Control structure efficiency enhancement for predictive video coding

Summary form only given. To achieve superior video compression performance it is generally necessary to base the encoding decisions on a local scale, instead of the traditional frame-by-frame approach. Partitioning schemes for optimal frame subdivision have been proposed. Unfortunately, a full local approach that leads to inhomogeneous partitioning of the frame encounters the serious problem of dealing with an increase of side information, particularly in the area of very low bit-rate compression where a relatively large overhead can possibly erode the benefits of the local analysis. Our approach adopts a moderate position, which can still retain some of the advantages of variable block size while keeping the overhead to a bare minimum. The idea is to subdivide a frame in a fixed number of square panels that we call "megablocks." All decisions relative to intra (I) or inter (P) coding are made at the megablock level. These decisions pertain to motion block size and degree of error quantization in the P-mode as well as vector and scalar quantization of subbands in the I-mode. The key to the improved performance of the megablock partitioning scheme is the joint cost minimization of motion field and quantized error information. In very low bit rate video coding, the cost of transmitting motion vectors consumes a significant fraction of the total bit budget. The megablock structure permits the use of variable size motion blocks without incurring the cost associated with motion segmentation or a full quadtree approach.