Binomial linear predictive approach to texture modeling

We present a parametric texture analysis and synthesis method based on a sparse autoregressive model. The method is aimed at reducing the number of parameters necessary to define the stochastic behavior of a random field representing a given texture. For this purpose, it systematically selects the dominant contributions to the pixel cross-correlation. The method is employed in the first stage of a nonlinear texture model. Dimensional and computational advantages as well as application limits are discussed.