Colorimetric Skin Tone Scale

• The CST Scale is a quantitative framework that defines human skin tone using empirical CIELAB data and standardized imaging protocols.
• It employs calibrated colorimeter readings, regression-fitted swatches, and clustering algorithms to minimize subjectivity and improve fairness.
• CST is applied in clinical assessments, dataset annotation, and AI fairness, demonstrating higher reproducibility and lower bias than legacy methods.

The Colorimetric Skin Tone (CST) Scale is a quantitative framework for assessing and categorizing human skin tone using precise colorimetric measurements, predominantly in CIELAB or related color spaces. Unlike older scales based on perception or text-based questions, CST approaches explicitly ground skin tone class boundaries and scale swatches in empirical colorimetric data. CST has been implemented using both direct colorimeter readings, calibrated image analysis pipelines, and data-driven methods that decompose image color using physical reflection models. The CST framework is widely used to reduce subjectivity and demographic bias in clinical assessment, image dataset annotation, and AI model fairness analysis.

1. Conceptual Foundations of the CST Scale

The CST scale’s core design principle is the direct mapping between measured skin color and class labels, using device-independent colorimetric spaces. Most CST implementations utilize CIELAB coordinates, with $L^*$ as the principal lightness axis and $a^*$, $b^*$ encoding chromatic subtleties. Early motivation for CST arises from the limitations of legacy scales (Fitzpatrick Skin Type, Monk Skin Tone) that rely on self-report or visual matching—modes demonstrably influenced by context, race, and device variation (Cook et al., 2024).

CST can be instantiated as a discrete palette (with each bin anchored by an empirically sampled $L^*,a^*,b^*$ trio), as a continuous normalized value (e.g., min-max or z-score of reflectance), or as a multidimensional index (e.g., tone plus hue angle) (Thong et al., 2023). In all forms, CST is grounded in the measurement of light reflected from the skin under standardized illumination.

2. CST Scale Construction and Empirical Basis

The paradigmatic reference for CST construction is large-scale, in-vivo colorimeter measurement. In one canonical protocol, bilateral readings (dorsal hand and facial zygomatic arch) are made for each volunteer using a calibrated DSM III colorimeter under D65 illumination. Mean $L^*,a^*,b^*$ coordinates are computed per subject and used both for palette design and annotation ground-truth (Cook et al., 2024). Scale swatches are derived by fitting quadratic regression to the real distribution of measured skin colors (hue and chroma as smooth functions of $L^*$), generating typically 10 bins at uniform $L^*$ intervals (e.g., $L^*=70$ for lightest to $L^*=20$ for darkest), each with its calculated $a^*,b^*$ and corresponding color patch.

Discrete class boundaries in CST are thus empirically set, not via perception, and the scale covers the full gamut of real human skin color, including intermediate tones poorly represented in alternatives (e.g., MST). No ad-hoc or perceptual thresholds are inserted; every bin is anchored in colorimeter data (Cook et al., 2024).

For imaging applications, CST scale construction may leverage dominant-cluster extraction and perceptual color differences—using CIEDE2000 ($\Delta E_{00}$) as a mapping metric between clustered, measured skin color and discrete reference blocks (Alyoubi et al., 20 May 2025).

3. Algorithmic Pipelines for Automated CST Extraction

Several computational pipelines for CST extraction from images have been demonstrated:

• Smartphone Colorimetry and ITA-based CST: A validated workflow uses a high-spec smartphone (e.g., iPhone 11) with controlled geometry and lighting, disables all automatic corrections, and captures calibrated sRGB images of anatomical sites (dorsal/palmar finger) (Burrow et al., 2024). Pixels are converted from sRGB to CIE XYZ, then to CIELAB, and finally an Individual Typology Angle (ITA) is computed per pixel: $$\mathrm{ITA}(x,y) = \arctan \left( \frac{L^* - 50}{b^*} \right) \cdot \frac{180}{\pi}$$. The mean ITA over a region of interest supports assignment to CST bins—e.g., “Very light” ($50^\circ < \mathrm{ITA} \le 90^\circ$) through “Very dark” ($-90^\circ \le \mathrm{ITA} \le -50^\circ$) (Burrow et al., 2024).
• Cluster-based Facial Skin Tone Classification: In high-throughput annotation or AI applications, face regions are detected and segmented, non-skin elements are masked, and dominant skin tone is extracted via clustering—e.g., X-means in HSV space, followed by conversion to CIELAB and assignment to the closest of eight calibrated reference swatches using minimum $\Delta E_{00}$ (Alyoubi et al., 20 May 2025). This approach optimizes robustness to variable lighting and imaging devices, especially when Gaussian blur is used for denoising.
• Diffuse Reflection-based Continuous CST (SREDS): For in-the-wild and unconstrained environments, the SREDS paradigm applies a dichromatic reflection model to decompose RGB skin patches into specular and diffuse bases via non-negative matrix factorization (NMF) (Bahmani et al., 2021). The diffuse magnitude score, normalized across a cohort, forms a continuous CST scale (e.g., CST$$_{01} = \frac{S_{\mathrm{raw}} - S_{\min}}{S_{\max} - S_{\min}}$$). SREDS demonstrates superior intra-subject reproducibility across changes in lighting, background, and pose compared to ITA or RSR measures.
• Multidimensional CST Including Hue: To capture a wider space of variation and address fairness, CST can be constructed as a two-dimensional vector $(L^*,h^*)$, where $$h^* = \mathrm{atan2}(b^*,a^*) \cdot \tfrac{180}{\pi}$$ distinguishes “red” ($h^* < 55^\circ$) versus “yellow” ($h^* > 55^\circ$) skin (Thong et al., 2023). This captures both tone and chromatic characteristics with implications for bias in vision models.

