TOFD Defect Image Recognition Based on YOLOv7

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