Asymmetric and Normalized Cuts for Image Clustering and Segmentation

Over the last few years spectral clustering has emerged as a powerful model for data partitioning and segmentation. Spectral clustering techniques use eigenvalues and eigenvectors of the matrix representation of a suitable graph representing the original data. In this paper a new spectral clustering method is proposed: the asymmetric cut. It allows extraction of relevant information from a dataset by making just one cut over the database. The approach is tailored to the image classification task where a given image class is to be extracted from an image database containing an unknown number of classes. The main goal of this paper is to show that the proposed technique outperforms standard spectral methods under given circumstances. The technique is compared against the conventional and well-known normalized cut algorithm