A New Graph-Based Method for Class Imbalance in Surface Defect Recognition

Abstract

Surface defect recognition is an important technology to guarantee the quality of products in modern manufacturing systems, but the class imbalance in real-world production greatly influences its performance. Recently, some deep learning-based methods have been developed to address the class imbalance. Because of the interclass similarities and the intraclass variations (ISIV) of surface defect, it is hard to apply them to solve the coupling between class imbalance and ISIV. Graph-based methods, including graph convolutional network (GCN), have the potential for ISIV. Therefore, this article proposes a new graph-based method, named anchor-based class-balanced GCN (ACB-GCN), to solve the class imbalance in surface defect recognition. First, the proposed method constructs a class-balanced graph to address the problem that excessive information from majority classes influence the performance of graph convolution. Second, the proposed method defines anchor vectors in each defect to reduce the influence of ISIV on graph construction. The experimental results on four famous data sets with class imbalance demonstrate that the proposed method can effectively address the coupling between class imbalance and ISIV, and thus extract the discriminative features. Meanwhile, the proposed method achieves better performance than traditional methods and the original GCN, especially on minority classes.