Aesthetic of Colour: A Machine Learning Approach of Palette

Generation and Aesthetic Classification

Ananya

, Batchu B. V. Aashish

, Kunda Chudanath

, Geetha K N

and Shali S

Dept of Mechanical Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Bengaluru,560035, India

Dept of Mathematics, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Bengaluru,560035, India

Keywords: Dominant Colour Extraction, Principal Component Analysis (PCA), Clustering, Support Vector Machine

(SVM), Colour Palette Generation, Aesthetics Classification, Colour Harmony, Computational Design.

Abstract: In this paper, a mathematical and computational framework is developed for the automatic generation of

aesthetic colour palettes from images using Principal Component Analysis, clustering, and Support Vector

Machines. Methodologically, the procedure is started with the key role played by Principal Component

Analysis in extracting colours with the aim of bringing dimensionality reduction while important chromatic

information is preserved. Secondly, the grouping of similar colours and the identification of predominant hues

or common base palettes are achieved using clustering algorithms. Then, the assignment of a retrieved colour

palette to a target aesthetic category is performed through classification by Support Vector Machines, and

further palettes of the same aesthetic are generated. In this way, an avenue is opened for possible application

in design, visual arts, and user interface personalization, enabled by data-driven insights into the quality of

colour harmony and aesthetic perception. Finally, the consistency and aesthetic qualities of the created palettes

over image datasets used are demonstrated through the presentation of experimental results.

1 INTRODUCTION

Colour is an integral part of visual communication

and has a deep effect on emotions, perception, and

decision-making processes. From art and design to

branding and user interfaces, it plays a ubiquitous

role. The choosing of harmonious colour palettes has

traditionally depended on the intuition and experience

of artists and designers, but computational techniques

have now brought the possibility of automating and

refining such a process, hence bridging the gap

between creativity and technology.

Colour aesthetics is the study of principles that

make colour combinations visually appealing and

emotionally resonant. Knowledge of this area is not

only important for making designs appealing but also

for creating engagement and communicating

messages effectively. In a number of fields, including

advertising, user experience design, and content

creation, aesthetic coherence and adaptability have

https://orcid.org/0009-0006-3318-1110

https://orcid.org/0009-0007-1825-8105

https://orcid.org/0009-0009-5162-9451

become critical concerns, further increasing the

demand for systematic colour palette generation.

The proposed method is a new way for the

creation of attractive colour palettes: through a

combination of mathematical models with machine-

learning techniques for discovering dominant colours

from an image and classifying these to aesthetic

categories. The presented framework improves

efficiency, enabling a systematic approach toward

this traditionally intuitive process in creating palettes.

At the heart of the framework is the application of

Principal Component Analysis (PCA) to reduce

dimensionality, retaining the most dominant

chromatic features of the image. These features are

then clustered in a meaningful way using a clustering

algorithm, allowing for the identification of dominant

colours forming the base palette.

Support Vector Machine (SVM) classifiers is

being used to assign these palettes to aesthetic themes

like "minimalist," "neon," or "pastel." Once

classified, more palettes of the same aesthetic can be

268

Ananya, , Aashish, B. B. V., Chudanath, K., K N, G. and S, S.

Aesthetic of Colour: A Machine Learning Approach of Palette Generation and Aesthetic Classiﬁcation.

DOI: 10.5220/0013590600004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 268-276

ISBN: 978-989-758-763-4

generated, allowing for applications in automated

design, adaptive user interfaces, and content

generation. This method provides an effective bridge

between the mathematical underpinnings of colour

analysis and their practical applications in design.

Through a combination of benefits from PCA

reduction, clustering for interpretability, and SVM for

precise classification, this framework simplifies

palette creation while allowing customization and

scalability. This structure responds to some of the key

areas in computational design, visual aesthetics, and

machine learning by providing a strong and adaptive

solution for the generation of colour palettes based on

aesthetic objectives.

2 PROBLEM STATEMENT

The traditional process of selecting harmonious

colour palettes has long relied on the subjective

intuition and experience of artists and designers,

which often proved inconsistent and inefficient.

