Total variation constrained graph regularized NMF for medical image registration

Abstract

Non-negative matrix factorization (NMF) is a good partsbased representation in computer vision. However, it fails to preserve or enhance the features and details of the data. To resolve this problem, we propose a novel sparse matrix factorization method for medical image registration, called Total Variation constrained Graph regularized Nonnegative Matrix Factorization (TV-GNMF). We incorporate total variation (TV) to control the speed of diffusion based on the gradient information, to preserve or enhance the features and details of the data. Manifold graph regularization is also incorporated to discover intrinsic geometric and structural information in the data. Experimental results demonstrate the effectiveness of our approach compared to state-of-the-art algorithms on medical images.