A novel image registration method is proposed based on non-negative matrix factorization by using the absolute feature vectors of the spectrum of graph. First, the non-negative weight matrices of the undirected weighted graph by the feature points of the images are constructed, from which the feature basis image obtained can contain the global features of the original images using non-negative matrix factorization. The initial value of non-negative matrix factorization is evaluated iteratively using the absolute feature vectors of the spectrum of the non-negative weight matrices, which can reflect the structural characteristics information of the graph, and improve the image matching rate. Finally, the correct feature correspondences are found in the feature basis vectors space between the weighted graphs. The experiments on both synthetic images and real-word images indicate that the proposed method is feasible for image registration.