EGCR-Net: Edge-Guided Cross-Region Network for Semi-Supervised Soldering Defect Detection

Surface defects in the printed circuit boards (PCBs) manufacturing process can affect product quality, thereby compromising the stability and reliability of the equipment. The ambiguity of defect boundaries and the insufficiency of annotated samples severely limit the detection performance of defect detection models. To address the above challenges, a semi-supervised segmentation network named EGCR-Net is proposed, which leverages a feature enhancement module and a cross-attention mechanism to enable effective multi-scale feature interaction under limited sample conditions, thereby improving the accuracy of solder defect segmentation. First, a memory module is introduced to store normal features and support anomaly comparison, thereby alleviating the problem of limited annotated samples. Second, a boundary-enhanced prompt (BEP) module is designed, which uses the Laplace operator to enhance the edge response of blurred defects. Third, a cross-region focus (CRF) module is proposed, which uses a cross-attention mechanism to capture long-distance context dependencies. The latter two modules work together to optimize the localization of defect boundaries, effectively addressing the detection challenges posed by blurred boundaries. We constructed the SolderPCB dataset, covering seven common solder defects, to evaluate real-world performance. The dataset will be publicly released after being organized. The experimental results show that EGCR-Net performs better than the current state-of-the-art methods in welding defect detection tasks and has been successfully deployed in automated PCB inspection lines to achieve quantitative detection of solder joint defects. In a real industrial environment, this method achieves an accuracy rate of 99%, effectively meeting standardized measurement requirements.