Existing image-difference measures show excellent accuracy in predicting distortions, such as lossy compression, noise, and blur. Their performance on certain other distortions could be improved; one example of this is gamut mapping. This is partly because they either do not interpret chromatic information correctly or they ignore it entirely. We present an image-difference framework that comprises image normalization, feature extraction, and feature combination. Based on this framework, we create image-difference measures by selecting specific implementations for each of the steps. Particular emphasis is placed on using color information to improve the assessment of gamut-mapped images. Our best image-difference measure shows significantly higher prediction accuracy on a gamut-mapping dataset than all other evaluated measures.

Image processing methods that utilize characteristics of the human visual system require color spaces with certain properties to operate effectively. After analyzing different types of perception-based image processing problems, we present a list of properties that a unified color space should have. Due to contradictory perceptual phenomena and geometric issues, a color space cannot incorporate all these properties. We therefore identify the most important properties and focus on creating opponent color spaces without cross contamination between color attributes (i.e., lightness, chroma, and hue) and with maximum perceptual uniformity induced by color-difference formulas. Color lookup tables define simple transformations from an initial color space to the new spaces. We calculate such tables using multigrid optimization considering the Hung and Berns data of constant perceived hue and the CMC, CIE94, and CIEDE2000 color-difference formulas. The resulting color spaces exhibit low cross contamination between color attributes and are only slightly less perceptually uniform than spaces optimized exclusively for perceptual uniformity. We compare the CIEDE2000-based space with commonly used color spaces in two examples of perception-based image processing. In both cases, standard methods show improved results if the new space is used. All color-space transformations and examples are provided as MATLAB codes on our website.

Small suprathreshold color differences around five CIE color centers were investigated on a typical liquid crystal display (LCD) with fluorescent backlight using the method of constant stimuli. The results were evaluated using probit analysis and compared with surface-color differences of the RIT-DuPont dataset. We focused especially on the relationship between T50 distances obtained from LCD and surface-color stimuli and on the influence of the display's narrowband primaries and its relatively low luminance level on interobserver uncertainty. The low luminance level of the LCD decreases the perceived color differences. However, considering the visual uncertainty of the experimental data, we could not reject the hypothesis that T50 distances from the RIT-DuPont and our experiment agree up to a constant scaling factor. In addition, we found significantly higher interobserver variability in the estimation of small color differences if the colors are viewed on an LCD. There are some indications that color-difference perception might be influenced by individual color-matching functions and, thus, by the spectral power distribution of the stimuli. We provide the experimental data, including all spectral stimuli shown to the observers, on our website.

We propose a method to improve the prediction performance of existing color-difference formulas with additional visual data. The formula is treated as the mean function of a Gaussian process, which is trained with experimentally determined color-discrimination data. Color-difference predictions are calculated using Gaussian process regression (GPR) considering the uncertainty of the visual data. The prediction accuracy of the CIE94 formula is significantly improved with the GPR approach for the Leeds and the Witt datasets. By upgrading CIE94 with GPR we achieve a significantly lower STRESS value of 26.58 compared with that for CIEDE2000 (27.49) on a combined dataset. The method could serve to improve the prediction performance of existing color-difference equations around particular color centers without changing the equations themselves.

No abstract