Image sequence coding by vector quantization

The authors investigate image sequence coding by means of a classified vector quantization technique in which 3-D cubes of the image sequence are coded. A four-image-color sequence (512×768×4) is divided into 32×32×4 or smaller cubes, which are classified as to horizontal, vertical, and temporal activity. Once the final form of the cube has been determined, a 3-D discrete cosine transform is applied, and the resulting coefficients are zone-sampled. These coefficients are then vector-quantized, and the resultant codebooks are optimized by both the LBG (Linde-Buzo-Gray) method and a simulated annealing technique. The average bit rate is in the range of 0.34-0.45 bits/pixel for active sequences with signal/noise ratio values (peak signal to root-mean-square noise) in the range of 36 to 41 dB overall