Welding Joints Inspection via Residual Attention Network

Welding joints inspection for surface quality evaluation remains a lot of challenging work because of the difficulty in extracting suitable features. We propose a novel residual attention network for automatic inspection of the quality of welding joints. Our network has the ability to extract more useful features while maintaining a compact structure. Owing to the regularization effect of the alpha robust loss we design, our model has enough generality capabilities with limited training samples. In the end, we evaluate the performance of our network on a dataset consisting of welding joints images with score-labelled imperfections, and our proposed method achieves satisfying results in terms of welding joints inspection by predicting quality scores accurately.