Optimization of C2PSA Module in YOLOv11 Architecture for Semiconductor Wafer Inspection

Conventional YOLO-based methods for wafer defect detection increasingly struggle to satisfy the precision and efficiency demands of advanced semiconductor manufacturing processes. To tackle this challenge, this study proposes a refined YOLOv11 architecture integrated with an enhanced C2PSA module, aiming to simultaneously improve detection accuracy and inference speed. Specifically, the key innovations are as follows: replacing standard normalization layers with DynamicTan (Dyt) to significantly enhance training stability and efficiency, incorporating Token Statistics Self-Attention (TSSA) to reduce computational complexity from quadratic to linear while strengthening the modeling of long-range dependencies, implementing a Multi-cognitive Visual Adapter (Mona) to optimize feature distributions and thus improve visual signal processing performance, and augmenting the C2PSA module with a Spatially-Enhanced Feed-forward Network (SEFN) to refine the modeling of local spatial details. Experimental validation on the WM-811k benchmark dataset shows that, compared with the baseline YOLOv11n, the proposed architecture achieves respective improvements of 4.28% in mAP@0.5 and 7.79% in recall rate. These results confirm the superior generalization capability of the proposed method and validate the effectiveness of the designed enhancements for industrial-scale semiconductor inspection systems.