Weld defects detection based on improved YOLO-V7

Abstract The automated manufacturing process frequently calls for welding operations, but welding mistakes can cause weld defects to appear. If weld defects are not promptly and accurately detected, they can cause property damage and endanger personal safety. For the first time, this study introduces a model based on the enhanced YOLO-V7 detection system to identify eight categories, which comprise seven defects categories and a good class of welding. This model addresses the existing poor accuracy of weld defects identification, and detection of fewer types of issues. Firstly, Wise-IOU is employed to swap out the original boundary frame loss function CIOU, and the quality of the anchor frame is evaluated using "outlier" instead of the original IOU, and a wise technique for gradient gain allocation is provided to enhance the overall performance of the detector. Next, the Distributed Shift Convolution (DSConv) is introduced to replace the original module to form a new ELAN-D module, which achieves lower memory usage and higher computing speed. Then, the lightweight CARAFE upsampling operator is used to replace the original model upsampling operator. Finally, the YOLO-V7 detection head is replaced with the YOLO-V6 decoupled detection head, but the part of the original model with implicit knowledge learning is retained to make its head detection more efficient. The experimental findings demonstrate that the approach achieves an F1 factor of 84.0%, a precision of 84.1%, a recall of 84.4%, and a mAP of 88.9% on the weld defects datasets, and the combination of the above aspects shows that the current enhanced model outperforms other models.