Abstract: We present methods to estimate perceptual uniformity of color spaces and to derive a perceptually uniform RGB space using geometrical criteria defined in a logarithmic opponent color representation.
Introduction
In digital imaging, a perceptually uniform color space is one where a mathematically equal change in color values corresponds to a perceptually equal change in human vision. Standard RGB spaces, such as sRGB, often lack this property; for instance, adjusting brightness can result in unintended hue shifts. While models like CIELAB attempt to address this through complex biological modeling of the human visual system, this research explores whether perceptual uniformity can be achieved through pure geometric optimization. The goal is to derive an RGB space where hue lines are straight, angles between hues are uniform, and chroma distances are consistent.
Methodology
The researchers approached the derivation of a uniform color space as a large-scale optimization problem. The core of their method involves a logarithmic opponent representation, which converts RGB values into a space invariant to brightness and system non-linearities (gamma). In this representation, hue is defined by simple ratios, simplifying the geometric analysis.
To identify the optimal 3x3 matrix for converting XYZ color data into the new RGB space, the team employed a spherical sampling technique. This allowed for the exhaustive testing of over 300 million potential linear transforms by sampling primary color configurations on a sphere.
The optimization process utilized three primary geometric criteria:
- Hue Constancy: Ensuring color data points for a single hue fit onto a straight line using Singular Value Decomposition (SVD).
- Hue Angle Uniformity: Spacing the 40 principal hues at equal 9° intervals.
- Chroma Uniformity: Maintaining consistent distances between adjacent color samples across the space.
The study utilized the Munsell Renotation Data and the Hung and Berns Data as benchmarks for perceptual equality.
Results and Discussion
The optimization yielded three specific transforms, each excelling in different geometric criteria. These optimized spaces demonstrated significant improvements over standard encodings like sRGB and ROMM. For example, "Transform 2" achieved a hue angle variance of 0.0021, which is substantially better than sRGB (0.0242) and even outperforms more complex models like IPT (0.0045).
A key insight from the study is that perceptual uniformity is highly dependent on the training dataset. The "optimal" space shifted depending on whether Munsell or Hung and Berns data was used, suggesting that perceptual uniformity is somewhat subjective. Furthermore, the results indicate that geometric "straightening" of the color space can be as effective as complex biological modeling.
Limitations and Future Directions
A significant hurdle for practical implementation is gamut practicality. The most mathematically uniform transforms often result in "imaginary" primary colors or gamuts that do not align well with physical display or sensor capabilities. Consequently, while these spaces are ideal for color difference measurements and mathematical corrections, they are not yet suitable as standard image encodings. Future research will focus on incorporating gamut extent as an additional optimization criterion and exploring power functions as alternatives to logarithmic representations to further enhance space uniformity.
Summary updated