Few-Shot Steel Surface Defect Recognition via Self-Supervised Teacher-Student Model with Min-Max Instances Similarity

Steel surface defect (SSD) recognition plays a critical role in guaranteeing the quality and economic benefits of industrial metal products. At present, deep-learning-based recognition methods have achieved excellent performance within related fields. However, these methods usually depend on the availability of extensive training data and label information. For SSD recognition, the SSD data are often insufficient and labels are scarce, making it hard to train a supervised deep model. In this article, we proposed a novel few-shot SSD recognition method based on self-supervised contrastive learning (self-SCL), which can effectively learn data representation with unlabeled data in the pretraining stage and learn categories from a few labeled samples in the fine-tuning stage. In this method, we proposed an information-guided teacher–student (TS) pretrained model with min–max instances similarity (MMIS) regularization. In the pretraining stage, we utilized the multicrop mechanism to build a teacher encoder (TE) with multiview information and leveraged the TE to guide the student encoder (SE) with information. Meanwhile, the SE also provided the TE with “questions” during the update process, which the SE also shared updated information with the TE for further refinement. Specifically, we introduced MMIS to address the problem of ambiguity between intraclass and interclass distances in SSD data. During the fine-tuning phase, we fine-tuned the pretrained model obtained by teacher–student with min–max instance similarity (TSMMIS) with a few labeled data and tested the SSD recognition accuracy in four SSD datasets. As the experimental results indicate, our method significantly outperforms other state-of-the-art methods on the four SSD datasets.