Source-linked AI summary
From Perception to Pixels: Human-Derived Visually Distinct Colors for Improved Dominant Color Extraction
Uldana Shyndali, Nurdaulet Toleugaliyev, Kuat Zharykbayev, Zarina Abildasheva, Pakizar Shamoi
TL;DR
Conventional dominant-color methods can miss human-perceived color differences, despite the task’s applications in visual computing. The paper builds a Human Color Dictionary from a perceptual experiment and integrates it into extraction pipelines. Evaluation reports improved agreement with expert palettes, alongside findings on gender and color distinguishability.
Problem
Dominant-color extraction is useful in visual-computing applications, but RGB-based K-means can rely on distances that do not reflect how people perceive color differences.
Source: §1 · p. 1Method
The authors derive a 64-color Human Color Dictionary from a 120-color perceptual experiment and map pixels to dictionary colors using ∆E00 in CIE Lab within K-means and histogram methods.
Source: §1 · p. 2Results
The integration consistently improves agreement with expert palettes across tested configurations; the study also reports no meaningful gender differences and greater distinguishability for highly saturated than dark or deep colors.
Source: §6 · p. 10Takeaways & Limitations
The findings support using experimentally derived perceptual representations in color extraction to make results more closely reflect human visual perception.
Source: §6 · p. 10Future work will explore cultural variation in color naming and integration with deep-learning-based color quantization.
Source: §6 · p. 11
Abstract
from arXiv · showhide
Dominant color extraction is a fundamental task in computer vision, with applications in design, image retrieval, and content recommendation. However, conventional algorithms based on device-oriented color spaces (e.g., RGB) and fixed quantization often fail to capture how humans perceive color differences. In this study, we propose a human-perception-based framework for dominant color extraction, built on a novel Human Color Dictionary (HCD) derived from a perceptual experiment. Participants (n=110) were sequentially shown 120 colors and asked to identify those they perceived as unique. Based on their responses, we constructed a 120 by 120 co-occurrence matrix and applied a greedy selection algorithm to derive 64 perceptually distinct colors, forming the HCD. This dictionary was integrated into RGB color histogram and K-means color extraction algorithms by initializing centroids and mapping pixel values to their nearest perceptual color using the deltaE 00 distance. Evaluation on a dataset of 4,562 images paired with expert-designed palettes demonstrated that our method improves perceptual alignment, as measured by Earth Mover's Distance (EMD) and best-match distance. Our findings suggest no meaningful gender differences in the number of uniquely perceived colors, indicating that sex does not significantly influence color perception in this task. Another key finding is that highly saturated colors are perceived as the most distinguishable, whereas dark and deep colors tend to perceptually collapse and are less reliably differentiated. The results confirm that incorporating experimentally derived human color representations ensures the generated color clusters align more closely with human perception, thus improving the usability of color palette generation in practical applications.
1. Introduction
Dominant-color extraction matters across visual-computing applications, but conventional methods can diverge from human judgments because they group colors using pixel statistics and device-oriented representations. This paper introduces a perceptually derived color dictionary and integrates it into established extraction methods, evaluated against expert-designed palettes.
1. Introduction
- Dominant-color extraction supports image retrieval, automatic palette generation, visual data analysis, and content-based recommendation, while alignment with human perception remains challenging.Color also shapes how people interpret visual information across design, marketing, digital art, and interface aesthetics.Source: §1 · p. 1
- Conventional RGB and statistical grouping can split visually similar shades, underrepresent perceptually dominant hues with fewer pixels, and make K-means depend on initial centroids and cluster counts.K-means commonly relies on RGB Euclidean distance, which may not reflect perceived color differences.Source: §1 · p. 1
- The framework maps pixels to their nearest HCD color using ∆E00 in CIE Lab and integrates the dictionary with K-means and RGB histograms.The stated evaluation uses best-match distance and Earth Mover’s Distance against palettes for 4,562 images.Source: §1 · p. 2
- A 64-color Human Color Dictionary (HCD), derived from responses by 110 participants, supplies experimentally selected perceptually distinct hues for extraction.The study also presents a reproducible construction method and a benchmark of 4,562 images with expert-designed palettes.Source: §1 · p. 2
2. Related Work
Prior work spans clustering, perceptual color modeling, and application-focused methods; the section identifies a need for experimentally grounded perceptual representations that can integrate with existing clustering frameworks.
2.1. Clustering-Based Approaches to Color Extraction
- Clustering methods are widely used for dominant-color extraction, but pixel statistics, RGB distance, and predefined cluster counts can produce perceptually inconsistent palettes.Reported approaches also examine preprocessing and color-space choices, while configuration can substantially affect K-means output.Source: §2.1 · p. 2
2.2. Perceptual and Human-Centered Color Modeling
- Perceptual color spaces, perceptual attributes, fuzzy models, and color descriptors have been used to better represent human discrimination or visual prominence.Examples include CIELAB/CIELUV, saturation and contrast cues, and MPEG-7’s combination of histograms with perceptual relevance.Source: §2.2 · pp. 2–3
2.3. Application-Driven, Interactive, and Fuzzy Extensions
- Application-specific and interactive palette methods target design workflows, while color-naming harmonization and fuzzy approaches address palette coherence or uncertainty.These methods may rely on predefined algorithms or color sets, handcrafted rules, or application-specific assumptions.Source: §2.3 · p. 3
2. Related Work
- The review motivates an experimentally grounded, human-derived color representation that can be integrated into existing clustering frameworks.It characterizes prior perceptual and fuzzy models as rarely grounded in large-scale human experiments.Source: §2.3 · p. 3
3. Methods
The methods use expert-designed image palettes as ground truth, describe standard RGB K-means and histogram baselines, and integrate a human-derived color dictionary using perceptual color distances.
