<title>Maximum Likelihood Estimation For Image Registration</title>

Abstract

This paper describes an image registration technique which utilizes a maximum likelihood estimate of translational shifts between images. Through use of a detailed error model, insight is gained into the suitability of common registration methods such as correlation. The relationship between the optimal statistical-based registration algorithm and algorithms which have been reported in the literature is also presented. Models of the two images (search and reference) to be registered are developed, assuming the search image is a combination of the reference image, additive i.i.d. Gaussian noise, and areas often referred to as background or clutter. These clutter areas may or may not have statistics similar to the reference areas. The image models differ from past models which implicitly assume a reference image of infinite dimension. Where the reference image is of infinite dimension, it is shown that Gaussian statistics lead to a likelihood equation yielding a square difference template-matching technique.