High-Precision Welding Defect Detection Practice: An Innovative YOLOv12 Model with Pinwheel Convolution and Adaptive Attention

This research introduces an innovative approach to enhance the accuracy and speed of weld defect detection by integrating attention mechanisms and Pinwheel-Shaped Convolution (PConv) into the YOLOv12 framework. Addressing the limitations of traditional CNNs and transformer-based models, which often struggle with either missing subtle defects or incurring high computational costs, this study achieves precise, real-time detection, outperforming existing solutions in complex industrial environments. A systematic methodology was adopted to train and validate the proposed model on four benchmark datasets, employing diverse data augmentation techniques such as Mosaic, Mixup, and Copy-Paste to improve generalization. Quantitative performance evaluation was conducted using established metrics, including recall, F1-score, and mean Average Precision (mAP), enabling rigorous comparison with baseline YOLO models and attention-based architectures under realistic industrial inspection conditions. The innovative PConv based YOLOv12 model demonstrated outstanding performance in weld defect detection. It achieved an F1-score of 0.941 and mAP@0.5 of 0.989 in single-class detection, and an F1-score of 0.848 with 0.887 recall in multi-class detection. Mosaic augmentation boosted mAP@0.5:0.95 to 0.854, enhancing generalization. The model converged rapidly, reaching mAP@0.5 of 0.905 in just 10 epochs and stabilizing near 0.996, proving its robustness and suitability for industrial real-time applications.