Images can be segmented using a pretrained CNN classification network. It is not a good idea for segmentation, but it is a very effective tool for visualization of the classifier.
The logic behind this is very simple, Gradient in the image space signifies the contribution of each pixel toward prediction and this information can be used for image segmentation.
https://medium.com/@negidhirendersingh26/image-segmentation-using-classifier-82f2085f5bc6
Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
The Apache License 2.0. Please see the license file for more informat