Abstract

Modeling and retrieving the transform parameters that characterize the underlying deformation field is the main crux of the problem in automatic image registration domain which involves employing a similarity measure in an image pair and a robust model estimator. Model estimators can be either a least square fit or an optimization method which finds minimum of a cost function. In this work, a stochastic mutual information based adaptive gradient descent optimizer is proposed in which transforms such as translation, affine and free form deformations are accurately retrieved in the process of image registration and only a percentage of population of intensities is used to estimate mutual information without losing accuracy in a stochastic way. Better than one tenth of a pixel accuracy is achieved in image registration by retrieving different geometric transformations accurately.