Token-textured object detection by pyramids

Abstract

Synthetic textures, on which human vision is tested, are usually composed of tokens such as dots, lines, Ts or crosses that are easily visible to subject tested 1. Humans discriminate easily between textures that are statistical mixtures of grey-level dots, if they differ enough in their statistical parameters 2. Discriminating between textures that differ in higher moments of their distribution is hard to reconcile with the common pyramidal multiresolution approach. We propose a new kind of pyramid that achieves good discrimination of statistically distributed grey-level and token-textures: discrimination between distributions having equal means, that is impossible to achieve with the intensity pyramid paradigm. Textures composed of statistical mixtures of tokens other than dots were tested on human subjects and found to produce results similar to those of grey-level dots. To discriminate by machine between token-textures the pyramidal approach is combined with a paradigm for composing the responses of several standard detectors. Human performance was tested on the same set of textures.