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

Yanshan Li, Huajie Liang, Hongfang Zheng, Rui Yu

Abstract: Highlights•We propose contrast-ranking class activation mapping algorithm (CR-CAM) with Inter-class Mapping Contract Block (IMCB) and Ranking Block.•By comparing features’ distances in manifold space, IMCB gradually removes irrelevant interference, obtaining a more accurate significance map.•The Ranking Block, through comparison, reduces surrounding area weight by measuring feature map distance in manifold space while considering similarity between feature maps. This makes the saliency map focus more on the core area’s characteristics.