Evolutionary algorithms have been used recently as an alternative in image registration, especially in cases where the similarity function is non-convex with many local optima. However, their drawback is that they tend to be computationally expensive. Trying to avoid local minima can increase the computational cost. The purpose of authors' research is to minimise the duration of the image registration process. This paper presents a method to minimise the computational cost by introducing a machine learning-based variant of Harmony Search. To this end, a series of machine-learning regression methods are tested in order to find the most appropriate that minimises the cost without degrading the quality of the results. The best regression method is then incorporated in the optimisation process and is compared with two well-known ITK image registration methods. The comparison of authors' image registration method with ITK concerns both the quality of the results and the duration of the registration experiments. The comparison is done on a set of random image pairs of various sources (e.g. medical or satellite images), and the encouraging results strongly indicate that authors' method can be used in a variety of image registration applications producing quality results in significantly less time.