Surface defects on printed circuit board (PCB) adversely affect the product quality and the stability and reliability of equipment performance. The diversity of tiny defect types, minute and high-density layouts of PCB, and complexity of their backgrounds are significant challenges in accurately identifying surface defects, making them particularly difficult to detect using traditional methods. To address these challenges, this paper proposes a novel YOLOLW-Net based on an improved YOLOv10 framework and a designed Legendre multi-wavelet Attention (LWA) module for the effective PCB tiny defect detection. More precisely, the backbone employs the LWA mechanism to enhance texture and detail feature extraction and deepen information interaction. Then, an improved Large Separable Kernel Attention (LSKA) module is devised to enhance the model’s ability to highlight defect locations against highly similar PCB backgrounds. Thirdly, the improved Global Multi-scale Coordinate Attention (GMCA) module is implemented to replace the Pyramid Segmentation Attention (PSA) module to fully aggregate contextual semantic information within large receptive fields of PCB images. Finally, extensive experiments are conducted on two public PCB defect datasets to demonstrate that the proposed YOLOLW-Net provides significant advantages over the state-of-the-art models with a mean average precision (mAP) reaching 99.2%. Statistical significance is rigorously validated through Friedman test and Nemenyi post-hoc analysis across multiple baselines, with Wilcoxon signed-rank tests over five random seeds confirming robustness.
Zheng et al. (Fri,) studied this question.