4. Validation, Accuracy, and Empirical Performance

Validation of CST methodologies has been conducted via both human studies and instrument cross-comparisons:

• Human Color-Matching Experiments: Direct annotation using CST swatches yields lower color-matching error ($\Delta E$) than alternatives across all skin types, and scales linearly with measured $L^*$. CST achieves median $\Delta E \approx 10$ vs. MST $\approx 15$, and greater step sensitivity ($\sim4.9~L^*$ units per step versus $7.4$ for MST and $14.7$ for FST) (Cook et al., 2024). Intra-class correlation for CST image ratings by device is $0.90$–$0.92$, compared to MST’s $0.81$–$0.89$.
• Statistical Comparison to Industry Standards: Smartphone CST estimation vs. reference colorimeter achieves mean absolute error (MAE) $\lesssim 1.2^\circ$ for ITA and Pearson $r \geq 0.98$ over diverse skin types (Burrow et al., 2024). Bland–Altman analysis reveals minimal mean bias ($+0.2^\circ$) and tight 95% limits of agreement.
• Classification Accuracy: Perceptually driven clustering pipelines, when paired with robust color-difference metrics (e.g., $\Delta E_{00}$ in HSV), reach up to 0.80 accuracy/$F_1$ in 8-class CST assignment, even under varying conditions (Alyoubi et al., 20 May 2025).
• Stability Across Illumination: SREDS-based CST exhibits the lowest intra-subject standard deviation in repeated measures, outperforming ITA even when lighting varies widely (Bahmani et al., 2021).

5. Bias, Limitations, and Fairness Considerations

CST reduces but does not eliminate bias inherent in subjective skin tone annotation. Known sources of error include:

• Background and Display Effects: Presentation background (white vs. gray) can shift palette-based rating by $\sim5-6~L^*$ units, evidence of simultaneous-brightness contrast (Cook et al., 2024).
• Race of Rater and Subject: Self-rating is consistently lighter for White raters and darker by Black raters ($\sim6~L^*$ bias in CST). Raters also systematically judge skin of White-identified subjects lighter at the same measured $L^*$, with differences up to $13.3~L^*$ (Cook et al., 2024).
• Imaging Device: Actual $L^*$ captured can differ by $>20$ units across devices, but scale ratings by human raters only vary by $5-6~L^*$, suggesting partial perceptual normalization (Cook et al., 2024).
• Quantization and Sensor Effects: Image-based CST resolution is limited by bit-depth and sensor spectral response. Absolute accuracy may drift across smartphone or camera models (Burrow et al., 2024).

Automated and instrument-based CST extraction pipelines—particularly those grounded in diffuse reflectance or calibrated color clustering—offer reduction but not elimination of demographic or context biases. Manual rating should be adjusted using linear/mixed-effects models, and hardware-independent pipelines must incorporate device/color card calibration where feasible (Cook et al., 2024, Thong et al., 2023).

6. Practical Implementation and Best Practices

For clinical, research, or data annotation applications, the following practices are substantiated:

• Use empirically grounded CST palettes with swatches corresponding to validated $L^*,a^*,b^*$ values (Cook et al., 2024).
• Standardize imaging geometry and disable all automatic camera corrections when acquiring image data for CST extraction (Burrow et al., 2024).
• Implement robust skin segmentation, outlier rejection, and perceptually motivated clustering/assignment algorithms in image-based pipelines (Alyoubi et al., 20 May 2025, Thong et al., 2023).
• For highest reproducibility, perform calibration with white reference tiles or reflectance cards, and choose manual exposure and lighting settings.
• Archive CST/class values in metadata for both dataset annotation and downstream fairness auditing, using longitudinal or cross-device comparison (Burrow et al., 2024).
• When possible, supplement human ratings with ground-truth colorimeter readings to provide validation and identify demographic/contextual bias (Cook et al., 2024).
• For fully objective labeling, automated pipelines must normalize for illumination and device effects by anchoring to external references or utilizing physically motivated reflectance decomposition (Bahmani et al., 2021).

7. Applications and Future Directions

CST is integral in several research and practical domains:

• Clinical Quantification: CST provides standardized covariates for pulse oximetry calibration, reducing skin-tone driven error in noninvasive diagnostics (Burrow et al., 2024).
• Dataset Annotation: CST allows reproducible, bias-aware annotation of large-scale image datasets, facilitating equitable evaluation of biometrics, computer vision, and generative AI systems (Cook et al., 2024, Thong et al., 2023).
• Machine Learning Pipelines: CST bins or continuous values serve as grouping variables for fairness metrics, performance audits, and synthesis monitoring, especially when bias along secondary hue axes is considered (Thong et al., 2023).
• Beauty Technology and Personalization: AI systems exploit CST to deliver perceptually congruent skin, hair, and undertone matching (Alyoubi et al., 20 May 2025).

Expanded research is recommended for broader population sampling, cross-device calibration, and multidimensional CST systems. Incorporation of chromatic and reflectance axes is emphasized for fully capturing phenotypic diversity and ensuring fairness across demographics (Cook et al., 2024, Thong et al., 2023).