Improved YOLOv8-based Vision Inspection Method for Weld Defect Classification

With the rapid development of industrial manufacturing technology, there has been an increased emphasis on defect detection in steel weld seams to ensure the quality and safety of steel products. However, in the welding seam defect detection process, the foreground and background of the welding seam being inspected may have overlap, blending, blurring, and other phenomena, which give challenges to the detection of welding seam defects. To enhance the efficiency and accuracy of the algorithm for weld seam defect detection, this paper proposes the YOLOv8-SS algorithm by introducing the SE-Attention module and the SIoU loss function based on the YOLOv8 algorithm. SE-Attention module can suppress the influence of background noise on target detection and improve the accuracy of small targets. SIoU loss function is used to optimize the output predicted box boundaries, which can enhance the algorithm’s accuracy and generalization ability. The experimental results on the same weld seam defect dataset show that compared to the YOLOv8 baseline model (YOLOv8n), the improved YOLOv8-SS model has no difference in floating-point calculations and parameter quantity while achieving an improvement of $0.7 \%$ in $\mathrm{mAP} \text{@} 0.5$ and $1.3 \%$ in [email protected]:0.95. Compared with YOLOv6, the improved YOLOv8-SS model has $68.6 \%$ less of the floating-point calculations and $34.5 \%$ less of the parameters of the YOLOv6 model, with respective improvements of $17.0 \%$ in $\mathrm{mAP} \text{@} 0.5$ and $17.6 \%$ in [email protected]:0.95. This indicates that the YOLOv8-SS model enhances the accuracy of weld seam defect recognition and maintains high efficiency. Furthermore, the results of ablation experiments verify the effectiveness of adding modules for enhancing performance.