

机构地区：[1]湖北工业大学电气与电子工程学院,武汉430068 [2]湖北第二师范学院计算机学院,武汉430205

出　　处：《中南民族大学学报（自然科学版）》2023年第6期788-798,共11页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：国家自然科学基金资助项目(61601176)。

摘　　要：基于现有瓶盖缺陷检测方法,提出了一种结合低照度增强、语义分割和异常检测的旋开盖缺陷检测方法.受拍摄光照和目标特征多样性的影响,传统语义分割方法对低照度图像分割不准确.为解决此问题,通过基于最大熵的Retinex模型增强低照度图像,选取OCR-Net语义分割网络分割去除背景.在检测缺陷时,半监督异常检测GANomaly网络解决了正常瓶盖样本和缺陷瓶盖样本不平衡的问题.但图像重建效果差,为此基于十字交叉注意力和最小二乘损失函数改善GANomaly网络对图像的重建能力.实验结果表明:低照度图像增强和语义分割解决了瓶盖图像因亮度低而分割不准确的问题,改进的GANomaly网络在瓶盖缺陷检测中,改善了图像重建效果,AUC值达到了0.71,且在MvTec AD数据集上表现优越,具有较好的应用价值.A screw-on cap defect detection method is proposed based on existing cap defect detection methods,combining low illumination enhancement,semantic segmentation and anomaly detection.Affected by the shooting illumination and the diversity of target features,the traditional semantic segmentation method is inaccurate for low illumination image segmentation.To solve this problem,low-illumination images are enhanced by a Retinex model based on maximum entropy,and the OCR-Net semantic segmentation network is selected to segment and remove the background.For detecting defects,the anomaly detection GANomaly network solves the imbalance between normal and defective bottle cap samples.However,the image reconstruction effect is poor,so the Criss-Cross Attention and least-square loss function are used to improve the image reconstruction ability of GANomaly network.Experimental results show that low-illumination image enhancement and semantic segmentation solve the problem of inaccurate bottle cap segmentation due to low illumination.The improved GANomaly network improvs image reconstruction in bottle cap defect detection with an AUC value of 0.71 and performs well on the MvTec AD dataset,which has excellent application value.

关 键 词：旋开盖缺陷检测 低照度图像增强 语义分割 改进GANomaly 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]