Content-Based Expansion for Image Matching

Abstract

This paper challenges the two popular approaches used in image matching: the gradient approach and the phase approach. We demonstrate from both theoretical and experimental points of view that the gradient approach has large model-incurred errors when the images contain significant high frequency information, and the phase approach has the same kind of errors when the images contain mainly low frequency information. Both are due to inappropriate signal expansion models. Based on such a unified perspective on those approaches, we propose a content-based expansion scheme, in which the exact form of the signal expansion depends on the content of band-passed signals. We show that such a scheme eliminates the risk of large model-incurred errors. Finally, we implement the content-based approach based on FFT (Fast Fourier Transform) and compare it with both gradient and phase approaches in performance.