Locally countable properties and the perceptual salience of textures

The human ability to discriminate structured from uniformly random binary textures has been shown to exploit third- and higher-order pixel correlations. We examine this ability in an experiment using a large number of texture families that can only be distinguished on the basis of these higher-order correlations. This study investigates statistical models based on possible explanatory variables involving spatial interactions of up to four pixels. Some of these explanatory variables have been recently associated with natural images, and others are somewhat less intuitive and are used here for the first time, to our knowledge. Our models are constructed using intraclass and cross-class feature selection by means of lasso/elastic net optimization and extensive cross-validation. We focus on a special set of locally countable image measures that seem to parsimoniously capture the observed discrimination performance. Among the measures underpinning the best models, we highlight a concept that can only exist in nine-pixel or larger image patches, but nonetheless is calculable based on the multiplicity of specific four-pixel patches in a texture. We show that this single geometric concept provides significant clues to explain texture discrimination.