Despite the increasing demands for visually

appealing and emotionally resonant designs in

brandings, user experience, and content creation, only

a few systematic methods help to streamline palette

creation in a more efficient manner. Most of the

existing computational techniques focus on object

classification or scene analysis, making aesthetic

colour analysis an under-explored domain in the area

of colour perception. This gap points toward a

framework that can extract dominant colours from

images in an automated way and categorize them into

aesthetic classes, allowing for scalable, consistent

palette generation. At the same time, approaching this

problem requires bridging the divide between

computational efficiency and creative demands in the

design.

3 LITERATURE REVIEW

The combination of machine learning and

dimensionality reduction techniques has been a focal

point in such diverse fields as image processing,

security systems, and aesthetic analysis. Methods like

PCA and SVM have been constantly used to achieve

efficient feature extraction and classification and,

therefore, are directly relevant to the proposed

research on colour palette generation and aesthetic

classification.

Jiang et al. (2023) investigated the fusion of

multiple features in image classification. The

application of PCA for dimensionality reduction and

SVM for classification showed how these methods

could be applied to complex datasets—very close to

the requirement in this study of extracting and then

classifying dominant colours from images. This really

demonstrates how PCA and SVM can simplify high-

dimensional colour data while preserving major

features that are important for aesthetic evaluation.

Qi and Wang (2014) highlighted the usefulness of

clustering and classification in improving image

categorization. The work on colour clustering directly

pertains to the proposed study, in which clustering

methods are applied to extract dominant colours.

Extending these methods, the proposed research

carries their application into the aesthetic realm, an

area that has been relatively unexplored.

Shieh et al. (2014) explored the application of

PCA and SVM in combination with PSO to real-time

face recognition. Even though the domain is different,

the flexibility of PCA and SVM to accommodate

different tasks shows their robustness and potential

suitability to the task at hand—colour palette

extraction and classification. More importantly, the

optimization techniques utilized in the study hint at

possible directions for optimizing the proposed

framework.

Malik and Waheed (2021) proposed an

unsupervised approach where PCA was used to

reduce dimensionality, and K-means clustering

addressed classification challenges. This

methodology provides a strong foundation for the

clustering-based extraction of dominant colours from

images, as proposed in this research. The parallels

between hyperspectral data classification and colour

data analysis emphasize the transferability of these

techniques.

Machine learning techniques have also been

extended to security domains, evidenced by

Varunram et al. (2021) Although this is focused on

intrusion detection, the exploration of PCA and other

dimensionality reduction methods shows their

effectiveness in extracting meaningful patterns from

high-dimensional data, which is a critical step in the

proposed research.

Deep learning applications can be seen in many

studies such as the one done by Nossam et al. (2024)

which used convolutional neural networks to detect

forgeries. While this study epitomizes the state of the

art in deep learning, its computational requirements

only strengthen the importance of lightweight

alternatives in the form of PCA and SVM, especially

for aesthetic applications that may not require the

complexity of deep learning models.

Singh and Babu (2019) introduced new methods

for analysing hyperspectral images, showing new

Aesthetic of Colour: A Machine Learning Approach of Palette Generation and Aesthetic Classiﬁcation

269

ways in which classification methodologies can be

developed. These insights go in tandem with the

proposed research, showing that feature extraction

and classification are key to getting accurate results.

Similarly, Kirola et al. (2022) further emphasizes the

importance of analysing visual data for a wide range

of applications, again validating the potential of

image-based approaches for aesthetic classification.

Although substantial progress has been made,

little emphasis has been placed on leveraging PCA

and clustering techniques specifically for extracting

dominant colours from images and classifying these

palettes into aesthetic categories using SVM. Current

research is mainly devoted to object or scene

classification and overlooks the creative and aesthetic

aspects of colour analysis. The proposed research

addresses this void by introducing a novel framework

that employs PCA and clustering to generate colour

palettes and then classifies them with SVM,

expanding these palettes based on aesthetic

categories. It contributes not only to the advancement

of colour analysis methodologies but also to closing

the gap between computational image processing and

the exploration of aesthetics.

