Abstract
In this paper, we propose a prototype of very low bit rate model-based image codec. It is designated to encode head-andshoulder image sequences, in which whose statistical and structural characteristics can be modeled based on the large amount of a priori knowledge. From the view of such image sequence coding, the subjects to be coded are mainly the 3-D head motions and the facial expressions. The encoder extracts the motion and expression parameters from input image sequences and transmits them to the decoder. The decoder synthesizes the output image sequences through the transmitted parameters. Clearly the transmitted information is much less than the original, that is, the compression ratio can be up to thousands of times. Therefore, it can be easily apply to the image sequence with large frame size, such as 512x512 or 640x480, etc. In our codec, a generic 3D head model with thousands of triangle polygons is used. For each image frame F of the given image sequence, we adjust the model by comparing feature points in the projected head model with that in F to construct a personalized pseudo-3D model for F. The block matching algorithm is applied to find the displacement vectors of feature points between two consecutive frames. These vectors will be coded and transitted to the decoder then. In decoding, we apply a simple 2D warping technique to mapping the polygonal textures (obtained in the first frame ) to reconstruct each image frame. In our experience, the total bitcount is less than lk per 640x480 frame if the residuals (for some acting areas) can be further coded to improve the quality. The residuals coding we used follows JPEG.