Improving Defect Classification Accuracy by Contrastive Learning Using Mask Data

This paper proposes a method to apply contrastive learning to defect recognition using mask data.In contrastive learning for defect recognition, learning by random cropping or similarity comparison does not work well because defect images consist of defective and non-defective parts.To solve this problem, the proposed method effectively uses supervised mask data.By using mask data, it is possible to know which classes are included in the area obtained by random cropping, and therefore it is possible to define a loss function for appropriate contrastive learning.In experiments, the effectiveness of the proposed method is verified using magnetic tile data from the CDS2K dataset.The conventional contrastive learning method and the proposed method were trained on 12 training datasets and compared, and the proposed method achieved higher classification accuracy in all results.