Research On The Stylization Of Image(Video)

Stylize rendering is a research field that combines computer technology and painting art. Its main aim is to simulate artist's painting effect by using computer. In consequence, it can efficiently generate different kinds of art style. This research has wide applications, including generation of fast cartoon, computer animation and so on. The thesis is mainly focusing on how to produce cartoon and oil painting effects. The main contribution of this thesis includes:(1) Proposing an importance map based image cartoon rendering algorithm. Firstly, the algorithm extracts importance map to capture contour features of objects in image. And the important map is used for avoiding over-segmentation problem. Secondly, Based on the segmentation result, the algorithm remove boundary glitch by using topology-preserving smoothing method. It can obviously improve the visual effect of cartoon results. Finally, the region filling is used to output the final cartoon result.(2) Improving W.Kang's oil painting algorithm. The existing algorithm has very obvious swirl problem in generating direction map. The improved algorithm smooths the direction map by using edge tangent flow. As a result, the direct map can describe the image structure feature for the next rendering. Then, K-means algorithm is to configure painting color. Finally, the oil painting effect is obtained by minimum energy.(3) Proposing a very convenient interactive video cartooning algorithm. The proposed algorithm only requires user's interaction on the first frame to improve the segmentation result. For other frames, the algorithms can automaticly inherit user's interactive operation by using Gaussian mixture model and motion tracking. The algorithm mainly consists of the following steps. Firstly, it builds multiple-region labeling by interactive segmentation. For each region, a Gaussian mixture model is defined by using color information. Secondly, it obtains region skeleton by motion tracking and updates Gaussian mixture model. Finally, the region labeling is obtained by optimizing energy function. And the cartooning result is easily outputted based on the region labeling and image cartooning algorithm.