Image segmentation by unsupervised sparse clustering

In this paper, we present a novel solution for image segmentation based on positiveness which regards the segmentation as a graph-theoretic clustering problem. Contrary to spectral clustering methods using eigenvectors, the proposed method tries to find an additive combination of positive components from an originally positive data-driven matrix. By using the positiveness constraint, we obtain sparsely clustered results which do not include cancellations by negative entries. Thus, we call this method sparse clustering. The proposed method adopts a binary tree structure and solves a model selection problem by automatically determining the number of clusters using intra- and inter-cluster measures. We tested our method with image sequences as well as single frame data such as points and gray-scale, color, and texture images. Moreover, in order to objectively evaluate the performance of our method, we compared the results of the proposed method with those of the human segmentation and the Ncut method using various images including the Berkeley datasets. Experimental results show that the proposed method provides very successful and encouraging segmentations.