Face-to-face video teleconferencing is very important for real timefl

communication. Current teleconferencing application uses standardfl

video codec, such as MPEG1/2/4, for the compression offl

face video. It either requires high bandwidth for high quality videofl

transmission, or the transmitted face video be blurred at low bitrate.fl

In this paper, we present a system for real-time coding offl

face video at low bit-rate. There are two main contributions. First,fl

we improve the technique of long term memory prediction by selectingfl

frames into the database in an optimal way. A new framefl

is selected into the database only when it is significantly differentfl

from those frames which are already in the database. In this way,fl

the database can cover a wider range of images. Second, we incorporatefl

the prior knowledge about faces into the long term memoryfl

prediction framework. The prior knowledge includes: (1) facialfl

motions are repetitive such that most of them can be reconstructedfl

from multiple reference frames; and (2) different components offl

the face and the background could tolerate different level of errorfl

because of different perceptual importance. Experiments showfl

that at similar PSNR the proposed system works much faster andfl

achieves better visual quality than standard H.264/JVT codec.

PDF file

In  Proc. of the IEEE Int. Conf. on Multimedia and Expo

Details