Abstract To address the challenges of scale inconsistency in defect detection, the presence of minor target defects, external environmental interference, and other factors affecting the detection of welded surface defects in marine T-profiles, as well as to meet the practical deployment requirements of detection models, a novel algorithm for welding surface defect detection based on YOLOv8 model is proposed. This algorithm offers enhanced detection accuracy, faster processing speeds, and lowered computational requirements, providing a more effective solution for production workshops. Firstly, the loss function is modified from CIoU to MPDIoU, effectively reducing missed detections caused by overlapping defects. Secondly, this study introduces an innovative detection head structure, termed the Co-Detail-Enhanced Convolution Detection Head (CoDECD). By leveraging parameter sharing and detail-enhanced convolution, CoDECD reduces the complexity and computational demands of the original model, whilst concomitantly achieving a substantial enhancement in the detection performance of small-target defects. Finally, the incorporation of the Vision Transformer with Deformable Attention (DAT) has been demonstrated to enhance the efficiency and precision of the system in the execution of complex visual tasks, enhancing the model’s detection accuracy even further. The test findings demonstrate that the improved model achieves a mAP@50 of 89.1% in detecting defects on the welded surfaces of T-profiles, representing a 3.5% increase compared to the YOLOv8 model. Furthermore, the GFLOPs and parameter count of the model are reduced by 24.1% and 24.9%, respectively.
Zhou et al. (Fri,) studied this question.