Quality inspection of steel plate welding is critical in industrial manufacturing. However, weld proximity defects often present diverse morphologies, overlapping regions, and dense distributions, posing challenges to accurate industrial defect inspection. Therefore, we propose an industrial detector based on the EP-YOLOv7. First, an Efficient Multiscale Channel Attention (EMCA) is introduced to strengthen multi-scale feature perception and improve the model?s focus on weld proximity defects. Second, the EMCA module is integrated into the Efficient Layer Aggregation Network to enhance feature fusion and defect representation. Finally, a Partial-Bottleneck Decoupling Predictor Head (P-BD Head) is designed to significantly improve localization accuracy and reduce missed detections of small targets. Experimental evaluations on both a self-built weld proximity defect dataset and a public generalization dataset show that EP-YOLOv7 achieves mAP of 85.2%/56.2% and F1 scores of 80.3%/43.3%. Meanwhile, the model size increases by only 0.6 MB (total 37.9 MB), demonstrating that the proposed approach delivers substantial accuracy gains while maintaining lightweight computational complexity, suitable for practical industrial applications.
Zhang et al. (Thu,) studied this question.