PCBs play a critical role in electronic manufacturing, and accurate defect detection is essential for ensuring product quality and reliability. However, PCB defects are often small, irregularly shaped, and embedded in complex textures, making them difficult to detect using traditional methods. In this paper, we propose CM-UNetv2, a semantic segmentation network designed to address these challenges through three architectural modules incorporating four key innovations. First, a Parallelized Patch-Aware Attention (PPA) module is incorporated into the encoder to enhance multi-scale feature representation through a multi-branch attention mechanism combining local, global, and serial convolutions. Second, we propose a Dual-Stream Skip Guidance (DSSG) module that decouples semantic refinement from spatial information preservation via two separate skip pathways, enabling finer detail retention. Third, we design a decoder module called Frequency-domain Guided Context Mamba (FGCMamba), which integrates two novel mechanisms: a Spatial Guidance Cross-Attention (SGCA) mechanism to enhance the alignment of spatial and semantic features, and a Frequency-domain Self-Attention Solver (FSAS) to compute global attention efficiently in the frequency domain, improving boundary restoration and reducing computational overhead. Experiments on the MeiweiPCB and KWSD2 datasets demonstrate that the CM-UNetv2 achieves state-of-the-art performance in small object detection, boundary accuracy, and overall segmentation robustness.
Guo et al. (Sat,) studied this question.