4 DATA PREPRATION

4.1 Image Preprocessing

 Input Image: The process begins with the

input image, which is typically in RGB

colour space.

 Resizing: To optimize computation and

reduce memory requirements, the image

might be resized to a smaller resolution

while maintaining its colour characteristics.

 Flattening the Image: The RGB value of

every pixel in the image is converted into a

2D array where each row represents one

pixel and the columns represent the RGB

colour channels, that is, a matrix with the

size N×3 where N is the total number of

pixels.

4.2 Data Collection and Labelling for

Training Dataset

 Input Data: Data was collected as a group of

colour swatches assigned with various

aesthetic classes. Classes include "neon, "

"pastel, " "monochrome, " "vintage, "

"modern".

 Labelling: Links a colour swatch in a dataset

to an aesthetic class based on a qualitative

description of their appearances. The labels

are the ground truth for Support Vector

Machines (SVM) training.

4.3 Feature Extraction from Colour

Palettes

 Dominant Colours: This can be achieved by

Principal Component Analysis (PCA) and

clustering, for example, Apply k-means on

all images or palettes to pull out the

dominant colours. Those are the dominant

colours that will turn out to become the most

important features in the classification.

 Representation of Colour Palette: Each

colour palette obtained from a given image

is represented as a set of features. The

common feature set of course will comprise

of RGB or HSV values of the dominant

colours. Given that to extract K dominant

colours from an image, the feature vector for

this palette would be a vector of K×3 values-

assuming each colour is represented in the

RGB space.

As an example, for the case of K=5

dominant colours, it would have a feature

vector like, Feature Vector = [R1, G1,

B1,B1, R2, G2, B3, R3, G3, B3, R4, G4, B4,

R5, G5, B5]

Figure 1: Importance of various features

 Additional features: Depending on the

colour harmony, contrast, saturation, and

distribution in the palette, additional features

like hue, mean luminance, etc. that may

enrich the feature set of the classifier can be

added.

INCOFT 2025 - International Conference on Futuristic Technology

270

5 DATA NORMALIZATION

5.1 Standardization for Applying PCA

Before running Principal Component Analysis

(PCA), the colour data needs to be standardized so as

to have zero mean and unit variance. This ensures that

data is centred around zero; thus, the results from

Principal Component Analysis (PCA) will be

effective.

𝑍=

𝑣𝑎𝑙𝑢𝑒 − 𝑚𝑒𝑎𝑛

𝑠𝑡𝑎𝑛𝑑𝑎

𝑟

𝑑

𝐷

𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛

)

5.2 Preprocessing the Data for

Training

 Scaling Features: SVMs scale better when

the input features are in a similar range.

Since RGB colours have values ranging

from 0 to 255, the features need to be

normalized or standardized to range within

the same scale. This is usually achieved by

rescaling the value to fall within the range 0

to 1.

For example,

𝑋 𝑠𝑐𝑎𝑙𝑒𝑑 =

𝑋

(2)

 Splitting Data: Split the labelled dataset into

train and test sets, for instance, using 80-20

split to evaluate the SVM model's

performance.

In doing so, the framework will ensure that the input

to the subsequent machine learning steps is optimized

by carefully preparing and normalizing the data,

which will further enhance the accuracy and

robustness of the aesthetic classification and palette

generation.

6 IMPLEMENTING PRINCIPAL

COMPONENT ANALYSIS (PCA)

AND K-MEANS CLUSTERING

6.1 Apply Principal Component

Analysis (PCA)

 Transformation: The transformed data

provided applies Principal Component

Analysis (PCA) on standardized colour data.

PCA's main aim is to reduce the

dimensionality of the data without losing the

most important features. Here, PCA will

identify the directions in which the colour

data varies the most and then project it on a

smaller number of components, often 2 or 3

(Jaadi, 2024).

Figure 2: 3D plot of all RGB pixels

 Extract Principal Components: The colour

data of the image is represented in a reduced

dimensional space. For example, when 2

principal components are taken into

consideration, the data of the image is now

reduced to 2 features that describe the

dominant colour variations of the image. It

helps to focus on the most relevant colour

features and eliminates the irrelevance of

others.

