Feature Analysis and Texture Synthesis
Yuanting Gu, Enhua Wu · 2007
Most texture synthesis algorithms explicitly or implicitly adopt Markov random field or similar distribution as their basic model to guide the synthesis process. However, MRF-like models can 't handle textures well with large scale structure or unstable structure due to their inherent local and stable assumptions. To make improvement in this regard, we propose a new texture analysis/synthesis framework that combines two main ideas. Firstly, in material space we decompose the texture contents into units with "basic shape " and "feature vector". Based on this, the space spanned by a set of sampled textons is constructed to help introduce additional changes upon textons. Secondly, in pattern space, using the idea of "feature texture " acquired from texture swatch for different properties especially for distribution rules of textons, we may capture and manipulate the global structure flexibly. By this formulization, we are able to obtain a satisfactory texture appearance, and also a rich controlability as well.