INTELLIGENT IMAGE INTERPRETATION FOR HIGH-COMPRESSION HIGH-QUALITY SEQUENCE CODING

Video transmission at very low bit rate has got growing attention in recent years. In this paper we propose an approach for the compression of 144 x 176 pixels Q-CIF video conference sequences. The compression ratio well exceeds 200 : 1 (thus leading to bit-rates under 10 kbit/s for 10 frames/s) with very good psycovisual quality of the reconstructed images. The algorithm integrates and improves different feature extraction and image coding techniques. At first the speaker's head is detected by means of active snakes. A new form of internal energy is defined that allows a very robust and fast head tracking. After head detection, internal facial features (i.e., eyes, nose, and mouth) are located by means of a new algorithm. The image is decomposed into different parts with different psicovisual relevance. This information is used to guide in an intelligent way the subsequent processing of the motion compensated difference image. The important areas are coded more accurately while the less relevant areas are coded in a coarser way. This approach grants very high compression while the image quality remains high. Subpixel block matching is used to obtain a motion compensated difference image. This image is segmented into homogeneous regions that are then coded by means of a technique based on differential chain code.