To suppress local minima, we smooth the cost function of image registration with a randomized smoothing (RS) technique.

A convolution-based smoothing technique in which the original cost function is convolved by a Gaussian distribution with an efficient randomized optimization procedure.

The RS technique is applied to translation, rigid, affine and B-spline transformation models.

Experiments on 2D artificial images, 2D cell images, 3D lung CT, and 3D brain MRI scans confirm the benefits of the method.