Aesthetic of Colour: A Machine Learning Approach of Palette Generation and Aesthetic Classiﬁcation

271

Figure 3: Eigenvectors aligned in the 3D pixel plot

Figure 4: 2D scatter plot along the PCA components

6.2 Create Clusters

 K-means Clustering: Once colour data is

reduced to principal components, a

clustering algorithm such as k-means is used

to group similar colours together. The k-

means algorithm works by partitioning the

data into K clusters, representing the most

dominant colours, based on Euclidean

distances between points in the principal

component space.

It initializes the problem using K centroids,

then assigns a closest centroid to each pixel

or data point and recalculates the centroids

based on the mean of all points assigned to

each cluster. The process is repeated until

the centroids converge (Yasini, 2023).

 Identify K: The value of K is user-defined,

which means it specifies how many

dominant colours will be found. There are

two ways to decide on K: image complexity

and the degree of colour granularity needed.

In most applications, K is set to a low value,

such as 5 or 10, so as to reflect the major

colour groups within the image.

Figure 5: Clusters formed in PCA space

6.3 Assignment of Dominant Colours

 Colour Identification: After the cluster

process, every cluster should be assigned a

dominant colour, which is generally

considered to be the centroid of that cluster.

The centroid represents the average colour

of all the pixels in that cluster, which is

therefore referred to as the "dominant"

colour for that region of the image.

 Creation of the Palette: The final output is a

list of K dominant colours. Such colours

often appear in a colour space like RGB or a

more perceptually uniform colour space

such as HSL or LAB.

6.4 Advantage of Applying PCA

Verification that applying Principal Component

Analysis (PCA) is helping can be done using the

following:

 K-Means Clustering Inertia Comparison:

It basically Measures compactness of clusters. Lower

is better.

INCOFT 2025 - International Conference on Futuristic Technology

272

Inertia =

∑

oints

(a)

(3)

a: distance to nearest centroid

 Silhouette Score:

Measures separation between clusters. Higher is

better.

𝑆𝑆 =

𝑏−𝑎

𝑀𝑎𝑥(𝑎, 𝑏)

(4)

a: Mean distance to points in the same cluster.

b: Mean distance to points in the nearest different

cluster.

7 IMPLEMENTING SUPPORT

VECTOR MACHINE (SVM)

7.1 Train the Support Vector Machine

(SVM) Model

7.1.1 Kernal Selection

SVMs use various kernels to transform the input

space. The most common in general use are:

 Linear Kernel: It is perfect for linearly

separable data.

 Radial Basis Function (RBF) Kernel: This

one comes into general use where data is not

linearly separable.

 Polynomial Kernel: Applied when data

follows some non-linear patterns.

The linear kernel works well when the classes are

somewhat distinct and separable with a straight line

(or hyperplane in higher dimensions), making it a

simpler and faster choice compared to the more

complex RBF kernel (Patle and Chouhan, 2013).

SVMs can only handle binary classification, so to

deal with multiclass problems One-vs-One (OvO) is

being used.

7.1.2 How OvO binary classification works

 Train a separate SVM classifier for each pair

of classes.

 For C classes, train C×(C−1)/2 classifiers,

each on a different pair of classes.

 In this case, all classifiers classify the

sample into one class, and the class that

receives most of the votes is assigned to the

sample

7.1.3 Training the SVM

Train the SVM model on the training set. In the

training step, the SVM will try to find that optimal

hyperplane which best separates the various aesthetic

classes in the feature space.

Figure 6: Training Data in PCA space

The model learns to classify each colour palette

based on its extracted features and the corresponding

label, namely aesthetic category.

7.2 Model Evaluation

 Testing the Model: After training, test the

performance of SVM on testing set to see

how good it is in identifying unseen colour

palettes. It is an interesting business metric,

as below:

 Accuracy: Percentage of correct predictions.

 Precision, Recall, F1 Score: These metrics

give further insight into how well the model

does in handling each aesthetic category,

specifically where data is imbalanced.

