The focus of this research is on perceptually lossless or very low distortion, low-rate image coders using subband vector quantization and models of the human visual system (HVS). The use of visual models in the image coders attempts to remove the psychophysical redundancies present in the images in addition to the statistical redundancies that most image coders attempt to remove. Two different strategies that make use of the current knowledge of the HVS are being employed in the data compression systems: (1) Quantize the intensity image after processing it with a homomorphic visual model. The idea behind this approach is that the intensity image, when processed with the visual model, will look more like the image that the "eye sees" and therefore minimizing the distortion in the transformed image is appropriate. (2) The second approach is to define a masking function for the image to be quantized. This function defines threshold values for each pixel of the image such that distortions of magnitude smaller than the threshold function will not be perceived by the eye. The approach then is to develop a data compression system such that the quantization produces errors of magnitude smaller than this threshold function. The specific problems that are being studied are: (1) Development of a subband vector quantizer that is equipped with a multichannel, homomorphic model of the HVS. (2) Development of a subband vector quantizer that used a perceptual masking function. A combination of the two schemes, as well as incorporation of predictive vector quantization and/or optimum pre- post-processing filters into the subband coder is being explored. Preliminary work on extending the ideas to color image sequence coding problems are also being conducted.
