Research And Application On User-controllable Multi-exemplars Texture Synthesis

Textures can describe a wide variety of natural phenomena. And texturing is a core process for computer graphics applications, which can enhance the realistic rendering greatly. With the rapidly increasing demands of realistic rendering in the field of movie making and video games, single texture synthesis cannot meet the needs. Multi-Exemplars synthesis is a challenging research topic, which can increase the richness of texture details, with a wide range of application scenarios. In this paper, we just focus on the multiple exemplars synthesis. The algorithm implements Multi-Exemplars synthesis under the guide of control map provided by users. We propose a new kind of input structure, known as "exemplars tree", in order to better integrate the information of multi-exemplars to form a single larger texture.The "exemplars tree" reorganizes the multi-exemplars based on the scale relationship as the input. The multi-exemplars synthesis algorithms before generally relay on the automated synthesis, as a result the resulting image is poorly controllable. In this paper, the highest level of the exemplars tree can be a control map image provided by the users. A variety of texture information can be synthesized according to the users' instructions by control map, which increased the controllability and flexibility.To further improve the quality of synthesis results, we use bilateral filtering technology to improve the texture analysis stage. The improved exemplar stack is called the "bilateral stack". Bilateral filtering can maintain the boundary well while smoothing the image, which feature can optimize the effect of the algorithm."Bilateral stack" can improve the boundary of the hybrids of multi-exemplars, as well as make the texture elements in the composite image more clearly.In the paper, we accelerate the algorithm. In order to calculate the best match of each pixel, the neighborhood information is needed according to MRF (Markov Random Field) model. We accelerate neighborhood matching by projecting the 5×5 pixel neighborhoods into a truncated 6D principal component analysis (PCA) space. And we demonstrate optimizations for GPU implementations of our method to make the algorithm parallelization and to improve the program efficiency.