Proposal for improving SimCLR using image synthesis for defect recognition tasks

This paper proposes an improvement of SimCLR for defect recognition tasks by image synthesis using weighted averages. There are studies on applying contrastive learning to defect detection in industrial products. This is because the number of defective products is quite small compared to non-defective products, and contrastive learning is a method that allows you to train a model with a small dataset by augmenting images and comparing them. However, problems with random trimming have been reported for the combination of defect detection and contrastive learning. Since defect images consist of defect areas and non-defect areas, augmentation by random cropping does not work well. To solve this problem, this study proposes the addition of image synthesis using weighted averaging to the conventional SimCLR’s augmentation method. The proposed method avoids wasteful learning that attracts feature vectors between cropped defect and non-defect areas. In the experiment, a CNN was trained on a small dataset of 32 images, and our proposed method improved AUC by 15% compared to the conventional method.