A Novel Approach for Visualization of Class Activation Maps with Reduced Dimensions

Abstract

Explaining how deep neural networks work is a new and challenging area for computer vision projects. The deep learning models are seen as Black-Box models because of the number of hidden layers, neurons, and activation functions. Class Activation Map (CAM) is a method that allows highlighting the most important features utilizing the last convolution layer. Since the last convolutional layer has lots of filters, it causes to create unfocused CAM outputs. Applying the Principal Component Analysis method to the filters for the purpose of uncovering the most important filters the filter size is reduced from 512 to 10 in this study. The results show that when the reduced filters are used, more focused CAMs are obtained. These maps can be used for weakly-supervised applications such as object detection and image segmentation. © 2022 IEEE.