Figure 7: Confusion Matrix

Aesthetic of Colour: A Machine Learning Approach of Palette Generation and Aesthetic Classiﬁcation

273

 Confusion Matrix: A confusion matrix

shows the number of correct and incorrect

classifications for each class-aesthetic

category-helping to visualize the

performance of the model.

 Cross-validation: Use cross-validation to

enhance the reliability of the model's

performance estimate by training and testing

on different subsets of the data.

7.3 Aesthetic Prediction

 Classification of New Colour Palettes:

Using the learned SVM model, one can then

classify new colour palettes. For example,

given a new collection of dominant colours

that have been extracted from an image, this

model would predict the aesthetic category

that best describes the image in terms of the

learned patterns.

Figure 8: SVM decision boundary in PCA space

 Results Visualization: After the model

generates an aesthetically pleasing class

prediction, this gives room to visualize the

colour palette with its aesthetic label. These

suggestions can then be used to generate

dynamic colour palette recommendations

based on the aesthetic one wants, thus

making the process henceforth more

personalized or tailored.

 Propose other visual palettes: Given a

classification of some colour style palette, it

then proposes other palettes that share

similar aesthetic properties. That is done by

querying a database for other palettes with

the same predicted aesthetic.

8 RESULTS AND DISCUSSIONS

The proposed system has been tested on images

containing nature, architecture, and artwork for

checking the dominant colour extraction, aesthetic

classification, and palettes generation. The results are

as follows:

Firstly, this image was uploaded.

Figure 9: Uploaded Image

From the uploaded image, colour data is retrieved

from each pixel and it is reduced in dimension by

using Principal Component Analysis (PCA). So

instead of three components (RGB), only two

components are being used.

These pixels are then clustered using K-means

clustering and 5 dominant colour are received from

the uploaded image. These colours are displayed as

copyable hex codes as well as a visual palette.

Figure 10: Colour palette Generated

The advantages of using Principal Component

Analysis (PCA) can be mathematically seen.

Reduction in clustering inertia indicates more

compact clusters being formed after using PCA.

Increased silhouette score indicates more distance

between different clusters after using PCA.

INCOFT 2025 - International Conference on Futuristic Technology

274

Figure 11: Inertia Comparison and Silhouette Score results

Finally, the model is correctly able to predict the

image aesthetic as pastel and provides the user with

similar pastel palettes as well.

Figure 12: Predicted Aesthetic and Similar Palettes

 Dominant Colour Extraction: Principal

Component Analysis (PCA) effectively

reduced the dimensionality of colour space

while preserving the most impactful colour

features. Overall, K-Means was the best

overall clustering algorithm that could

segment the dominant colours to a general

average accuracy of 95% in identifying key

visual tones.

 Aesthetic Classification: The Support

Vector Machines (SVM) classifier is also

trained on a labelled set of aesthetic

categories such as "minimalistic," "vivid,"

and "monochromatic," achieving a

classification accuracy of 94%. It, therefore,

exemplifies the robustness of the system in

projecting dominant colour palettes into

subjective aesthetic labels.

Results indicate that there is hidden potential in

using the Principal Component Analysis (PCA),

Clustering, and Support Vector Machines (SVM)

together in an aesthetic-driven system for building a

palette since good performance was obtained from the

identification of colours and categorization into

Figure13: Classification report

aesthetic subjective groups. The feasibility of

applying such systems in domains which have

demonstrated distinct need for clear computational

precision and sensitive aesthetic perception was

confirmed.

Therefore, what is required are refined algorithms

for clustering or even additional features such as

texture and spatial information for further

performance.

Another useful output was the generation of novel

palettes for the same aesthetic class. Many design

applications, such as automated design systems for

branding and content creation, would be useful with

this system. Sometimes, the output would not be

especially novel, but aesthetically coherent. Future

versions of the system could use generative models

such as GAN (Generative Adversarial Networks) to

make the system much more creative.

Accordingly, the proposed system bridges the gap

between computational colour analysis and aesthetic

interpretation. In this manner, it is extremely useful

for applications depending upon perception and

engagement by users highly based on colour.

