Perceptual Color Difference Modeling Using Machine Learning and Human Similarity Judgments

Elnara Kadyrgali; Muragul Muratbekova; Adilet Yerkin; Nuray Toganas; Ayan Igali; Malika Ziyada; Aruzhan Burambekova; Jamaladdin Hasanov

cs.CVSep 30, 2026

Perceptual Color Difference Modeling Using Machine Learning and Human Similarity Judgments

Authors: Elnara Kadyrgali, Muragul Muratbekova, Adilet Yerkin, Nuray Toganas, Ayan Igali, Malika Ziyada, Aruzhan Burambekova, Jamaladdin Hasanov, +1 more

Organizations: School of Information Technology and Engineering (SITE), Kazakh-British Technical University, Almaty, Kazakhstan · School of Information Technology and Engineering (SITE), ADA University, Baku, Azerbaijan

Abstract

Accurate assessment of color differences is essential for applications ranging from digital design to quality control. While existing color difference metrics, such as CIEDE2000, aim to approximate human perception, they may still exhibit inconsistencies with perceptual judgments. In this study, we investigate a data-driven approach to color-difference estimation based directly on human evaluations. We collect similarity judgments for 2,000 systematically generated color pairs, each rated by seven observers using a four-point ordinal scale. These judgments are then used to train regression models using different color representations, including RGB channel differences, HSI differences, and COLIBRI fuzzy linguistic categories. Experiments with five regression algorithms show that the choice of color model has a greater influence on prediction performance than the choice of regression algorithm. Using COLIBRI features alone, linear regression achieves an R2 of 0.595, outperforming RGB and HSI representations, which achieve R2 values of 0.479 and 0.493, respectively. The best performance is obtained by LightGBM using the combined representation, reaching an R2 of 0.703. The results indicate that human perceptual color differences are better captured when numerical color coordinates are complemented by graded perceptual categories, highlighting the potential of data-driven models for perceptually aligned color-difference estimation.

Figures & tables

Study	Separation conditions	Color pairs	Color difference range	Assessment method	Key finding (separation effect)
Vision science foundations
Boynton et al., 1977 [ 7 ]	Juxtaposed; with gap	—	Threshold	Threshold detection	Foundational gap effect study: gap impaired luminance discrimination but improved chromatic discrimination; explained by contour enhancement and spatial averaging
Sharpe & Wyszecki, 1976 [ 8 ]	Juxtaposed; with gap	—	Threshold & supra	Ratio comparison; liminal det.; color matching	Separation impairs lightness more than chromaticness; proposed a proximity factor for color-difference formulas
Eskew, 1989 [ 9 ]	Juxtaposed; narrow gap (isoluminant fill)	—	Threshold	Threshold detection	Luminance or chromatic gap prevents spatial integration, enhancing chromatic sensitivity; gap had little effect for flashed stimuli
Danilova & Mollon, 2006 [ 10 ]	Juxtaposed; separated up to 10°	—	Threshold	2AFC threshold	Discrimination optimal at small gap; thresholds rose moderately with increasing separation; gap effect more pronounced in parafovea
Psychophysical experiments — surface & printed colors

Table I : Summary of Prior Studies on the Effect of Sample Separation on color-Difference Evaluation. Studies are grouped by stimulus type. The number of observers ranged from 4 to 46 across studies; see original sources for details. NS = no separation; — = not reported in original source.

Panel of Figure 1 — Figure 1 : Color pairs visualization

Panel of Figure 2 — Figure 2 : Interfaces used in the color similarity judgement experiments

Figure 3 : Proposed framework for perceptual color difference modeling

Dataset characteristic	Value
Participants	7
Unique color pairs	2 000
Total evaluations	28 600
Human similarity rating distribution
Not similar	9 501 (33.22%)
Somewhat similar	7 192 (25.15%)

Table II : Summary of the perceptual color similarity dataset and grouped cross-validation design.

Metric	Feature Set	Linear Regression	Decision Tree	Random Forest	LightGBM	XGBoost
RMSE	RGB	$0.2653\pm 0.003$	$0.2794\pm 0.006$	$0.2519\pm 0.004$	$0.2438\pm 0.004$	$0.2530\pm 0.005$
	HSI	$0.2616\pm 0.004$	$0.2546\pm 0.007$	$0.2312\pm 0.003$	$0.2268\pm 0.001$	$0.2344\pm 0.002$
	COLIBRI	$0.2337\pm 0.004$	$0.2654\pm 0.009$	$0.2341\pm 0.006$	$0.2274\pm 0.005$	$0.2364\pm 0.003$
	RGB + HSI	$0.2552\pm 0.004$	$0.2409\pm 0.005$	$0.2155\pm 0.003$	$0.2116\pm 0.003$	$0.2161\pm 0.003$
	RGB + COLIBRI	$0.2248\pm 0.005$	$0.2422\pm 0.003$	$0.2118\pm 0.001$	$0.2033\pm 0.002$	$0.2091\pm 0.002$
	HSI + COLIBRI	$0.2315\pm 0.004$	$0.2413\pm 0.002$	$0.2143\pm 0.002$	$0.2104\pm 0.002$	$0.2156\pm 0.001$

Table III : Performance comparison of regression models across different feature sets using five-fold grouped cross-validation. Values are shown as mean

\pm

standard deviation across folds.

