Digital image registration by sequential analysis

Abstract This paper proposes a new approach to translational image registration problems, based on the theory of sequential tests of hypotheses (S.T.H.). This leads to the development of two different methods: the first one is based on the Gaussian assumption and uses the fact that the variance of the error between two images to be registered tends to be low on the registration point. The second method uses binary images derived from the original ones. The statistical model for the resulting accumulated error is a binomial distribution and the registration position is characterized by a low probability of the binary error being one. In both methods two sequences of thresholds are employed: one leading to the rejection of the point and the other one to the eventual acceptance of it. Experimental results with both methods are presented. They include registration of a LANDSAT image against noisy versions of it, matching of different channels of the same multispectral image as well as matching of segments of two images taken at different dates. Successful registration was achieved in most cases even in low signal to noise ratio conditions, with a modest amount of computational effort.