Large block VQ for image sequences

Rate distortion theory indicates that a well-defined signal source can be compressed closely to the rate distortion bound, provided that the coding block length is large enough. This is the motivation of vector quantization (VQ) that quantizes a group of signals together as a vector. However, due to its exponentially increasing computational complexity, in the application area of image/video processing almost all unconstrained VQ work has been constrained to a vector size of 4×4. As the computer becomes much faster, it becomes computationally affordable to increase this standard vector dimension. The authors investigate the effect on VQ performance when the vector dimension is increased. Experimental results show that a significant improvement can be achieved by increasing the vector dimension from 4×4 to 8×8. Both the PSNR and visual performance are improved by increasing the vector dimension, but the computational complexity and codebook overhead both increase exponentially. Several algorithms have been tested to reduce the computation and a new approach to reduce the codebook overhead has been developed. Experiments show that both the computational complexity and the codebook overhead can be reduced significantly without losing much performance. Also, a new coding scheme-large block VQ with shift-is introduced as a way to improve the performance further