Model Dependency	Feature	Coefficient (Mean $\pm$ SD)
		RGB	HSI	COLIBRI	RGB + HSI	RGB + COLIBRI	HSI + COLIBRI	RGB + HSI + COLIBRI
	Intercept	$0.6930\pm 0.005$	$0.6792\pm 0.005$	$0.7851\pm 0.005$	$0.6954\pm 0.005$	$0.7899\pm 0.005$	$0.7799\pm 0.005$	$0.7937\pm 0.005$
	Red Channel	$-0.0027\pm 0.0$	-	-	$-0.0014\pm 0.0$	$-0.0015\pm 0.0$	-	$-0.0018\pm 0.0$
RGB	Green Channel	$-0.0029\pm 0.0$	-	-	$-0.0016\pm 0.0$	$-0.0013\pm 0.0$	-	$-0.0016\pm 0.0$
	Blue Channel	$-0.0024\pm 0.0$	-	-	$-0.0013\pm 0.0$	$-0.0011\pm 0.0$	-	$-0.0014\pm 0.000$
	Hue	-	$-0.7848\pm 0.007$	-	$-0.4954\pm 0.010$	-	$-0.2282\pm 0.008$	$0.0950\pm 0.016$

Table IV : Linear regression coefficients for perceptual color similarity across different feature representations using five-fold grouped cross-validation. Values are shown as mean

\pm

standard deviation across folds.

Figure 4 : Experimental results for Experiment 1 and 2 side to side

Category	Experiment 1	Experiment 2
Very Similar	$\Delta E\leq 4.221$	$\Delta E\leq 2.069$
Similar	$4.221<\Delta E\leq 11.406$	$2.069<\Delta E\leq 6.743$
Somewhat Similar	$11.406<\Delta E\leq 24.266$	$6.743<\Delta E\leq 18.889$
Not Similar	$\Delta E>24.266$	$\Delta E>18.889$

Table V : Perceptual similarity decision boundaries thresholds

Figure 5 : Comparison of the proposed model with conventional color-difference measures.

Pair	Model Difference	Model Similarity	Model Interpretation	$\mathbf{\Delta E_{00}}$	CIEDE2000 Interpretation
1–2	0.539	0.461	Somewhat similar	18.066	Very much
3–4	1.000	0.000	Not similar	68.840	Strongly
5–6	0.982	0.018	Not similar	74.104	Strongly
7–8	0.403	0.597	Similar	3.047	Appreciable
9–10	0.444	0.556	Similar	25.478	Strongly
11–12	0.498	0.502	Similar	9.713	Much

Table VI : Linguistic comparison between the proposed model and CIEDE2000.

Explore similar work

Human-AI Perceptual Alignment by Playing Hues and Cues
Aug 7, 2026Nuria Alabau-Bosque, Jorge Vila-Tomás, Paula Daudén-Oliver +3Perceptually Artificial Intelligence Alignment
Image Processing Lab, Universitat de València, Carrer del Catedrátic José Beltrán Martinez, Paterna, 46980, Spain
Beyond Color Geometry: Evaluating Human-Like Color Representations in Vision Models
Jul 15, 2026Ayan Igali, Pakizar ShamoiVision Foundation Models Vision Transformer
School of Information Technology and Engineering Kazakh–British Technical University Almaty, Kazakhstan
Perceptual Asymmetry Between Hue Categories: Evidence from Human Color Categorization
May 10, 2026Elnara Kadyrgali, Nuray Toganas, Muragul Muratbekova +1Asymmetry
School of Information Technology and Engineering Kazakh-British Technical University, Almaty, Kazakhstan 050000

Aug 7, 2026, cs.CVJ/K move · Enter open · S save

Human-AI Perceptual Alignment by Playing Hues and Cues

Nuria Alabau-Bosque, Jorge Vila-Tomás, Paula Daudén-Oliver, Pablo Hernández-Cámara+2

Image Processing Lab, Universitat de València, Carrer del Catedrátic José Beltrán Martinez, Paterna, 46980, Spain

Perceptually Artificial Intelligence Alignment

Open paper AskPDF arXiv

Evaluating the perceptual alignment between Contrastive Vision-Language Models (CVLMs) and humans is typically constrained by traditional benchmarks that overlook fine-grained semantic and cultural nuances. In this work, we propose a novel evaluation framework that leverages the gamified, discrete color space of the board game Hues and Cues. By mapping the board's 480 color cells to the CIE xy chromaticity diagram, we calculate empirical perceptual distances across a carefully curated 100-word vocabulary spanning seven semantic categories. To properly contextualize model performance, we establish an empirical lower bound of expected error-the Human Consistency baseline-calculated via Leave-One-Out (LOO) cross-validation on a dense dataset of color associations collected from 325 human observers through a custom digital interface. We evaluate 162 models across multiple architectural families and pre-training datasets to assess their semantic color grounding. Our results demonstrate that while CVLMs successfully replicate human cognitive biases, such as idealized memory colors for concrete physical referents (e.g., food and plants), they systematically diverge from the human baseline in abstract, subjective, and pop-culture domains. We identify two distinct failure modes in severely misaligned concepts: semantic misclassification and a systematic uncertainty collapse into a default blue coordinate. Furthermore, we reveal that highly curated pre-training datasets are significantly more effective than massive, uncurated corpora in mitigating these severe misalignments. Ultimately, this work highlights that despite their broad categorization capabilities, current CVLMs still fail to capture the nuanced, localized consensus of human color memory, emphasizing the value of gamified tasks in exposing underlying model biases. The data and code are publicly available to test other metrics.

Abstract

Explore similar work

Human-AI Perceptual Alignment by Playing Hues and Cues

Beyond Color Geometry: Evaluating Human-Like Color Representations in Vision Models

Perceptual Asymmetry Between Hue Categories: Evidence from Human Color Categorization