A Lightweight Modified YOLOX Network Using Coordinate Attention Mechanism for PCB Surface Defect Detection

Abstract

Surface defect detection for the printed circuit board (PCB) is essential in PCB manufacturing. Existing defect detection networks have several problems: low detection efficiency, high memory consumption, and low sensitivity to small defects. To address these issues, we propose a new lightweight deep-learning-based defect detection network, YOLOX with a modified CSPDarknet and coordinate attention (YOLOX-MC-CA). YOLOX-MC-CA is developed on the YOLOX and uses the coordinate attention (CA) mechanism to improve the recognition capability of small PCB surface defects. The backbone network in YOLOX is also modified into a new CSPdarknet structure with some inverted residual blocks. The modified CSPDarknet (MC) backbone network helps the YOLOX decrease the number of parameters on the premise of guaranteeing the feature extraction ability. We evaluated the YOLOX-MC-CA with an augmented dataset based on a public PCB surface defect dataset. Compared to the squeeze-and-excitation (SE) module, convolutional block attention module (CBAM), and other approaches in previous research, the CA mechanism improves the network with more detection precision for the small PCB surface defects. The experimental results demonstrate that our network is superior to other state-of-the-art (SOTA) networks for PCB surface defect detection, scoring 99.13% on mean average precision (mAP) and 47.6 frames per second (FPS) on detection speed, only occupying a parameter space of 3.79 million (M). It demonstrates that the proposed network is more suitable for deployment on embedded systems.