| Summary: | 碩士 === 國立成功大學 === 電機工程學系碩博士班 === 90 === In order to synthesize and simulate face activity and expression, MPEG-4 defined the face object. MPEG-4 specified three types of data for a face object: the Facial Animation Parameters (FAPs), Facial Definition Parameters (FDPs) and FAP Interpolation Table (FIT).
This thesis proposes a method for simplifying FAPs. The method is based on the statistics obtained by analyzing feature points. The statistics is divided into two types: global and local. The global statistics is concerned with extracting high-activity feature points of an entire head model. The proposed method employs block-matching to calculate the motion vector of a global feature for each frame. If the magnitude of a motion vector is above the specified threshold then it is regarded active; otherwise, not active. According to the global statistics, 3 feature points: head pitch, yaw, and roll are selected to constitute a T shape coordinate system for a face. The corresponding MPEG-4 FAP for these 3 global feature points are then coded. In local statistics, we focus on the mouth and eyebrows area. From global statistical results and the FAPs defined in MPEG-4 we select the above two areas to obtain local statistics. The local statistics is concerned with the local movement of facial muscle and the correlation between them. The proposed method employs the warping concept to gather the motion statistics of local features. According to the local statistics, 5 corresponding MPEG-4 FAPs are selected and other FAPs are interpolated according to their local correlation with the former. The proposed system totally uses 8 FAPs to control a synthetic face model, including 3 global FAPs and 5 local FAPs.
To test the quality of the synthesized face expression we compare the PSNR between using 68 FAPs and using 8 FAPs. The experiments showed that the PSNR of the proposed system is above 30 dB and the bit-rate is less than 1.38 kbits/sec.
|
|---|
