FEATURE-SPECIFIC VECTOR QUANTIZATION OF IMAGES

A new interband vector quantization of a human vision-based image repr esentation is presented, The feature-specific vector quantizer (FVQ) i s suited for data compression beyond second-order decorrelation, The s cheme is derived statistical investigations of natural images and from from the processing principles of biological vision systems, The init ial stage of the coding algorithm is a hierarchical, size- and orienta tion-selective, analytic bandpass decomposition, realized by even- and odd-symmetric filter pairs that are modeled after the simple cells of the visual cortex. The outputs of each even- and odd-symmetric filter pair are interpreted as real and imaginary parts of an analytic bandp ass signal, which is transformed into a local amplitude and a local ph ase component according to the operation of cortical complex cells, Fe ature-specific multidimensional vector quantization is realized by com bining the amplitude/phase samples of all orientation filters of one r esolution layer, The resulting vectors are suited for a classification of the local image features with respect to their intrinsic dimension ality, and enable the exploitation of higher order statistical depende ncies between the subbands, This final step is closely related to the operation of cortical hypercomplex or end-stopped cells, The codebook design is based on statistical as well as psychophysical and neurophys iological considerations, and avoids the common shortcomings of percep tually implausible mathematical error criteria, The resulting perceptu al quality of compressed images is superior to that obtained with stan dard vector quantizers of comparable complexity.