Figure 1
Visual texture is a perceptual property of image regions. In this respect texture is analogous to perceived brightness, COLOR, DEPTH, and MOTION. Just as we can describe a region of an image as bright or dark, red or green, near or far, or moving up or down, we can describe it as having a mottled, striped, or speckled texture. Note that none of these descriptions is redundant or inconsistent with any of the others; saying that a region is bright red, distant, and moving up does not in any way constrain its textural properties. In natural images visual texture frequently is related to qualities of surfaces indicating roughness or smoothness or other physical properties.
The English word texture derives from the Latin texere, meaning "to weave." "Texture" was first used to describe the character of a woven fabric as smooth, ribbed, twilled, and so forth. Similarly, in describing visual texture we refer to such properties of an image region as granularity, periodicity, orientation, and relative order or randomness.
Interest in the perception of visual texture has both applied and theoretical roots. Texture rendition is important for creation of natural-looking images. However, because of its high level of spatial detail, digital representation of image texture can also be expensive in terms of storage and transmission bandwidth. Therefore, it is useful to understand what characteristics of texture are perceptually important so that these can be represented efficiently. From the standpoint of understanding visual processes texture is also of interest because texture perception differs in some ways from more general visual perception.
Figure 1
The most basic manifestation of this difference can be seen in the fact that regions containing visually discriminable structure do not always form distinct textures. For example, as noted by Beck (1966; see also an example in Bergen 1991) a bipartite region containing upright T and L shapes appears as a uniformly textured area while a bipartite region containing upright and tilted T shapes (or tilted T and upright L shapes) shows a clear division into two visually distinct areas (see figure 1). This greater effectiveness of the orientation difference is not predicted by pattern discrimination properties: T, L, and tilted T shapes are all easily distinguishable when inspected as individual patterns within the texture image. This phenomenon is referred to variously as "texture segregation," "texture-based segmentation" or "preattentive," "instantaneous," or "effortless" texture discrimination. It can also be described as the tendency for some pairs of textured stimuli to induce the formation of an illusory contour between regions of different spatial structure (see SHAPE PERCEPTION, GESTALT PERCEPTION, SURFACE PERCEPTION). One of the major areas of activity in the study of texture perception has been to try to isolate the stimulus characteristics that support this phenomenon. Descriptions of these characteristics have included local features (e.g., Beck 1966, 1982; Julesz 1981), pixel statistics (e.g., Julesz 1962, 1975), and linear filters (e.g., Harvey and Gervais 1978, 1981; Richards and Polit 1974; Beck, Sutter and Ivry 1987; Bergen and Adelson 1988). For reviews of this work see Bergen (1991) and Graham (1991).
A rather ubiquitous idea found in this body of work is that texture segregation is based on a simplified analysis of spatial structure. According to this view, the reason that some texture pairs segregate and others do not is that the texture analysis process does not possess the powers of discrimination present in more sophisticated processes such as VISUAL OBJECT RECOGNITION, AI. Thus, the nonsegregating textures (such as those composed of the T and L shapes in the example) look the same to the (hypothetical) texture analysis process even though they look different to the (hypothetical) pattern discrimination process. Possible reasons that the visual system would perform such a simplified analysis involve issues of speed and complexity. It may be more important to have rapid (or "early") information about texture structure than to respond to all visible structure differences. This would be the case if texture-based segmentation has an orienting or ATTENTION control function. Alternatively, it may simply be too expensive of processing resources to make a more detailed analysis of texture structure. Texture stimuli have the inherent characteristic of spatial density as compared to the relative sparsity of other visual characteristics. Dense, more complex processing may not be a biologically feasible option.
Texture segregation is most strongly driven by simple differences in local spatial structure such as granularity (coarseness or spatial frequency content), orientation, and sign-of-contrast. These are also the kinds of spatial characteristics that most strongly determine the physiological responsiveness of cells in the early stages of mammalian VISUAL CORTEX (see also VISUAL ANATOMY AND PHYSIOLOGY). This coincidence raises the intriguing possibility that texture properties are actually computed at this rather early stage of visual processing. This possibility has been associated with all of the different styles of texture descriptions described earlier; at a qualitative level local feature extraction, analysis of image statistics, and analysis of spatial frequency content are all plausible (although crude) descrip-tions of early cortical processing.
