Study On Non-photorealistic Rendering Theory And Method Of Artistic Style Image

Non-photorealistic rendering (NPR) is a new problem in computer graphics andimage processing, belonging to computer art category. It is an effective combination ofcomputer science and art, which simulate artistic rendering style by designing algorithmsand creating mathematical models with the help of the computer technologies. It doesn'tlook for the photorealistic but emphasizes people's subjective feelingsTechniques are classified according to source data into object-space methods that use3D model of a scene to create the rendered image, and image-space methods that workdirectly on the 2D image. This thesis restricts itself on the topic of painterly stylesrendering theory and method based on 2D image. Based on the theory of fuzzymathematics, statistics, agent and bionics, umpty problems on color transferring, texturesynthesis-based artistic style transferring and stroke-based rendering are studied. Somenovel models and algorithms have been proposed to effectively and efficiently renderpainterly styles based on the existed achievement of previous researchers, which aredetailed as follows:(1) It proposes an adaptive color transfer algorithm based on fuzzy colorclustering. The source and target images are first transformed into taft space from RGBspace. Then fuzzy color clustering is applied to obtain different clustering fields.According to the target statistical properties, the matching weights are calculated todetermine the direction of local color transfer. Finally, membership degree control isintroduced to color transfer equations to realize the adaptive color transfer between images.The experimental results show that the improved algorithm can get more natural and moreharmonious effects than former color transfer algorithms.(2) It proposes a fast style transfer algorithm. A new neighborhood similaritymetric is builded by improving the classical Ashikhmin's approach and using the statisticalcharacteristic belonging to source image, target image and their blurred images. In addition,producing extra candidates in random positions solves the circs that Ashikhmin's approachtends to create repeated candidates. And meanwhile, visual integrality is maintained byapplying the bi-directional scanning strategy into style transferring. Enough experimentsand results have proved that the improved algorithm can realize style transferring betweenimages faster and more effectively than former algorithms.(3) It proposes a multi-agent rendering algorithm. In this algorithm, renderingtask is divided into three sub-missions: contour rendering, sketch rendering and washing-over. Each stage is implementd by one type of rendering agents. Rendering agentis defined as an agent with attributes of position, lifecycle, type designator and visual range.These rendering agents interact with rendering environment and take different actionsaccording to different local environment incentive information. They can finish respectiverendering tasks by searching their characteristic pixels and drawing them. All marks formfinial-rendering results. Experimental results show this algorithm can simulate the paintingprocess efficiently.(4) In this paper, a NPR method based on ant colony algorithm (ACS) withequilibrium of distribution is proposed. Based on bionics idea, this method gives theconcept of rendering ant and it is defined as an agent with attributes of pheromone, mark,lifecycle and so on and with behaviors of single-walking, rendering marks, updatingpheromone, jumping and dying. The algorithm can direct rendering ants to select the mostoptimal rendering trail and adjust their behaviors such as single-walking, rendering marks,jumping and dying by updating pheromone continually. The marks that all rendering antslay on the trails are just the rendering result. Because rendering ants select trails purelyaccording to the selected probability, the tracks will converge by degree, which will lead tolocally rendering. To avoid this phenomenon, information weight is brought in todynamically adjust the equilibrium of selected probability. And this brings more naturalrendering results. Experimental results show that this method can simulate the paintingprocess effectively and can get pleasant rendering effects.(5) It presents an adaptive LIC rendering algorithm for fluid artistic style. First,image structure vector field is derived from source luminance orthonormal vector sets. Andusing neighbor vectors to improve the weak gradient fields produces the structure vectorfield that is needed by LIC rendering. Then, LIC convolution step-length is adjustedaccording to local curvature in the structure vector field. LIC convolution length iscalculated according to the local statistical characteristics of the texture image. Finally, thealgorithm transfers the reference image's color characteristics to the rendering result tomake it full of rich color like oil paintings. Experimental results comfirm that thisalgorithm can create vivid and flexible fluid artistic style such as van Gogh's style.