Research On Sample-Based Texture Synthesis

The texture synthesis is widely used in computer vision, computer graphics and image processing. The high-quality and fast sample-based texture synthesis method is one of the most important methods, which can avoid some shortcomings of the traditional texture mapping and procedural texture synthesis, for example complicated parameters. Therefore, this method is used most widely. Although various types of texture synthesis algorithms are appeared, but the texture’s essential attribute-self similarity is still incomplete, and the evaluation of the texture quality is also ignored.This thesis first surveys the existing state-of-art work on texture types, texture synthesis algorithm and the result evaluation. Based on this survey, this thesis present effective self-similarity definition method, an novel idea of texture synthesis and an numerical evaluation method about texture synthesis results. The contributions of the thesis mainly include:First, taking into account the fact that the texture recognition can be based on the texture structure and texture appearance characteristics, the thesis gives the definition of self-similarity and the quantitative method which to measure the self-similarity. Based on the texture characteristics of texture self-similarity, this method introduces the structure self-similarity and appearance self-similarity. For the measure of the structure self-similarity, we first extract edge information in a certain scales, then refine edge, select the special positions and divide the sub-regions to obtain the shape difference factor and area of difference factor considering the radius and area. Then we get the texture structure self-similarity based on the most similar sub-regions. About the measure of texture appearance self-similarity we consider two aspects, firstly, we carry out the histogram distance between the sub-regions, and get the most similar sub-regions; secondly, we divide texture into sub-blocks, and calculate the differences between the sub-block and the entire texture within the threshold control range, and then compute the sub-block appearance similarity of texture.Second, achieving the self-similarity of texture, we present a novel seeding texture synthesis algorithm. This thesis analyse the texture sample, get the seeds of the sample, and then synthesize textures. The procedure mainly includes:seed selection, seed separation, growth pattern, planting, seed growth and correction. We introduce the algorithm survey, the algorithm description, the algorithm implement and the algorithm experimentsn. Experiments with a variety of textures demonstrate that our method is simple, clear and efficiently synthesize various textures.Third, the texture synthesis results evaluation and texture types division are very important problem, but it is neglected so far. Based on the definition of texture self-similarity,this thesis introduces the methods of texture classification. In addition, we design a result evaluation method based on studies on people’s perception. We compare the subjective evaluation results with the numerical method results. This study not only gets the effective texture classification method, but also enrichs texture synthesis system framework, and provides the novel idea for the evaluation of the texture synthesis results.