Research On Flow Motion Analysis And Synthesis In Visual Media

There are challenges associated with image analysis and video synthesis in the area of computer vision, which have bright future in the applications of digital entertainment and weather forecast. People need not only real sense of image, but also realistic motion. It is difficult to describe the process of real motion due to the complexity, so physical law of the real world is needed. Behind the reality is physics, which describes how objects move and interaction with each other. However, on one hand, the physical law of real world is complicated, the physical description of phenomena is not good and development of corresponding field is not mature. Building a perfect physical model is not an easy task. On the other hand, even there is an applicable model, direct use of the model is difficult because it require too much computational power. Another point is to enhance the control of synthesis, by which a sound control strategy to animators can be achieved. Artistic effect can be made through simple interface.Real time computing is more important than precise of performance in computer vision. Low computational cost will make people preview the possible effect conveniently, thus animating scene can be adjusted quickly and less emphasis is needed on the precision of the problem.Seldom existing methods for animating and synthesizing images satisfy the requirement of various scenes. In terms of visual media, this thesis is aimed at solving the problems of various, flexible video or animation scene based on flow mechanics, and proposes effective algorithm to satisfy the need of application. The major contributes of the thesis are as follows:1) Based on the theory of flow mechanics, we construct basic 3D patterns and flow models such as water and fog to demonstrate the basic motion of flow that contains structure and expression of basic mathematic models. With the mechanism of flow mechanics, we construct basic models such as uniform, source, sink and rotaion, genenrate basic movement in visual media and creat basic advection pattern using N-S model. 2) On the basis of mathematic models of flow, we combine complicated motion. Via adjustment, all of these patterns can be expressed dynamicly in visual media to creat various scene. We segment flows using image matting technique. Image inpainting is performed to fill the gaps to do the job of pre-process of the image.3) Using static images that contain flows as our research objects and after image-processing step, we project flow models onto static flow images through projection methods in computer vision such as pinhole camera model and vanish points and lines to creat continuous video. We propse seamless algorithm to make harmonious movement between flow and scene. Via the adjustment the physical model setting and modifying the basic flow patterns, good-looking animation is genenrated.4) In terms of video flow analysis, we learn from nature scenes to get more natural and flexible videos. We also get new flow motion through the analysis and synthesis of input flow video textures.