Estimating a true set of matching corresponding pixels has always been a challenge for image registration. This paper formulates the alignment problem as an estimation problem using expectation-maximization (EM) algorithm. With a given input of set of correspondence points and their distribution function a posterior probabilities of correspondence can be estimated and maximized until convergence. A simple thresholding can then distinguish between the inliers and outliers which can recover the coherent pixel correspondence between the image pairs. The experimental results have shown efficient results for images with rigid motion of translation and rotation.

image registration, expectation, maximum-likelihood, point correspondence, coherence