Image synthesis using multiscale model

Considers a specific class of textures which are stochastic, possibly periodic, two-dimensional signals displaying fractal-like or self-similar characteristics. Most natural textures belong to this class. We explore the use of autoregressive (AR) processes on trees as a texture model. This theory was originally proposed for multiscale signal analysis. The generalized lattice structures used for parametrization of AR processes on trees makes computer implementation fast and efficient. We have done extensive experiments on texture generation and study the effect of reflection coefficients and model order on the quality of the generated textures