Wide Class Activation Map for Generating Expanded Range of Activation Mapping

Abstract

In this paper, we proposed a wide CAM method for generating a wider range of activation maps than the CAM(Class Activation Map) method. The existing CAM method using the VGG-16 model extracts the activation map by applying GAP(Global Average Pooling) after the convolution 5_3 layer which extracts the highest dimension feature map. Because the CAM was applied to the lost feature maps rather than low-dimensional features such as edges. Not only does it determine based on objects in a much smaller area than the original image, it also only identifies objects that were relatively closed to each other in the image with many objects, or objects that are noticeable to humans. It cannot be determined. In order to improve this problem, the feature map of the convolution 4_3 layer, which was the previous step, was used to utilize the feature map of the lower level to utilize more feature of the heat map of the object. Experimental results show improved error rates in the classification top-1 and top-5 compared to the conventional methods.