We weakly train a CNN classifier with the rico dataset on existence of icons in mobile app screenshots.
We then compute the class activation mapping by implementing the Grad-CAM algorithm of the input test images to figure out the input pixels responsible for the activations of each output neuron.
After that we perform OTSU thresholding to get prediction candidates and blob analysis to choose the most likely correct prediction.
Finally we iteratively perform activation maximization to see what the trained network has learned. Example outputs for the arrow and menu classes.
