Human Texture Perception

Textures are important visual cues about surface property, scenic depth, surface orientation, and etc. Amazingly, human vision system utilises the information effectively in interpreting the scene and performs very efficient texture discrimination and segmentation. Researches show that texture perception in human vision is one of the early steps towards identifying objects and understanding scene [53,3].

Texture, and its effect on human visual perception has been the subject of interest to the vision community, which has been extensively studied in multiple disciplines including neuroscience, psychophysics, and computer science. In neuroscience and psychophysics, texture studies focus on neural processes involved in visual perception, with a great amount of efforts on understanding the mechanisms of texture detection and segregation. In computer vision, texture studies focus on simulating human texture perception via computing technologies and deriving appropriate mathematical representations of textures to facilitate computerised texture processing, classification and segmentation. Texture studies in computer vision are not mere mathematical problems, because they are closely related to studies in neuroscience and psychophysics.

Figure 4.6: A typical synthetic image used in pre-attentive texture discrimination experiments [3]. This texture contains three regions: the background region with L-shaped figures, the left region with X-shaped figures, the a right region with T-shaped figures. Notably, the left region is easily detected from the background while the right region is much harder to discriminate.

The very first steps toward computational texture analysis were conducted by Julesz et al. [53,54]. They empirically investigated the perceptual significance of various image statistics of texture patterns in order to determine how the human low-level visual system responds to the variation of a particular order statistic. In their experiments, carefully selected synthetic textures with either repetitive or randomly placed micro-patterns, such as lines, dots, and symbols, are used, each corresponding to a certain order statistic. Examples of order statistics include contrast (first-order), homogeneity (second-order) and curvature (third-order). Figure 4.6 shows a synthetic image that is commonly used in this and other similar experiments.

Julesz Conjecture and the texton theory are two main contributions resulted from their studies, which open up many other research avenues.

Julesz Conjecture

suggests that human cannot distinguish between textures with identical second-order statistics[53]. The conjecture was proven to be false by Julesz himself [55], but it established an important idea that texture might be modelled using low-order statistics. Today's main stream texture analysis approaches characterise texture with a set of sufficient statistics, which are at least conceptually based on the Julesz Conjecture. Due to limited computing power, many of the available approaches still rely on first- and second-order signal statistics. It should be noted that images like in Fig 4.6 display some local features, i.e. small strokes, that require higher-order pixel-wise signal statistics to represent.

Texton Theory

states that textons are ``the putative units of pre-attentive human texture perception'', related to texture's local features, such as edges, line ends, blobs, etc. Julesz observed that human texture discrimination could be modelled by first-order density of such textons [54].

Researches show that human visual system performs local spatial frequency analysis on retinal images which could be simulated by a computational model using a filter bank [57,21]. This theory has motivated mathematic models of filter-based human texture perception. For example, Bergen [3] suggests that a texture can be decomposed into a series of sub-band images using a bank of linear filters at different scales and orientations. Each sub-band image is related to certain texture features. Therefore, a texture is characterised by an empirical distribution of the magnitude of filter responses and therefore similarity metrics, e.g., distance between distributions, could be derived for discriminating textures.

As a whole, the study of visual texture perception is an important subject area in vision science. A variety of theories have been developed to understand the mechanisms about human texture perception. Texture research in neuroscience and psychophysics has greatly influenced the counterpart in computer vision. The Julesz Conjecture, for instance, inspired the statistical approach to texture analysis which characterises a texture by image statistical features. The texton theory led to a structural approach that extracts texture primitives as local features for texture description. The filter-bank model has also been introduced into computational texture modelling, resulting in methods that decompose a texture using filters and extend texture analysis into the frequency domain.