Adaptive perceptual coding of visual information

This dissertation is motivated by potential applications of models of human vision to improve the performance of still image and video compression techniques. It addresses the problem of local adaptation of the quantization stages based on just noticeable distortion (JND) thresholds.;Existing perceptual coders are not able to fully exploit the local variations of the perceptual masking provided by the visual information., This dissertation presents a new feedback paradigm for perceptual image coding which estimates the JND thresholds based on the already coded data. The proposed perceptual coding schemes include scalable techniques and schemes that support compression with a fixed perceptual target distortion as well as compression with a fixed target bit-rate.;Using transformations that decompose the signal into components of different spatial frequency and orientation, the new coders derive local perceptual masking thresholds that are based on models of human visual properties. Based on the perceptual feedback paradigm, a locally-adaptive perceptual quantization scheme is developed that does not require additional side information. The resulting locally-adaptive quantization technique can be applied with subband-based image decompositions as well as with block-based image transform methods. An adaptive perceptual image coder that relies on the new quantization technique is designed for a 4 x 4 subband decomposition.;The performance of coders based on the new adaptive quantization technique can be improved by introducing a perceptual distortion measure which is based on probability summation. An adaptive coder with the perceptual distortion measure is designed for the 8 x 8 block-based DCT, and it is shown that the feedback paradigm for perceptual coding of visual information results in perceptible coding gains.;Additional work is presented that expands perceptual coding to embedded zerotree coding. It is shown that embedded zerotree coding of perceptually weighted transform coefficients results in minimization of a distortion measure that is more relevant with respect to human perception than mean-square error. The performance of embedded zerotree coding in terms of perceived image quality is improved and a spatially scalable extension of embedded zerotree coding is introduced.;The dissertation also proposes a perceptual video coding architecture which extends the perceptual feedback paradigm to hybrid video coding and allows locally-adaptive perceptually optimized motion prediction as well as locally-adaptive perceptual motion residual coding.