The relationship between models of early visual processing (particularly those based on some understanding of mammalian cortex) and texture perception phenomena has been the basis for computational and psychophysical investigations. Examples include Caelli (1985), Turner (1986), Malik and Perona (1990), Bergen and Landy (1991), Chubb and Landy (1991), Graham (1991), and Landy and Bergen (1991). The underlying information representation in most of these studies is a collection of linear filters sensitive to different scales and orientations of spatial structure, possibly with a preceding static nonlinearity. These models (sometimes referred to as "energy models" or "filter models") provide a good description of texture phenomena for certain classes of stimuli. Heeger and Bergen have demonstrated that for textures of relatively "random" structure such models can be used as the basis for synthesis of textures that match a target texture in appearance. This procedure also demonstrates the limitations of this kind of representation: for textures with nonlocal or quasi-periodic structure, physiologically motivated filter models do not successfully capture texture structure. Promising extensions to this approach can be found in the work of Popat and Picard (1993) and De Bonet (1997), which uses more explicit spatial information to represent more coherent structure.
It is possible that the image characteristics that are commonly referred to as "texture" actually fall into two different categories: texture and pattern. In this case a representation comprising both components may be necessary to give a complete description. This may also mean that attempts to associate visual texture perception phenomena with the properties of visual brain components may succeed in the sense that aspects of texture perception can be related to individual or group properties of neurons, but fail in the sense that there is no single neural system displaying all of the properties of "texture".
Beck, J. (1966). Effect of orientation and of shape similarity on perceptual grouping. Percept. Psychophysics 1:300-302.
Beck, J. (1982). Textural segmentation. In J. Beck, Ed., Organization and representation in perception. Mahwah, NJ: Erlbaum.
Beck, J., A. Sutter, and R. Ivry. (1987). Spatial frequency channels and perceptual grouping in texture segregation. Computer Vis. Graphics Image Processing 37:299-325.
Bergen, J. R. (1991). Theories of visual texture perception. In D. Regan, Ed., Spatial Vision. London: Macmillan, pp. 114-133.
Bergen, J. R., and E. H. Adelson. (1988). Early vision and texture perception. Nature 333:363-364.
Bergen, J. R., and M. S. Landy. (1991). Computational modeling of visual texture segregation. In M. S. Landy and J. A. Movshon, Eds., Computational Models of Visual Processing. Cambridge, MA: MIT Press, pp. 253-271.
Caelli, T. M. (1985). Three processing characteristics of visual texture segregation. Spatial Vision 1:19-30.
Chubb, C., and M. S. Landy. (1991). Orthogonal distribution analysis: A new approach to the study of texture perception. In M. S. Landy and J. A. Movshon, Eds., Computational Models of Visual Processing. Cambridge, MA: MIT Press, pp. 291-301.
De Bonet, J. S. (1997). Multiresolution sampling procedure for analysis and synthesis of texture images. Computer Graphics Proceedings, Annual Conference Series 1997, ACM SIGGRAPH, pp. 361-368.
Graham, N. V. (1991). Complex channels, early local nonlinearities, and normalization in texture segregation. In M. S. Landy and J. A. Movshon, Eds., Computational Models of Visual Processing. Cambridge, MA: MIT Press, pp. 273-290.
Harvey, L. O., and M. J. Gervais. (1978). Visual texture perception and Fourier analysis. Percept. Psychophysics 24:534-542.
Harvey, L. O., and M. J. Gervais. (1981). Internal representation of visual texture as the basis for the judgment of similarity. F. Exp. Psychol. Hum. Percept. 7:741-753.
Julesz, B. (1962). Visual pattern discrimination. IRE Trans. Information Theory 8:84-92.
Julesz, B. (1975). Experiments in the visual perception of texture. Scient. Am. 232:34-43.
Julesz, B. (1981). Textons, the elements of texture perception and their interactions. Nature 290:91-97.
Landy, M. S., and J. R. Bergen. (1991). Texture segregation and orientation gradient. Vision Research 31:679-691.
Malik, J., and P. Perona. (1990). Preattentive texture discrimination with early vision mechanisms. Journal of the Optical Society of America A 7:923-931.
Popat, K., and R. W. Picard. (1993). Novel cluster - based probability model for texture synthesis, classification, and compression. Proceedings of SPIE Visual Communications and Image Processing: 756-768.
Richards, W., and A. Polit. (1974). Texture matching. Kybernetik 16:155-162.
Turner, M. R. (1986). Texture discrimination by Gabor functions. Biological Cybernetics 55:71-82.
Copyright © 1999 Massachusetts Institute of Technology