Patch matching with polynomial exponential families and projective divergences

Abstract

Given a query patch image, patch matching consists in finding similar patches in a target image. In pattern recognition, patch matching is a fundamental task that is time consuming, specially when zoom factors and symmetries are handled. The matching results heavily depend on the underlying notion of distances, or similarities, between patches. We present a method that consists in modeling patches by flexible statistical parametric distributions called polynomial exponential families (PEFs). PEFs model universally arbitrary smooth distributions, and yield a compact patch representation of complexity independent of the patch sizes. Since PEFs have computationally intractable normalization terms, we estimate PEFs with score matching, and consider a projective distance: the symmetrized γ-divergence. We demonstrate experimentally the performance of our patch matching system.