Truncated and Integrated Class Activation Maps for Weakly Supervised Defect Detection

Abstract Deep learning is progressively being utilized on a vast scale for surface defect
 detection, which is pivotal for the implementation of automated industrial quality
 control. However, the acquisition of large-scale, high-precision labeled data is
 challenging, significantly impeding the intelligent advancement of defect detection
 tasks. In response to this challenge, this work proposes the use of weakly supervised
 semantic segmentation methods, enhancing key technologies based on Class Activation
 Maps (CAM). A primary limitation to the performance of class activation maps is the
 low spatial resolution of the feature maps in the last layer of the convolutional neural
 network. Therefore, there is an expectation to generate high-resolution feature maps
 that yield high-quality semantic information. This paper reconsiders the properties
 of semantic information in shallow feature maps, discovering that shallow feature
 maps still contain fine-grained, non-discriminative features while mixing a considerable
 amount of non-target noise. Furthermore, a simple gradient-based denoising method is
 proposed to filter the noise by truncating the positive gradient. The proposed scheme
 can be easily integrated into other CAM-related methods, facilitating these methods
 to obtain higher-quality class activation maps. To substantiate the effectiveness
 of our approach, a series of experiments were conducted within the context of a
 weakly-supervised semantic segmentation task, with a particular emphasis on defect
 segmentation. The results of these comprehensive experiments consistently validate
 the effectiveness of our approach for defect detection tasks.