Multi-size Scaled CAM for More Accurate Visual Interpretation of CNNs.

The search for decision bases for image classification neural networks is one of the popular research directions in deep neural network interpretability. These studies highlight regions of interest to image classification models by generating a saliency map that assign contribution values to each pixel in the input image. However, these current methods cannot accurately locate the key features of the target object and tend to include other irrelevant objects in the salient region, resulting in unreliable saliency maps. In addition, the problems of noise and low resolution have also plagued the researchers. To address the above issues, we propose a saliency map generation method based on multi-size scaling of the input image for the CNN-based model. The method scales the input images in multiple sizes and extracts different resolution feature maps and their corresponding gradients from specific convolutional layers, and fuses them as masks of the input images. Then the masked input images are input to the model separately to obtain the weights of each mask, and finally the masks are combined with linear weighting to obtain the saliency map. Experiments show that our method can produce more detailed saliency maps and accurately target regions of interest to the model. It has significant advantages and higher application value than the current CAM-based (class activation mapping) methods.