Product Defect Detection Based on Multi-Class Classification

Takumi Meguro, Shota Nakada, Qiangfu Zhao

Abstract: In anomaly detection tasks, missing anomalies is a problem that must be avoided as far as possible. It has very serious consequences for companies that manufacture products if products with anomalies are delivered to the users. On the other hand, we should also minimize the over-detection rate while keeping the number of missed anomalies to zero or almost zero. To achieve these objectives, we propose an image-based multi-class classification method that can not only detect but also localize the anomalies. Since the proposed method can pay more attention to the anomalies, images of the normal parts will provide a less negative effect on the performance. Interestingly, by setting a proper threshold for one of the outputs, we can separate normal and abnormal data more clearly compared with the straightforward two-class approach. The effectiveness of the proposed method is verified using datasets provided by the partner company.