PCMBA-YOLO: Pinwheel Convolution and Multi-Branch Aided FPN with Shape-IoU for Electroluminescence Defect Detection in Semiconductor Laser Chips

To ensure that laser chips meet stringent reliability standards in practical applications, comprehensive limit testing and reliability verification must be performed before deployment. This paper proposes an electroluminescence (EL) imaging-based detection method for Catastrophic Optical Mirror Damage (COMD) and Catastrophic Optical Bulk Damage (COBD). A novel model, PCMBA-YOLO, is developed on the YOLOv12 framework, integrating a Multi-Branch Aided Feature Pyramid Network (MBAFPN) and a Pinwheel Convolution (PConv) structure to enhance weak-signal feature extraction and expand the receptive field with minimal parameters. Furthermore, a Shape-IoU-based regression loss is introduced to model bounding-box shape and scale, improving localization precision and convergence. Experimental results show that PCMBA-YOLO achieves 99.4% mAP@0.5, 97.6% Precision, and 98.7% Recall, with a 14% reduction in parameters compared to the baseline. The proposed method demonstrates superior accuracy, efficiency, and robust generalization, providing a high-performance solution for automated visual inspection in semiconductor manufacturing.