Research On Sample-based 2D Texture Synthesis

The technique of sample-based 2D texture synthesis is an important research topic in computer graphics recently. Based on the strong relativity of neighboring points, Efros and Leung presented an algorithm with non-parametric sampling, which is an effective synthetic method for many textures. The Efros and Leung's algorithm was mended by Wei and Levoy through sampling directly from swatch picture and avoiding to construct explicit probability function, which make the synthetic process more efficiency. Based on relativity principle, the Ashikhmin algorithm improved synthetic speed extremely by limiting the search range in the neighborhood of current point, and the algorithm get nice quality for nature textures. XU Xiao-gang presented a multi-seed fast texture synthesis algorithm, which search matching point using helical searching instead of the full searching based on scan line to accelerate matching process. In 2001, Efros presented an image quilting algorithm, which changes the basic synthesis unit from single pixel to texture block in sample. The algorithm can get nice quality for structural texture which can't be transacted ideally in other algorithm. But it is still a difficult work to explore a algorithm which has both synthetic efficiency and quality. Furthermore, no algorithm can transact all textures at present.Firstly, the major applications about texture synthesis are introduced in this paper. Then based on the studies about some typical 2D texture synthetic algorithms, Some comparisons and summarizations about the ideas and transacting range and defects of these texture synthetic algorithms are drawn. An accelerated method for Wei and Levoy's basic algorithm (for short, WL algorithm) is presented. Plenty of experiments show that it can effectively improve the synthetic speed of WL algorithm at the same quality. A method for automatic recognizing the size of neighborhood based on Statistics is also presented, which can decrease the number of input parameters and be convenient for user's operating in texture synthetic algorithms sample-based using MRF model.Finally, the future works about the technique of texture synthesis are introduced.