Graph Regularized Non-Negative Matrix Factorization for Image Retrieval

TL;DR: This work encoded the geometrical information contained in the data by constructing a nearest neighbor graph and perform matrix factorization based on this graph structure and compared NMF and GNMF based method in the context of image retrieval.

Abstract: This paper presents a novel matrix factorization method for effectively performing the image retrieval in large image databases Non-negative Matrix Factorization (NMF) provides a parts-based representation of data by finding two non-negative matrices whose product can well approximate the original data matrix Although it has been applied successfully for several applications, simply using NMF for image retrieval provides low performance results which results from the fact that NMF fails to consider the geometric structure that is contained within the data To solve the above problem, we encode the geometrical information contained in the data by constructing a nearest neighbor graph and perform matrix factorization based on this graph structure In this work, 500 images from Corel dataset have been taken into consideration We compared NMF and GNMF based method in the context of image retrieval Experimental results demonstrate the effectiveness and robustness of GNMF based approach