Perceptual Fidelity for Digital Color Imagery

摘要

The problem of measuring the fidelity of digital color images in a manner thatcorresponds to human perceptual assessments is addressed. Experiments are performed to validate human visual system (HVS) models, which provide access to a 'perceptual space' in which visual distortions may be measured, and then a model is proposed for assessing the perceptual fidelity of digital color image. Color Mach bands are produced in the first experiment, demonstrating that, as in the brightness channel, low spatial frequency attenuation occurs in the chromatic channels of the HVS. In the second experiment, a correlation between the chromatic channels of the HVS model and color discrimination axes of color blind observers is demonstrated. Removing variation from one of the chromatic channels of a natural image produces a color-distorted image which the color blind subjects cannot distinguish from the original. Removing variation from the other chromatic channel produces an image that appears colorful to normally-sighted observers, but monochrome to the color blind observers. The third experiment shows that a Gabor filter-based HVS model produces illusory contours in several illusory contour stimuli. These results provide a unique validation of multiple-channel HVS models which process the image in multiple spatial frequency bands that are tuned to match measured sensitivities of neurons in the primary visual cortex of cats and monkeys. Finally, the multiple-channel processing used in the illusory contour experiment is combined with the color vision model from the first two experiments to produce a multiple-channel, color HVS model for measuring perceptual fidelity of color images. A demonstration of the model shows that the structure of the new model is correct. However, inaccurate parameter values for the multiple-channel processing of the chromatic channels cause over-predict.