| ||Definition:|| ||Sub-band coding for images has roots in work done in the 1950s by Bedford and on Mixed Highs image compression done by Kretzmer in 1954. Schreiber and Buckley explored general two channel coding of still pictures where the low spatial frequency channel was coarsely sampled and finely quantized and the high spatial frequency channel was finely sampled and coarsely quantized. More recently, Karlsson and Vetterli have extended this to multiple subbands. Adelson et al. have shown how a recursive subdivision called a pyramid decomposition can be used both for compression and other useful image processing tasks.
A pure sub-band coder performs a set of filtering operations on an image to divide it into spectral components. Usually, the result of the analysis phase is a set of sub-images, each of which represents some region in spatial or spatio-temporal frequency space. For example, in a still image, there might be a small sub-image that represents the low-frequency components of the input picture that is directly viewable as either a minified or blurred copy of the original. To this are added successively higher spectral bands that contain the edge information necessary to reproduce the original sharpness of the original at successively larger scales. As with DCT coder, to which it is related, much of the image energy is concentrated in the lowest frequency band.
For equal visual quality, each band need not be represented with the same signal-to-noise ratio; this is the basis for sub-band coder compression. In many coders, some bands are eliminated entirely, and others are often compressed with a vector or lattice quantizer. Succeedingly higher frequency bands are more coarsely quantized, analogous to the truncation of the high frequency coefficients of the DCT. A sub-band decomposition can be the intraframe coder in a predictive loop, thus minimizing the basic distinctions between DCT-based hybrid coders and their alternatives.|