Visual detection and perception of distortion in wavelet-compressed images

Image compression has played a pivotal role in facilitating the widespread use of visual-communication and digital-imaging technologies. Lossy image compression, in particular, has made possible the rapid exchange and storage of visual information over limited-bandwidth and limited-memory channels. Such algorithms operate by exploiting both redundancy in the data and properties of the end-user; namely, when the end-user is a human, lossy compression is guided by properties of the human visual system. This thesis presents our work in uncovering some of those properties, and then using the results to design better quantization strategies.; Eight psychophysical experiments, designed to investigate visual detection and perception of wavelet subband quantization distortions, are presented. To quantify differences in detection of traditional targets versus detection of wavelet distortions, Experiments I and II measured detection thresholds for simple wavelet distortions both in the unmasked condition and when masked by natural images. To quantify summation of responses to wavelet distortions, Experiments III and IV measured detection thresholds for compound wavelet distortions in the unmasked condition and when masked by natural images. To further investigate masking, Experiments V and VI measured detection thresholds for wavelet distortions in the presence of textures and radiographs, respectively. Finally, to quantify the effects of natural images on the perception of suprathreshold distortions, Experiments VII and VIII measured perceived contrasts of wavelet distortions both in the unmasked condition and when masked by natural images. These experiments revealed that images induce frequency- and image-selective effects on detection and perception of wavelet distortions, effects which preclude the direct application of commonly used psychophysical results for compression.; Based on these psychophysical data, this thesis presents three contrast-based quantization strategies, and an algorithm for predicting detection thresholds, for use in wavelet-based image compression. The DCQ-LS algorithm, designed for visually lossless compression of natural images, quantizes subbands such that the contrasts of the quantization-induced distortions exhibit the experimentally determined threshold contrasts in the reconstructed image. The DCQM-LS algorithm uses a similar approach for visually lossless compression of medical images. For suprathreshold compression, the DCQ algorithm is presented, which attempts to provide the best-looking image given constraints on the bit-rate.