Analysis of appearance space attributes for texture synthesis and morphing

Texture synthesis and morphing are important techniques for efficiently creating realistic and visually attractive textures. A popular class of synthesis algorithm are pixel-based techniques, which search in a given 2D exemplar for a pixel with a similar neighbourhood to the pixel currently being generated. The methods have the advantage that they are fast, they can be easily generalised to higher dimensions, and synthesised textures can be represented as references to the exemplar which allows definition of additional channels, such as displacement maps, at no additional cost. The quality of pixel-based techniques depends on the metric used to compare pixel neighbourhoods. Lefebvre and Hoppe introduced the term appearance space for measures describing pixel neighbourhood similarity. In this paper we introduce new appearance space attributes and evaluate them and existing attributes for texture synthesis and morphing. Our results show that our proposed gradient estimate significantly improves synthesis and morphing quality with little additional cost.