On the effect of the attention mechanism for automatic welding defects detection based on deep learning

Attention mechanism has been widely used deep learning applications for automatic welding defect detection. Literature suggested that the attention mechanism slightly improved defect detection accuracy. In most cases, it was used along with other strategies, such as transfer learning and data augmentation. However, the solo effect of the attention mechanism on the automatic welding defects detection has not been thoroughly examined. Therefore, this study considers two attention mechanisms, including channel attention mechanism and spatial attention mechanism, into the basis of binary classification network to analyze and compare their effect. The analysis is conducted from three aspects: (i) visualizing and quantifying the extracted feature, (ii) tracking the salient pixels of welding defects, and (iii) comparing the clusters of defective and non-defective features. The results suggest the spatial attention mechanism improves the information entropy of extracted features, enhance the model to focus on the salient pixels of welding defects, and prompt the separation of the defective and non-defective features clusters.