Image Registration Using Isometric Projection of Graph

This paper proposes a novel graph spectral method for image registration based on isometric projection, which utilizes purely structure information of the graph to solve a graph matching problem. We demonstrate how two graphs can be compared exactly and robustly through their isometries. To do this, firstly, a new projection method is given, by which the nodes of two graphs are projected to the normalized rows of critical point of a matrix function. Secondly, the uniqueness of the projections is proved by a theorem. In this way, correspondences can be found in a unit sphere. We attempt the utility of the method on a number of synthetic and real-world images.