Binary classification of welding defect based on deep learning

The current study investigated the effects of data augmentation, convolutional layers depth, and learning rate on inspection accuracy of a deep learning model which detects welding defects. The experimental results suggested that simultaneous use of these two methods improved the model's performance more than the sum of using each of the two methods alone. As the number of convolutional layers increases above 10, the network cannot extract defect features effectively, and the model's accuracy will decrease. Learning rate could significantly influence convergence rate and accuracy, and using an initial learning rate of 1e-4 and decaying it at epoch 250 could reduce the loss function and increase model accuracy.