FAP-Based Subtle Expression Synthesis

To synthesize realistic expression animation, subtle expression (for short SE, such as wrinkle, crow's-feet and dimple, etc.) is a crucial factor. But the only geometric transform of the facial feature is considered in the traditional expression synthesis, completely ignoring texture variations on face. In the current studies on the subtle expression synthesis, only SE from extracted expression samples can be synthesized, so their reality is poor. This dissertation investigates SE synthesis of facial animation in this problem. The works of the dissertation are as follows: Firstly, it has carried out to obtain the SE and the facial motion synchronization data by feature-tracking method. The accuracy motion data according with MPEG-4 and avoiding the expensive equipments can be obtained, and face animation parameters (namely FAP) can be extracted from video. The automatic active shape model and the manual adjusting are adopted that can easily get the synchronization data. These assure credible data for setting up SE synthesis model. Secondly, a PCA-based parameterized method of expression ratio image parameters (for short ERIP) about subtle expression data is proposed. And calculate method of ERI is improved on. Many experiments show that the proposed expression ratio image parameter method can effectively extract subtle expression feature on a large amount of facial subtle expression data. Finally, a SVR-based method as a kind of optimized regression is proposed for SE synthesis model from FAP to SE parameters. This dissertation discusses this model based on four different SVR methods. Then, ε-insensitive SVM with radial basis function as kernel is adopted as prototype of this model. In subtle expression synthesis, the robustness of this model is shown when forecasting ERIP. Based on above mentioned works, SE synthesis model is accomplished, and it can generate statistically realistic expression and fit into MPEG-4. Firstly, the expression synthesis system based on this model is implemented; secondly, this model is used in mobile terminal; finally, in human facial illustrations, a simple application of this model is given. This model works well and satisfies the user requirement.