Abstract: Defect detection in large-scale manufacturing aims to find defective components,such as damaged,misaligned components,and components with printing errors.Due to unknown defect types and shortage of defect samples,industrial defect detection faces great challenges.To overcome the above difficulties,some methods utilize common visual representations from natural image datasets to extract generalized features for defect detection.However,there are distribution differences between the extracted pre-trained features and the target data.Using this feature directly will lead to poor detection performance.Therefore,ConPatch,a method based on contrastive representation learning is proposed.This method employs contrastive representation lear-ning to collect similar features or separate dissimilar features,resulting in goal-oriented representations of features.In order to solve the problem of lack of defect annotation,two similarity measures in data representations,pairwise similarity and global similarity,are used as pseudo labels.In addition,the method uses a lightweight memory bank and only stores the feature centers of all normal sample which are all defect-free sample in the memory bank,reducing the space complexity and the size of the memory bank.Finally,the normal features are brought closer to a hypersphere and the defect features are distributed outside the hypersphere to gather the normal features.Experimental results show that the I-AUROC and P-AUROC of the ConPatch model based on Wide-ResNet50 reaches 99.35% and 98.26% respectively in the industrial defect detection dataset MVTec AD.In the VisA dataset,I-AUROC and P-AUROC reaches 95.50% and 98.21%,respectively.The above results verify the effectiveness of the proposed model.

Key words: Industrial product defect detection, Contrastive representation learning, Similarity measure, Memory bank, Hype-rsphere