Truncated and integrated class activation maps for weakly supervised defect detection

Abstract Deep learning is now widely used for detecting surface defects, which is crucial for automated quality control in industries. However, getting lots of accurate labeled data is tough, and this slows down the progress of smart defect detection. To tackle this, we suggest using weakly supervised semantic segmentation (WSSS) methods, especially those based on class activation maps (CAMs). One issue with CAM is that the feature maps from the last layer of the neural network do not have high enough resolution. We want to create feature maps with more detail that can give us better semantic information. We take a new look at the semantic information in the early feature maps, finding that they have fine details but also mix in a lot of noise that is not relevant to our target. To fix this, we propose a simple way to reduce noise by cutting off positive gradients. This idea can be added to other CAM methods to help them get better CAMs. A large number of WSSS experiments were conducted on defect detection datasets. The results from these experiments consistently show that our method is effective for finding defects.