基于正样本的产品表面缺陷视觉检测方法  

A Positive Sample-Based Visual Inspection Method of Product Surface Defects

作　　者：左旺 陈仲生 李潮林 柳浩鹏 ZUO Wang;CHEN Zhongsheng;LI Chaolin;LIU Haopeng(College of Electrical and Information Engineering,Hunan University of Technology,Zhuzhou Hunan 412007,China)

机构地区：[1]湖南工业大学电气与信息工程学院,湖南株洲412007

出　　处：《湖南工业大学学报》2023年第1期69-74,共6页Journal of Hunan University of Technology

基　　金：国家自然科学基金资助项目(51975206)。

摘　　要：表面缺陷视觉检测是产品质检重要环节之一,而实际工业应用中,往往存在表面缺陷样本少、需要大量标注样本等问题。为此,将注意力机制引入Resnet50网络,提出一种基于正样本的产品表面缺陷视觉检测方法。首先,利用预训练网络Resnet50_CBAM学习到包含来自不同语义层和分辨率信息的嵌入向量,并利用多元高斯参数表示图片正常特征;其次,将缺陷图像输入到预训练网络Resnet50_CBAM,获得相应的嵌入向量和多元高斯参数;最后,采用马氏距离计算整张缺陷图像所有像素点的缺陷分数图,实现基于像素级的缺陷区域定位。实验数据集验证结果表明,与已有方法相比,所提方法需要的正常样本更少且检测精度更高,从而可以有效解决少样本的产品缺陷视觉检测难题。Visual detection of surface defects is one of the important aspects of product quality inspection.However,in practical industrial applications,in view of such flaws as few surface defect samples and the need for a large number of labeled samples,attention mechanism is introduced into Resnet50 network,with a positive sample-based visual inspection method proposed for product surface defects.First,the pre-training network Resnet50_CBAM is adopted to acquire the embedding vectors containing information from different semantic layers and resolutions,with the multivariate Gaussian parameters used for a representation of the normal features of the images.Next,the defective images are input to the pre-training network Resnet50_CBAM,thus obtaining the corresponding embedding vectors and multivariate Gaussian parameters.Finally,the Markov distance is used to work out the defective scores of all pixels in the whole defective image,so as to realize the defect area location based on pixel level.The experimental data set verification results show that compared with the existing methods,this proposed method is characterized with a low requirement of normal samples yet with a higher detection accuracy,which can effectively solve the problem of visual inspection of product defects with fewer samples.

关 键 词：正样本 少样本 马氏距离 CBAM 预训练网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]