Graduate CAM The goal is to discover which part of the image caused the problem. From convolutional neural network to final decision. The method consists of creating a heat map representing the activation classes of the images received as input. Each activation class is associated with a specific output class.
These classes are used to indicate the importance of each pixel in relation to the class in question by increasing or decreasing the pixel intensity.
For example, if an image is used in a dog and cat convolutional network, Grad-CAM visualization You can generate a heatmap for the “cat” class that shows the ranges to which different parts of the image correspond to cats, and a heatmap for the “dog” class that shows to which ranges the parts of the image correspond. dog.
For example, consider CNN for dogs and cats. The Grad-CAM method generates a heatmap of the cat object class that shows the extent to which each part of the image corresponds to a cat. It also generates a heatmap for the dog object class as well.
The class activation map is created at each location (x, y ). The resulting class activation map is resampled to the size of the input image. This is shown in the heatmap below.
