Self-Supervised Learning with Min-Max Distance Approach for Few-Shot Steel Surface Defects Recognition
作者单位：School of Electronic and Information Engineering Wuyi University School of Mechanical and Automation Engineering Wuyi University Technical DepartmentGuangdong Province Kejie Machinery Automation Company
会议名称：《第43届中国控制会议》
会议日期：1000年
学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 081104[工学-模式识别与智能系统] 08[工学] 080502[工学-材料学] 0805[工学-材料科学与工程（可授工学、理学学位）] 0835[工学-软件工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）]
关 键 词：deep learning self-supervised learning L1-norm
摘 要：Surface defects in steel are crucial for quality inspection in the manufacturing of industrial metal *** deep learning methods have been extensively studied in this field, the lack of large-scale annotated defect samples imposes significant labeling pressure, making training fully supervised deep models challenging. In response to this issue, we employ self-supervised learning for few-shot recognition,effectively learning representations from limited unlabeled data and achieving defect recognition with a small amount of labeled samples. In particular, we construct Other encoder with multiview information and Base encoder with single view information based on the distillation self-supervised learning framework through the multi-crops mechanism, and the Other encoders and Base encoders perform information share. Our main contribution is that, based on this framework, we propose a new regularization method to solve the problem of instances deviating from the center of self-supervised clustering by minimizing the maximum instance distance(MMD). In testing,fine-tuning the MMD pre-trained model with a small number of labeled data achieves superior recognition performance on the steel surface defect dataset NEU, significantly outperforming other testing methods.