Score-CAMpp: class activation map based on logarithmic transformation

Score-CAM is the first algorithm based on gradient-free model, which can provide convincing explanation for the decision made by neural network. Score-CAM use the confidence score to express the weight of activation maps. Compared with the gradient-based algorithm, Score-CAM can not only accurately locate a single object, but also show better performance when locating multiple similar objects. However, Score-CAM not only contains a lot of non-target information, but also the feature map has poor selectivity to feature areas. Therefore, we propose a Score-CAM m based on logarithmic transformation, namely Score-CAMpp. Firstly, the deep features of images are extracted by the deep neural network. And we need to process the extracted feature map by logarithmic transformation to enhance the effective area of the feature map. Then, the transformed feature map is used as a mask, multiplied by the input image and sent to the network to obtain the weight. Finally, the weight and the transformed feature map are linearly weighted and summed. The experimental results show that the proposed algorithm is superior to Score-CAM in visualization and data on evaluation standard.