Conditional Gaussian Models for Texture Synthesis

Abstract

An ideal exemplar-based texture synthesis algorithm should create a new texture that is perceptually equivalent to its texture example. To this goal it should respect the statistics of the example and avoid proceeding to a “copy-paste” process, which is the main drawback of the non-parametric approaches. In a previous work we modeled textures as a locally Gaussian patch model. This model was estimated for each patch before stitching it to the preceding ones. In the present work, we extend this model to a local conditional Gaussian patch distribution. The condition is taken over the already computed values. Our experiments here show that the conditional model reproduces well periodic and pseudo-periodic textures without requiring the use of any stitching technique. The experiments put also in evidence the importance of the right choice for the patch size. We conclude by pointing out the remaining limitations of the approach and the necessity of a multiscale approach.