9 CONCLUSIONS

This paper proposes a novel method of colour palette

generation based on the integration of dimensionality

reduction, clustering, and supervised classification. A

system is developed and successfully used to extract

dominant colours from the images using Principal

Component Analysis (PCA) and clustering, classify

palettes into aesthetic categories with Support Vector

Machines (SVM), and generate additional palettes

consistent with the identified aesthetic.

The results show this methodology has the

capability of achieving high accuracy in both colour

extraction and aesthetic classification and presents

itself as a robust tool to use in applications related to

design, branding, and even content creation. In

Aesthetic of Colour: A Machine Learning Approach of Palette Generation and Aesthetic Classiﬁcation

275

addition, the generated palettes were found to be

aesthetic coherent, thus underlining practical utility.

Despite the positive results, it still has drawbacks,

such as handling complicated textures and enhancing

the originality of generated palettes. Here, there are

many possible future research directions: in addition

to modern algorithms, GANs (Generative Adversarial

Networks) could be used to add generative creativity,

while other features help with more detailed aesthetic

classification.

The proposed system bridges the gap between

computational and subjective domains,

demonstrating the potential of leveraging

mathematical frameworks to address artistic

challenges. This work lays the groundwork for future

innovations in automated design systems and

aesthetic-driven computational tools.

ACKNOWLEDGEMENTS

We sincerely thank Dr. Geetha K N and Dr. Shali S

for their valued guidance throughout the semester and

for helping us with all their time and effort in shaping

this project. We express our gratitude for your

supervision and constructive feedback which brought

us to the completion of the project.

REFERENCES

Jiang, C., Chen, T., Tao, G. (2023). Gap image

classification based on PCA-SVM with multiple feature

fusion. 2023 4th International Conference on

Mechatronics Technology and Intelligent

Manufacturing (ICMTIM). IEEE.

Qi, L. Y., Wang, K. G. (2014). Information system in image

classification based on SVM and color clustering

analysis. Advanced Materials Research, 886, 572–575.

Shieh, M. Y., Chiou, J. S., Hu, Y. C., Wang, K. Y. (2014).

Applications of PCA and SVM‐PSO‐based real‐time

face recognition system. Mathematical Problems in

Engineering, 2014(1), Article 530251.

Malik, A., Waheed, M. (2021). Unsupervised classification

of hyperspectral images using PCA and k-means.

Yasini, S. (2023). Using machine learning to create custom

colour palettes. Towards Data Science. Available at:

https://towardsdatascience.com/using-machine-

learning-to-create-custom-colour-palettes-

acb4eeaa06aa.

Jaadi, Z. (2024). Principal Component Analysis (PCA): A

step-by-step explanation. Brennan Whitfield.

Patle, A., Chouhan, D. S. (2013). SVM kernel functions for

classification. 2013 International Conference on

Advances in Technology and Engineering (ICATE), 1–

Varunram, T. N., Shivaprasad, M. B., Aishwarya, K. H.,

Balraj, A., Savish, S. V., Ullas, S. (2021). Analysis of

different dimensionality reduction techniques and

machine learning algorithms for an intrusion detection

system. 2021 IEEE 6th International Conference on

Computing, Communication and Automation (ICCCA),

237–242.

Nossam, S. C., Katakam, R. A., Pulastya, G., Jayan, S.

(2024). Signature forgery detection and verification

using deep learning techniques. 2024 15th International

Conference on Computing, Communication and

Networking Technologies (ICCCNT), 1–6.

Singh, T., Babu, T. (2019). Fractal image processing and

analysis for classification of hyperspectral images.

2019 10th International Conference on Computing,

Communication and Networking Technologies

(ICCCNT), 1–6.

Kirola, M., Joshi, K., Chaudhary, S., Singh, N., Anandaram,

H., Gupta, A. (2022). Plants diseases prediction

framework: An image-based system using deep

learning. 2022 IEEE World Conference on Applied

Intelligence and Computing (AIC).

INCOFT 2025 - International Conference on Futuristic Technology

276