Image Segmentation using Sparse Subspace Clustering

Image segmentation is one of the fundamental problems in computer vision. Machine learning approaches to solving the problem of partitioning pixels into disjoint and meaningful subsets is a major area of research. Both supervised and unsupervised techniques have been applied on this pattern classification problem. Among supervised learning, sparse dictionary learning is a relatively new technique that has given good performance on image segmentation problem. Training a dictionary on each class of features extracted from the images has resulted in creating an overcomplete basis that spans the subspace to which the particular class feature set belongs. We are investigating to address the problem in the unsupervised setting, where prior information on the class features are unavailable. We address the problem by classifying the features based on the subspace to which they belong. This is done by graph-cut segmentation by defining similarity matrix based on subspace clustering. The results obtained are encouraging and strongly indicates the fact that higher dimensional features may be classified more accurately by utilizing subspace characteristics than by using functions of distance measures.