Texture synthesis using asymmetric 2-D noncausal AR models

The author investigates the suitability of two-dimensional (2-D), noncausal, autoregressive (AR) models with possibly asymmetric support for synthesis of images visually similar to natural textures. These models characterize the gray level at an image pixel as a linear combination of gray levels at nearby locations in all directions and an additive non-Gaussian, higher-order white noise variable. Existing results based upon the second-order statistics of the images assume that the model support is symmetric, whereas the author exploits higher-order statistics of the image to fit AR models with possibly asymmetric support. Experimental results of synthesis of 128*128 textures visually resembling several real life textures in the Brodatz album (and other sources) are presented. The synthetic textures are generated using models obtained from real images via inverse filter criteria.