PCB Defect Detection Based on Improved YOLOX Algorithm

Conventional object detection algorithms have low accuracy when detecting defects on printed circuit boards (PCBs) with small targets and complex backgrounds, making it difficult to detect defects accurately. This paper proposes a more suitable algorithm for detecting surface defects on printed circuit board so as to overcome this problem, called YOLOX-CAS. The method is based on the YOLOX network and introduces the convolutional block attention module (CBAM) at the neck of the model to increase the attention of the model to printed circuit board surface defects. In addition, the adaptive spatial feature fusion (ASFF) module is introduced at the end of the neck to reduce background interference and improve foreground detection accuracy. Finally, the slicing aided hyper inference (SAHI) framework is integrated into the model, which enhances the ability to detect small objects. Experiment results demonstrate that the proposed YOLOX-CAS algorithm outperforms YOLOX-s by 4.1% in terms of the (mAP), thereby improving the accuracy. It is evidence that the YOLOX-CAS algorithm presents promising application potential in industrial production environments.