Apparent color picker: color prediction model to extract apparent color in photos

Abstract

A color extraction interface reflecting human color perception helps pick colors from natural images as users see. Apparent color in photos differs from pixel color due to complex factors, including color constancy and adjacent color. However, methodologies for estimating the apparent color in photos have yet to be proposed. In this paper, the authors investigate suitable model structures and features for constructing an apparent color picker, which extracts the apparent color from natural photos. Regression models were constructed based on the psychophysical dataset for given images to predict the apparent color from image features. The linear regression model incorporates features that reflect multi-scale adjacent colors. The evaluation experiments confirm that the estimated color was closer to the apparent color than the pixel color for an average of 70%–80% of the images. However, the accuracy decreased for several conditions, including low and high saturation at low luminance. The authors believe that the proposed methodology could be applied to develop user interfaces to compensate for the discrepancy between human perception and computer predictions.