From Color Image Difference Models to Image Quality Metrics

Color difference research has recently culminated in the creation of the CIEDE2000 color difference equation. This equation has been shown to accurately predict perceived differences between simple stimuli on a solid background. A similar metric is desired to predict differences of complex stimuli, such as color images. Such a metric would be useful for predicting both threshold and supra-threshold differences. Threshold prediction is valuable for determining if two images are perceptually different, such as an original and a compressed image. Supra-threshold prediction can determine the magnitude of differences between two images, a precursor to perceived image quality. There are several existing spatial-contrast threshold models, though these models tend to ignore color and magnitude information. This presentation details a modular framework for a color image difference metric, based upon CIE color difference formulas. Some of the modules described include spatial filtering (similar to S-CIELAB), spatial frequency adaptation, local attention filtering, local and global contrast compensation, and visual masking. The ability to predict both threshold and magnitude errors can be combined with traditional psychophysical image scaling experiments to create a metric of perceived image quality.