3.1. Data Collection
- 4,562 images paired with manually selected five-color expert palettes provide the ground-truth dataset.The images were collected from designer websites.Source: §3.1 · p. 3
3.2. K-means Color Clustering
- K-means represents each image with five color clusters, assigning pixels to nearest RGB centroids and updating those centroids with cluster means until convergence.The method begins with randomly positioned centroids and repeats assignment and update steps until centroid positions stabilize.Source: §3.2 · pp. 3–4
3.3. RGB Histogram for Color Extraction
- The RGB histogram divides each channel into eight bins, ranks bins by pixel count, and returns the centers of the five most frequent bins.The bin centers represent the image’s statistically frequent colors.Source: §3.3 · p. 4
3.4. CIELAB Color Space and CIEDE2000
- CIELAB uses lightness, green–red, and blue–yellow axes, while ∆E00 serves as the study’s primary distance measure for perceptually aligned color comparisons and mappings.The study sets the ∆E00 parametric factors kL, kC, and kH to 1.Source: §3.4 · p. 4
3.5. Proposed Approach
- The proposed approach builds an HCD from participant selections, maps colors to dictionary entries, and uses the dictionary in K-means initialization and histogram-based extraction.The methods evaluate extracted palettes against expert palettes using Best-Match similarity and Earth Mover’s Distance.Source: §3.5 · p. 4
4. Experiment Design
The experiment presented participants with 120 colors and asked them to judge whether each color had been seen before. Participants viewed the full set of colors sequentially.
Stimuli
- The experiment presents participants with 120 distinct colors from the Muratbekova and Shamoi (2024) dataset, and Figure 2 shows the stimuli and their distribution in HSV color space.Source: §4 · p. 4
Procedure
- For each color, participants answered whether they had seen it before; “No” added a small color box for later recognition and comparison, while “Yes” did not.The displayed stimulus was a 200×200 px square, and saved boxes were 30×30 px.Source: §4 · pp. 4–5
- The task ended after participants had viewed all 120 colors.The experiment form’s start page included age-category and gender options and instructions about the task and privacy.Source: §4 · pp. 4–5, Figure 3 · p. 5
5. Results
The experiment produced a 64-color Human Color Dictionary from participants’ judgments, and the proposed dictionary-based methods improved palette agreement against expert-designed palettes. The discussion reports no meaningful gender difference in perceived color uniqueness and finds that saturation and intensity relate to distinguishability.
5.1.1. Human Color Dictionary
- The 110-participant experiment yielded an average of 64 distinct colors, or 53.5% of the presented stimuli, and a greedy algorithm used pairwise co-occurrence counts to select the 64-color dictionary.The 120 × 120 matrix records how often each color pair was marked distinct by the same participant; Figure 7 shows the resulting dictionary.Source: §5.1.1 · pp. 5–6, §5 · p. 5, Figure 6 · p. 6
5.1.2. Examples
- Examples show more perceptually even initial K-means centroids after refinement and compare classic with proposed K-means and histogram palettes mapped to the dictionary.The figures show centroid distributions and palette comparisons; the histogram example maps each pixel to its closest dictionary color before palette extraction.Source: §5.1.2 · p. 6, Table 2 · p. 6, Figure 8 · p. 7
5.1.3. Performance Evaluation
- Best-Match similarity for the mapped histogram increased from 64.83 to 70.78, while CIE Lab generally scored higher than RGB across K-means settings.Table 3 compares histogram and K-means variants against expert-designed ground-truth palettes using Best-Match and EMD; the passage also reports that mapping increased EMD.Source: §5.1.3 · p. 7, Table 3 · p. 8
5.1.4. Discussion
- The experiment found no meaningful gender difference in unique-color counts; the reported standard deviations were 19.83 for females and 18.14 for males.Participants were predominantly aged 18–25, and the discussion notes that this homogeneous age range may limit the emergence of age- and gender-related differences.Source: §5 · p. 5, §5.1.4 · pp. 7–8, Table 4 · p. 8
- High-saturation colors and unique hues were often distinguished, while dark and deep colors were least distinguishable; all saturation levels peaked at medium intensity.The discussion attributes the pattern to luminance contrast: dark intensity had the lowest distinguishability regardless of saturation, and high- and low-saturation colors were often distinguished from medium-saturation colors.Source: §5.1.4 · pp. 8–9, Table 6 · p. 9
5.2. Application Example
- In the application example, HCD-based color features produced image groups with similar color appearance in a hierarchical clustering of four painters’ artworks into 10 clusters.The example used the Best Artworks of All Times dataset and showed palette extraction for five images randomly selected from a 139-image subset.Source: Figure 13 · p. 9, §5.2 · p. 10, Figure 14 · p. 10
6. Conclusion
The framework embeds a human-derived color dictionary into standard clustering and histogram pipelines using perceptual color mapping. Evaluation reports stronger agreement with expert palettes, alongside findings on color distinguishability and future extensions.
6. Conclusion
- The dictionary integration consistently improves agreement with expert-designed palettes across tested configurations.It uses perceptual color mapping in CIE Lab with CIEDE2000 without changing the core structure of existing algorithms.Source: §6 · p. 10
- Highly saturated colors are most distinguishable, while dark or deep colors tend to perceptually collapse; no meaningful gender differences were observed.Source: §6 · p. 10
- The findings indicate that the human-derived dictionary makes extracted colors more reflective of human visual perception than traditional clustering algorithms.Source: §6 · p. 10
- Future work will explore cultural variations in color naming and integration with deep-learning-based color quantization.Source: §6 · p. 11