Optimal Transport Between GMM for Multiscale Texture Synthesis

Julie Delon, Agnès Desolneux, Laurent Facq, Arthur Leclaire

Abstract: Using optimal transport in image processing tasks has become very popular. However, it still faces difficult computational issues when dealing with high-dimensional distributions. We propose here to use the recently introduced GMM-OT formulation, which consists in restricting the optimal transport problem to the set of Gaussian mixture models. As a proof of concept, we use it to improve the texture model Texto based on optimal transport between distributions of image patches. Using GMM-OT in this texture model allows to deal with larger patches, hence providing results with better geometric details. This new model allows for synthesis, mixing, and style transfer.