CIE-Lab Color Space

• CIE-Lab Color Space is a device-independent color model that expresses colors using L*, a*, and b* values to achieve perceptual uniformity.
• It forms the basis for skin tone measurement and fairness evaluation in diverse fields such as computational imaging and dermatology.
• Key methodologies include converting sRGB readings to CIE-Lab, employing Delta E and ITA metrics, and calibrating imaging devices for robust analysis.

The CIE-Lab (CIELAB) color space is a device-independent, perceptually uniform colorimetric model defined by the International Commission on Illumination (CIE) in 1976. Designed for consistent quantification and comparison of surface colors regardless of device or lighting, CIE-Lab is foundational in research related to color measurement, computational imaging, and algorithmic fairness in medical and biometric applications. Its central property is perceptual uniformity: Euclidean distances in the space correspond, as closely as possible, to human visual difference perception.

1. Core Definition and Mathematical Structure

CIE-Lab expresses color as a three-dimensional vector $(L^*, a^*, b^*)$ defined with respect to a reference white illuminant, most often D65. The axes are:

• $L^*$: Lightness ($0= \text{black}, 100= \text{white}$)
• $a^*$: Green ($-a^*$) to red ($+a^*$)
• $b^*$: Blue ($-b^*$) to yellow ($+b^*$)

The CIE-Lab space is derived from the CIE XYZ tristimulus values by nonlinear transformation:

$$\begin{align*} L^* &= 116\,f\left(\frac{Y}{Y_n}\right) - 16 \ a^* &= 500 \left[ f\left(\frac{X}{X_n}\right) - f\left(\frac{Y}{Y_n}\right) \right] \ b^* &= 200 \left[ f\left(\frac{Y}{Y_n}\right) - f\left(\frac{Z}{Z_n}\right) \right] \end{align*}$$

where $L^*$0 for $L^*$1 and $L^*$2 otherwise; $L^*$3 are the tristimulus values for a white reference (Cook et al., 2024).

2. Perceptual Uniformity and Delta E Metrics

The dominant property of CIE-Lab is its effort towards perceptual uniformity. The Euclidean distance in this space, $L^*$4, is widely adopted as a metric for perceived color difference:

$L^*$5

Empirical repeatability of well-calibrated colorimetric devices for human skin yields noise limits on $L^*$6 (CIELAB units) at 95% confidence, setting a lower bound on colorimetric accuracy in biomedical skin measurement (Cook et al., 2024).

3. Applications in Skin Tone Measurement and Fairness Evaluation

CIE-Lab underpins objective skin tone measurement in both dermatology and fairness auditing for AI, facilitating continuous or ordinal labeling. Standard protocols, such as those utilizing the Cortex Technology DSM III Colorimeter, entail measurement of skin loci and direct conversion of raw sRGB readings to CIE-Lab under D65. Representative parameters extracted include:

• Lightness ($L^*$7): direct surrogate for skin brightness or reflectance
• Hue: defined as $L^*$8
• Chroma: magnitude $L^*$9

Regression models leveraging these parameters elucidate the correspondence between traditional, subjective ratings (e.g., Fitzpatrick Skin Type, Monk Skin Tone) and objective colorimetry, revealing substantial biases and low explained variance for subjective scales (FST: $0= \text{black}, 100= \text{white}$0; contextual mean error well above $0= \text{black}, 100= \text{white}$1) (Cook et al., 2024).

4. Image-derived Skin Color Metrics and Robustness Considerations

Many computational strategies estimate skin phenotype from images by mapping detected skin regions' sRGB values to CIE-Lab and summarizing $0= \text{black}, 100= \text{white}$2 or, more robustly, the Individual Typology Angle (ITA):

$0= \text{black}, 100= \text{white}$3

Lighting, camera calibration, and skin masking quality strongly influence the accuracy of image-derived CIE-Lab estimates. Only protocols with rigorous control over lighting, device standardization, and background correction approach instrument-grade reliability (corrected image protocols: $0= \text{black}, 100= \text{white}$4 with ground truth; uncontrolled: $0= \text{black}, 100= \text{white}$5) (Howard et al., 2021). Patch-based and uncalibrated workflows show systematic lighting biases, which require calibration or result in high estimation error (Benčević et al., 6 Apr 2025).

5. Mapping CIE-Lab to Skin Tone Scales: Continuous and Ordinal Strategies

CIE-Lab enables creation of both continuous and binned skin tone scales, transcending the subjectivity of legacy ordinal scales. The Individual Typology Angle (ITA), derived solely from CIE-Lab parameters, is mapped to Fitzpatrick types using established thresholds (e.g., I: ITA$0= \text{black}, 100= \text{white}$6, VI: ITA$0= \text{black}, 100= \text{white}$7 (Benčević et al., 10 Feb 2026)). The CST (Colorimetric Skin Tone) scale, introduced for increased sensitivity and reduced bias, directly bases its intervals on empirical CIE-Lab quantiles and demonstrates higher correlation with measured pigmentation than text- or palette-based scales (Cook et al., 2024).

6. Methodological Limitations and Bias Sources

Despite its mathematical rigor, application of CIE-Lab in phenotyping or fairness auditing is sensitive to methodological detail. Uncontrolled imaging environments introduce lighting and device-dependent biases. Manual assignments based on images, even after conversion to CIE-Lab, do not reach parity with calibrated, in-person measurement (Howard et al., 2021). In skin tone annotation, crowd-sourced protocols can yield high inter-rater reliability, but algorithmic annotation against CIE-Lab/ITA is less consistent with clinical ground truth (Groh et al., 2022).

7. Research Trends and Recommendations for Fair Skin Color Measurement

Current research critiques ordinal, UV-response-based scales (e.g., FST) for their limited colorimetric sensitivity and context/race-related biases, advocating for direct use of CIE-Lab-derived continuous and ordinal skin color metrics (Cook et al., 2024, Benčević et al., 10 Feb 2026). Synthetic datasets with controlled melanin and lighting (e.g., S-SYNTH) facilitate comparative benchmarking of CIE-Lab-based methods for robustness to extrinsic factors (Benčević et al., 6 Apr 2025). It is recommended to calibrate device pipelines, employ full-scene or region-based color correction, and rely on CIE-Lab/ITA-driven annotation in large-scale fairness analyses whenever feasible.

---

References:

(Cook et al., 2024, Howard et al., 2021, Benčević et al., 10 Feb 2026, Benčević et al., 6 Apr 2025, Groh et al., 2022)