Visual texture perception

The problem of visual texture perception is studied in this thesis. Three texture models are proposed. Each is aimed at exploring a different aspect of the problem. The first model assumes texture is composed of repeating elements. An algorithm is developed to detect, localize, and group these elements into a coherent piece of texture. The assumption of exact repeatedness enables a notion of point to point correspondence, which forms a strong link between the problem of texture and motion/stereopsis analysis. Though texture is widely studied as a 2D phenomenon (tangential texture), nature shows an abundance of relief texture. This 3D nature of texture relates appearance to surface shape in a different way from tangential texture. Moreover, occlusions, shadowing, and mutual illumination have to be taken into account. The geometric aspect of this problem is studied in this thesis through the perpendicular texture model. Intuitive relationships are derived relating the texture and surface properties to the appearance in the image. An example of surface shape recovery is also demonstrated. Finally, a general model for texture is proposed. This is called the texton model. The basic intuition is that, despite the difference in appearance, all textures are made up of symbols from the same vocabulary of tiny prototypical elements. This vocabulary can be learned from example textures. Using this model, a solution to the difficult problem of texture recognition from a single image under any lighting and viewing directions is proposed. This texton model is also used to predict the appearance of natural materials under any illumination and viewing conditions.