Abstract

this paper, we describe a synthesis technique that is capable of capturing both the structural and random aspects of textures. In particular, we charactarize textures in terms of a set of statistical measurements on a wavelet representation: (1) the local spatial correlation of coefficients within each subband, (2) the local spatial correlation of coefficient magnitudes, (3) the cross-correlation between coefficient magnitudes at adjacent scales and orientations, and (4) The first few moments of the pixel histogram. Note that we have assumed both spatial locality and stationary, but not Gaussianity. We develop an efficient algorithm for synthesizing images subject to these constraints via iterative projection onto solution sets. We show striking examples of texture synthesis and constrained texture synthesis, demonstrating the power and flexibility of the model

Citations