Image-Individualized Gamut Mapping Algorithms

czuk Zofia Barań, Peter Zolliker, Joachim GiesenJournal of Imaging Science and Technology 2010 · DOI

Abstract

In this article, the authors demonstrate that image quality measures (IQMs) can be successfully utilized to develop image-individualized gamut mapping algorithms (GMAs). By comparing various IQMs against psychovisual data from four recent studies, the researchers found that certain automated measures—specifically the Structural Similarity Index (SSIM)—predict human preferences as accurately as traditional scaling methods like Thurstone’s Law of Comparative Judgement. This is significant because IQMs operate independently of experimental data. The best-performing measure was used to create a "meta-algorithm" that automatically selects the optimal GMA for any given image, outperforming fixed, single-algorithm approaches.

Introduction

Every display and printing device possesses a unique "gamut," defined as the range of colors it can physically reproduce. When an image contains colors outside a target device's gamut, a Gamut Mapping Algorithm (GMA) is required to compress those colors into the available space. Historically, no single GMA has proven optimal for all images; some images require the preservation of vibrant hues, while others depend on subtle lightness gradients or fine textures.

Traditionally, identifying the most effective GMA for a specific image required exhaustive psychovisual testing, where human observers manually compared hundreds of image pairs. This process is time-consuming, expensive, and difficult to scale. This research addresses the need for an automated system capable of predicting human preference to select the best mapping strategy on an image-by-image basis.

Methodology

The researchers evaluated five distinct Image Quality Measures (IQMs) to determine their correlation with human observer choices:

  1. SSIM (Structural Similarity Index): Evaluates changes in luminance, contrast, and structure.
  2. MSE (Mean Square Error): A standard point-by-point difference in lightness.
  3. $Δ E$: The Euclidean distance between colors in the CIELAB space.
  4. LMSE (Laplacian Mean Square Error): A local measure focusing on edge differences.
  5. $Q_{LC}$ (Local Contrast): Based on the Michelson contrast of image patches.
Example test image 'a' used in the psychovisual studies to establish human preference ground truth.

These measures were validated using a "hit rate" metric—the percentage of times the automated measure correctly predicted the human-preferred image in a pair. The validation utilized data from four studies:

  • Basic Study (BS): 97 images with seven algorithms.
  • Image Gamut (IG): 75 images focusing on source gamut descriptions.
  • Local Contrast (LC): 77 images exploring detail enhancement.
  • Individual Study (IS): 20 specific images where comparisons were repeated 40 times to establish a highly reliable ground truth.

Results and Discussion

The study revealed that pointwise measures like MSE and $Δ E$ performed poorly, often yielding hit rates near 0.5 (equivalent to random guessing). This suggests that human preference in gamut mapping is not driven by simple color differences, but by the preservation of spatial details and textures.

Example test image 's' featuring high-saturation colors and complex textures, where structural preservation is critical.

SSIM emerged as the clear winner, achieving the highest hit rates across all datasets. In the Individual Study (IS), SSIM's performance approached the theoretical maximum hit rate (0.8 to 0.85) dictated by human inconsistency. Remarkably, SSIM—a "blind" measure requiring no human training data—predicted choices nearly as well as the Thurstone model, which is statistically derived from the observer data itself.

By implementing a meta-algorithm that calculates the SSIM for several candidate GMAs and selects the one with the highest score, the researchers achieved perceived quality levels significantly higher than those produced by any single fixed GMA, such as SGCK or HPminDE.

Conclusion

This research proves that image-individualized gamut mapping is superior to universal approaches. While the current study primarily focused on the lightness (L) coordinate of the CIELAB space, future work may involve developing a "generalized SSIM" that incorporates chroma and hue dimensions. Although calculating multiple IQMs adds computational overhead, the resulting gain in image quality provides a compelling case for automated, individualized color management pipelines.

Summary updated

Abstract

In this article the authors show that image quality measures can be successfully used to develop image-individualized gamut mapping algorithms. First the authors compare different image quality measures for the gamut mapping problem and then validate them using psychovisual data from four recent gamut mapping studies. The scoring function used to validate the quality measures is the hit rate, i.e., the percentage of correct choice predictions on data from psychovisual tests. Some of the image quality measures predict the observer's preferences as good as scaling methods such as Thurstones method, which is used to evaluate the psychovisual tests. This is remarkable because the scaling methods are based on the experimental data, whereas the quality measures are independent of these data. The best performing image quality measure is used to automatically select the optimal gamut mapping algorithm for an individual image.
BibTeX
@article{baran2010image,
  author = {czuk Zofia Barań and Peter Zolliker and Joachim Giesen},
  title = {{Image-Individualized Gamut Mapping Algorithms}},
  journal = {Journal of Imaging Science and Technology},
  year = {2010},
  doi = {10.2352/j.imagingsci.technol.2010.54.3.030201},
}
Figure