Spatial normalization of MRI brain images by nonlinear image registration is an essential task for many applications in the field of computational neuroanatomy. Here, a multilevel block matching technique adapted to the problem of intersubject registration is presented. The concept of generalized partial volume (GPV) interpolation, which was originally used in joint intensity histogram computation, is used here in regional similarity matching. The influence of kernel function selected for GPV interpolation on the quality of registration is studied in experiments which include simulated brain images and synthetic deformations.