A Generative Model of Natural Texture Surrogates

Natural images can be seen as patchworks of different textures for which local image statistics are roughly stationary within confined regions but otherwise can be highly diverse. In order to model the natural variety of textures, we sampled 64×64 patches of homogeneous textures from a large image database and described each patch by a set of texture parameters obtained with a popular texture algorithm. Texture variations across natural images can thus be expressed by the empirical joint distribution of the texture parameters. After suitable post-processing, we were able to fit this distribution with a multivariate Gaussian. Sampling from the model, one obtains random vectors that can be back-transformed into texture parameters from which new textures can be synthesized. These generated textures share with natural images not only the characteristics of the second-order statistics (power spectrum) but also resemble higher-order image correlations to considerable extent. Thus, we have devised a generative model of natural texture surrogates with which we can generate fully controlled ensembles of complex stimuli useful for probing the visual system for perceptually important high-order correlations of natural images. We demonstrate the descriptive power of our model by achieving state-of-the-art compression results based on texture synthesis using a simple quantization scheme of the texture coefficients. In addition, we demonstrate how our approach can be useful for evaluating the descriptive power of generative models of natural images. ar X iv :1 50 5. 07 67 2v 1 [ cs .C V ] 2 8 M ay 2 01 5