Understanding the effect of transfer learning on the automatic welding defect detection

The welding defect dataset is difficult to collect due to its cost and time-consuming, which is overcome by the transfer learning method in this work. In detail, the models are initialized with the weights and biases in the pre-trained model and then compared with the model trained from scratch. The results show that the transferred weight improves the model's accuracy from 92.76% to 96.70%, but the transferred bias reduces the model's accuracy from 92.76% to 91.0%. The analysis of the feature map suggests that the transferred weight increases the variance of the feature map and improves the contrast between each channel of feature maps. As a method to validate whether the transfer learning could help detect the pixels of welding defects more accurately, the gradient class activation map (Grad-CAM) is used to track the important pixels in radiographic images for the model's predicted results. The results show that the transfer learning method enhances the model's attention to the pixels of the welding defect.