Fig. 1. Camera and robot setup

Fig. 2. ResNet-18 model

Fig. 3. Convolutional layer

Fig. 4. Convolutions dynamics within CNN

Fig. 5. Pooling layer

Fig. 6. Six categories of images. (a) Good weld seam; (b) burn through; (c) contamination; (d) lack of fusion; (e) misalignment; (f) lack of penetration

Fig. 7. Data augmentation. (a) Origin; (b) rotation; (c) random horizontal; (d) random crop; (e) random brightness; (f) contrast adjustment

Fig. 8. Curves of accuracy and loss. (a) Accuracy curves; (b) loss curves

Fig. 9. Confusion matrix

Fig. 10. 2D spatial distribution of deep features. (a) Combination of Softmax loss and center loss; (b) Softmax loss only

Fig. 11. Feature maps of ResNet features from different layers. (a) Input image; (b) convex function optimization layer; (c) layer 1; (d) final layer

Fig. 12. Process of obtaining guided grad-CAM

Fig. 13. Example of guided grad-CAM for welding image. (a) Input; (b) CAM; (c) guided grad-CAM