中圖分類(lèi)號(hào)： TP183
文獻(xiàn)標(biāo)識(shí)碼： A
DOI： 10.19358/j.issn.2096-5133.2022.01.010
引用格式： 楊繼增，關(guān)勝曉. 面向CNN的類(lèi)激活映射算法研究[J].信息技術(shù)與網(wǎng)絡(luò)安全，2022，41(1)：63-68.

A class activation mapping algorithm for CNN

Abstract： Class activation mapping(CAM) is a straightforward method in the interpretation of convolutional neural networks(CNN), usually generated by the last convolution layer of CNN, which can highlight different object regions of the target class. Several previous CAM methods only depend on the final convolution layer, so the class activation map, generated by these, can only show the rough object position information. A new method called score-weighted & layer-wise class activation mapping(SL-CAM) was proposed to generate class activation maps by merging shallow to deep information of CNN. The class activation map generated from the shallow feature map and its corresponding gradient contains detailed and accurate location information with more noise. The activation map generated from the deep feature map contains less noise with rough location information. SL-CAM outperforms Grad-CAM, Grad-CAM++ and Score-CAM on ILSVRC2012 Val.

Key words : class activation mapping；convolutional neural networks；visualization