Texture synthesis via matching first and second order statistics of a wavelet frame decomposition

Abstract

In this novel approach to texture synthesis, a method which relies on matching the first and second order statistics of wavelet subbands is developed. Synthetic textures are produced using estimates of the correlation matrices and marginal statistics of a wavelet frame decomposition. Previous work has concentrated on iteratively matching the marginals of a subband decomposition, in the hope that the synthetic texture would capture the long range correlations. Our proposed algorithm goes beyond simple marginal matching by attempting to match the correlation coefficients of the subbands as well. We neglect any interband correlations. The algorithm is relatively fast, and has been developed with an eye to applications in regional compression.