Constrained Dual Graph Regularized NMF for Image Clustering

Abstract

Non-negative matrix factorization (NMF) becomes an important dimension reduction and feature extraction tool in the fields of scientific computing and computer vision. In this paper, for using the known label information in the original data, we put forward a semi-supervised NMF algorithm called constrained dual graph regularized non-negative matrix factorization (CDNMF). The new algorithm employs hard constraints to retain the priori label information of samples, constructs two association graphs to encode the geometric structures of the data manifold and the feature manifold, and incorporates the additional bi-orthogonal constraints to improve the identification ability of data in the new representation space. We have also developed an iterative optimization strategy for CDNMF and proved its convergence. Finally the clustering experiments on five standard image data sets show the effectiveness of the proposed algorithm.