Vector Quantization and Signal Compression

S-TREE: self-organizing trees for data clustering and online vector quantization

“…S-'l'REEl, like other tree-structured elustcring algorithms (Gersho & Gray, 1992;Hoffmann & Buhmann, 1995;Held & Buhmann, 1998), is suboptimal in the sense that the leaf node selected by the algorithm is not neeessarily the one dosest to the input vector. This occurs because branching at the higher levds of the tree biases the search, which may cause data points to be assigned to wrong dusters or weight vectors not to correspond to the duster centers.…”

“…Vector qua.ntiz;crB were first tested on the classical Gauss-I'via.rkov source benclunark, with construction f(lllowing Gersho and Gray (1992). Training sets were processed with input dimensions M = 1, 2, ... , 7, with each training set.…”

“…In order to overcome the computational burden associated with unconstrained clustering, structural constraints such as lattices and trees have been proposed (see Gersho and Gray (1992) for a review). In particular, tree-structured clustering methods have become popular in the vector quanti:;;ation literature.…”

Normalized image representation for efficient coding

“…The sign (or phase) of each coefficient is inherited from the sign of the corresponding linear coefficient. Note that this scheme is similar to the one used in transform coding, where first a linear transform is used to reduce the statistical dependence between the samples of A and then some additional non-linearity may be considered in order to simplify the quantizer design [3], [4].…”

“…In the conventional approach to transform coding, perceptual factors are taken into account only after the selection of the representation, in the quantizer design. Moreover, in order to apply the standard theory for bit allocation, the (perceptual) metric has to be diagonal in the representation to be quantized [4]. However, the above linear transforms do not completely achieve the desired independence from both points of view [2], [6].…”