AG 2, AG 4, MMCI   Info

AG Audience

English

45 Minutes

Saarbrücken

Automatically predicting the perceived difference between color images is useful in various image processing applications such as color image compression, halftoning or gamut mapping. In this talk, I will present a general framework for predicting the perceived difference between color images. It includes image normalization, feature extraction, and feature combination. Image normalization is performed by an image-appearance model that converts the pixel values to color attributes (lightness, chroma and hue) considering the viewing conditions (e.g. the viewing distance or the luminance level). The color attributes are arranged in a nearly perceptually uniform working color space with low cross-contamination. I will describe how to design this color space that plays a crucial role in the framework because it ensures that image features such as edges or gradients are not over- or underestimated, i.e., their computed magnitudes exceed their perceived magnitudes or vice versa. Finally, difference features based on the color attributes are extracted from the normalized images and combined to a single number reflecting the perceived image difference. I will describe how to test and compare the prediction performance of image-difference measures using visual data comprising gamut mapping distortions. At the end of the talk, I will show how to use the image-difference measure to optimize gamut mapping. These optimization results may be used, in turn, for further enhancing the measure. A multi-scale version of the resulting measure achieves the best agreement with human judgments on the TID2013 visual database comprising 3,000 distorted images and more than 500,000 paired comparisons.

Karol Myszkowski

--email hidden

passcode not visible

logged in users only