Aspects of Visual Texture Discrimination

Rick Gurnsey and Roger A. Browse

Abstract

On the simplest view, textural segmentation may be characterized as involving (a) the measurement of certain image properties and (b) detection of differences in these properties between neighbouring regions. Apparently following this simple view, questions have been asked about which image properties permit effortless discrimination of textures by humans. We show results which indicate that several other factors must be taken into account in a satisfactory theory of textural segmentation. As might be expected, the probability of discriminating two textures depends on how long subjects have to examine the textural display, Contrary to what might be expected, however, texture discrimination is not symmetric; the probability of discriminating two textures, in many cases, depends on which forms the foreground (disparate) region and which forms the background. It would seem then, that discrimination cannot be based entirely on local differences between textured regions; a difference signal should be indifferent to the "sense" of the boundary. Furthermore, the ability to discriminate two textures depends on the amount of practice that subjects have had with the materials and procedure. In general, with naive subjects, the probability of discriminating two textures and the probability of a discrimination asymmetry arising, is related to differences in rather simple image properties such as scale, contrast and orientation.

Texture may be considered an image property which, like colour, contrast, and disparity, provides a basis upon which images may be parsed into coherent regions prior to interpretation. This view is explicit in the computational vision literature and most textbooks on the subject (e.g., Ballard and Brown, 1982; Levine, 1985; Navatia, 1982) include chapters on the use of texture analysis in image segmentation. One motivation for the view of textural segmentation as a percursor to interpretation is the apparent ability of humans to spontaneously segment regions composed of different textures (e.g., Julesz, 1984).

Perhaps because of the work on image segmentation in the computational vision literature, the study of human textural segmentation has undergone somewhat of a renaissance (e.g., Beck, 1982; Beck, Pradzny and Rosenfeld, 1983; Caelli, 1982, 1985; Caelli and Moraglia, 1985; Enns, 1986; Grossberg and Mingolla, 1985; Gurnsey and Browse, 1987a; Julesz, 1981, 1984; Julesz and Bergen, 1983; Nothdurft, 1985a, 1985B, 1985C; Richards, 1979). Most texture experiments involve artificial textures, usually constructed from micropatterns chosen because they have certain theoretically relevant properties. Presumably, the textural properties found to permit discrimination in the laboratory are also important in more usual environments. In general, however, we shall consider artificial textures comprising two spatially disjoint regions.

Central to many investigations of texture processing is the rather simple question: What limits our ability to segment two textures? Answers to this question typically concentrate on the properties of the textures involved; however, we shall show that there are several other theoretically relevant aspects of human visual texture discrimination that are not addressed by current theories.

This research was supported by a grant from the Natural Sciences and Engineering Research Council of Canada (A2427). We are grateful to Dr. Brian Butler and Marion Rodrigues for their comments.