A Class-Incremental Learning Method for PCB Defect Detection

Abstract

Defect detection of printed circuit boards (PCBs), as a critical step in the manufacturing process, has achieved significant improvement with the help of deep learning techniques. However, existing research has focused only on the closed static detection scenario. This study aims to transfer the PCB defect detection task to the more practical incremental detection scenario. First, to cope with the performance requirements of industrial quality inspection, this article proposes a PCB-YOLOX detector for PCB defect detection by optimizing based on YOLOX-S. Specifically, a feature enhancement module (FEM) is designed to improve the feature representation of the model for small targets of defects, while an attention feature fusion module (AFFM) is designed to facilitate the efficient fusion of features at different scales. Then, the PCB-YOLOX is combined with an incremental learning method, elastic response distillation (ERD), to propose a class-incremental PCB defect detection method. Experimental results in the static detection scenario show that PCB-YOLOX achieves competitive performance in terms of detection accuracy compared to several state-of-the-art detectors, with 96.5% (mAP0.5) and 51.9% (mAPs), respectively. The model parameters, detection speed, model size, and computation of PCB-YOLOX are 12.8 M, 50.5 frames/s, 49.1 M, and 35.6 G, respectively, which can meet the needs of industrial inspection. In addition, the experimental results in the incremental detection scenario show that the method proposed in this article can effectively alleviate the catastrophic forgetting problem in the incremental learning process.