A Hierarchical Model for Texture Description and Synthesis

Existing models either describe textures as a non-hierarchical surface (by means of Markov-chains, time series and other stochastical methods) or distinguish only between micro and macro textures. Besides this, textures are used only for mapping colour information (usually derived from digitized photographs) on the object surface or for varying the normal vector of a given surface (bump mapping). In addition, the different models are strongly combined with special generation algorithms and the produced textures are exclusively raster images. As a consequence, the above models are not able to describe more than a few types of textures. First a definition of the term "texture" will be presented, then a survey and analysis of existing texture models will be given. Then a hierarchical texture model which is capable of describing a great number of visual textures will be described in detail. In this model we provide complete textures which will be mapped on geometrical objects and which consist of several slices. Each slice represents an optical surface property. These properties are approximated by the different parameters of the illumination model. The slices themselves are three-structured hierarchical compositions of several levels. Each intermediate node in the tree is a texture itself and is derived by operations (transformations and combinations) performed on the textures of the next lower level. The textures of the lowest level (the leaves of the tree) are either pre-existing raster fields or result from the application of one of several production rules on the texture elements (texels). The production rules determine the spatial organisation of the texels on the texture definition plane. A formal language, based on this model, enables the user to describe in a structural manner nearly every desired texture as well as to combine existing textures in order to generate a new one. These formal descriptions are transformed automatically into a sequence of algorithm calls, which will finally generate the requested textures. Finally first experiences with prototype implementations of this model will be discussed.(AGD