Facial Texture Coding at Low Bit Rates

This paper treats the topic of how to eeciently represent facial textures, for model based coding of human face images. The scheme is based on the Karhunen-Lo eve transform and involves an elaborate normalization process that enhances PSNR with around 0.5 dB compared to normalization with scaling, translation and rotation. It also seems to augment the perceived visual quality. Furthermore, a block tiling technique is introduced, improving quality to levels where comparison with waveform coders is worthwhile, showing a 5 dB improvement over JPEG. The nonstationarity of the signal is here exploited in a rate/distortion optimizing fashion, which further improves the quality by 0.4 dB. To further lower complexity for high quality images while preserving coding eeciency at very low bit rates, quadtree tiling of the eigenspace is introduced.