Abstract
Welding technology is an important link in shipbuilding, and ultrasonic diffraction time difference method is one of its important inspection means, but the identification of its TOFD image still mainly relies on manual evaluation. In this paper, in order to solve the problems of low accuracy and unbalanced categories of defect recognition in the existing automatic identification research, by reclassifying defects into primary defects, secondary defects, and tertiary defects according to the degree of harm to the weld, and constituting a TOFD image dataset containing 934 images, and then using the YOLOv7 model, by adding mixup enhancement combined with mosaic enhancement, the The mAP value of the improved YOLOv7 model was improved by about 42% to 93.41% compared with the original model, and the AP0.5 values of the YOLOv7 improved model for primary defects, secondary defects, and tertiary defects were improved by about 30%, 38%, and 58% compared with the original model, respectively, by introducing the CBAM attention mechanism method to optimize the problems of insufficient data sample size and the imbalance between foreground and background; the AP 0.75 values improved by about 16%, 30%, and 45%, respectively, compared with the original model; its detection speed was 7.8 sheets per second.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have