Unsupervised Segmentation of Images using CNN

摘要

Image segmentation literally means to classify each object or element in an image to different classes. Thus, we must cluster each pixels of the image to the appropriate class it should belong to. The segmentation of image is accomplished through unsupervised learning. Therefore, any training data or labels is not provided beforehand. A Convolutional Neural Network is used to assign each pixel in the input image to the appropriate cluster. The present system is considered as a Parent-Child track. Initially the Parent track deals with prediction of cluster labels with initialized network parameters. The Child track refines the cluster and calculates the SoftMax loss between predicted and refined labels. It is propagated back to parent track to adjust the network parameters such that the loss will decrease with each iteration.