TOFD Image Features Recognition Based on Improved YOLOv8

TOFD detection is critical to weld quality. TOFD image evaluation is highly labour-intensive, with low efficiency and unstable quality. In order to realize the intelligent recognition of TOFD image features, this study first introduced a self-adaptive TOFD image enhancement processing method to achieve noise reduction and feature contrast enhancement. Then, by analyzing the characteristics of TOFD graph, this study has improved the target detection algorithm of YOLO v8 network structure with the reduction of three detection heads to two and reduced the number of convolution processing. Recognition models of TOFD image boundary characteristics (latter wave, LW; bottom wave, BW) and five types of defects (Pore, Cluster Porosity, Slag, Slags and Crack) are established respectively. The verification results have shown that based on the proposed self-adaptive TOFD image enhancement processing method and the improved YOLOv8 model, the accuracy of boundary detection has reached 100%, and the accuracy of defect detection has exceeded 97%, among which the detection accuracy of the most harmful defect slags and crack is more than 90%.