Weld Surface Defect Detection Based on Improved YOLOv7

Welding is the most economical and effective permanent metal connection method. However, the surface defects caused by various factors, is the quality of welding products cannot be fully guaranteed. In order to improve the efficiency and accuracy of defect detection, this paper proposes a new weld defect detection algorithm based on YOLOv7. Firstly, the DCGAN model is used to enhance the data set of welding defects collected in the industrial field. Secondly, the Repvgg model architecture is analyzed, and the residual branch and 1 × 1 convolution architecture are added between each module of the high-efficiency layer attention network, and the CBAM attention module is integrated. Finally, Focal-EIoU is used to replace CIoU in the original YOLOv7 network model to optimize the loss function, which accelerates convergence, improves regression accuracy and network robustness. The experimental results show that the improved YOLOv7 network model has the highest average detection accuracy and the lowest model convergence compared with the original network and the classical target detection network model.