Model-based Coding Of Video Objects In Mpeg-4 Human Face

Low bit rate image compression has become an active research field because of the common use of computer network and increase of people's requirement to multimedia information. Due to the great improvement on bit rate and image quality, the model-based coding technology focuses researchers' eyes. The MPEG-4 standard represents this theory. It defines a face object and offers some tools to render the object.In this thesis, MPEG-4 model-based coding for facial video object is investigated. The whole codec system is generally discussed and divided into two parts of Analysis and Synthesis. In the process of image analysis, the color, gray and gradient information are exploited to find the facial feature regions, which resolves the initial placement problem of deformable template. Then the template is used to estimate the facial animation parameters, which is converted to FAP compatible with MPEG-4 and compressed later. Additionally, three head motion FAPs are established according the practice. The synthesis part reconstructs images. Parameters after error concealment drive the vertices of frame model by KD muscle model. Then the texture mapping method realizes a realizing image. A hierarchical structure is built to describe expression FAP, which is compatible with MPEG-4 and regarded as the interpolation rule to calculate other FAPs. A statistical experiment is achieved to show that it is able to do the image synthesis in real time.Besides image analysis and synthesis, model-based coding research still has an important direction to reconstruct an individual facial model. After investigating other researchers' work in this field, the thesis offers a practical scheme to match generic model to a specific face by the front-view facial image. And a ModelMatch tool is developed to achieve it. The model match process has two stages: first, match the control points by radial shrink method proposed in this thesis. Then suppose the frame model is elastic and divided into several feature regions. Each region has its own adjustive rule. So the controlled points can be adjusted by the predefined control points according the rule to establish an individual facial model including texture.Finally, the thesis indicates the model-based coding implement in developing video chat system Grimace Chat, and discusses the prospect of future work.