Image Clustering Based on Graph Regularized Robust Principal Component Analysis

Abstract

Image clustering has become one of the most popular themes in web based recommendation system. In this study, we propose a novel image clustering algorithm referred as graph regularized robust principal component analysis (GRPCA). Unlike existing spectral rotation or k-means method, no discretization step is required in our proposed method by imposing nonnegative constraint explicitly. Besides, in GRPCA an affinity graph is constructed to encode the locality manifold information, and the global graph structure is respected by applying matrix factorization. The proposed method is robust to model selection that is more appealing for real unsupervised applications. Extensive experiments on three publicly available image datasets demonstrate the effectiveness of our algorithm.