A simple texture descriptor and its applications to attention

Jan-Olof Eklundh (KTH Stockholm, Sweden)

Texture as a characteristic of visual appearance in recognition, classification and segmentation is a widely studied topic and numerous approaches to representing and describing it have been proposed. The most natural ones are based on applying a set of filters or local operators to the image and then extract a set of descriptors from the output of these. When complete filter bases are applied one of course retains all the information so in a sense there is little to add to such approaches. However, the descriptors are then themselves rather complex and high-dimensional, something that one wants to avoid in some situations. For instance, salience computations in attention are often based on variants of the feature integration model proposed by Treisman and Gelade. High-dimensional input is then difficult to integrate with e.g. color and edge information. In this talk we propose a simple texture descriptor that in the extreme case just consists of a real number. The descriptor represents the dependence between the lines (and columns) in a square local image patch and is computed so that the response to edges is suppressed. We'll discuss the properties of the descriptor, how it can be efficiently implemented and also show experimentally that it indeed give desirable results in texture discrimination tasks. We finally apply it in an attentional model demonstrating that it aids in attending to textured regions in addition to other salient structures.