CR-CAM: Generating explanations for deep neural networks by contrasting and ranking features

Abstract

The class activation mapping (CAM) algorithm is a visual interpretation algorithm that identifies the most discriminative regions for the target class in a classification task. However, existing CAM algorithms do not consider the differences between different categories and regions that are irrelevant to the target class during the feature extraction process. This results in interference from similar categories and irrelevant regions on the edges of the target class, leading to distorted saliency maps.To address this issue, we propose the Contrast-Ranking Class Activation Mapping (CR-CAM), including Inter-class Mapping Contract Block (IMCB) and Ranking Block. To gradually eliminate the interference regions that are irrelevant to the target class, IMCB is designed to compare distances between features in manifold space in order to generate more accurate saliency maps. To account for the similarity between different categories, ranking blocks adopt a comparative approach to measure the distances of feature mappings in the manifold space, thereby reducing the weights of surrounding regions. CR-CAM can be used on both CNN and GCN without modifying the algorithm to generate the class activation map. Experiments on both ImageNet and NTU RGB+D 60 datasets show highly competitive performance. In particular, CR-CAM outperforms alternative approaches by an average decrease of 0.0811 in Ave Drop, while exhibiting an average improvement of 0.011 in Ave Inc for CNN.