International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.

3, June 2023

CLUSTERING AN AFRICAN HAIRSTYLE DATASET

USING PCA AND K-MEANS

Teffo Phomolo Nicrocia, Owolawi Pius Adewale, Pholo Moanda Diana

Department of Computer Engineering, Tshwane University of Technology, Pretoria

(Soshanguve)

ABSTRACT
The adoption of digital transformation was not expressed in building an African face shape classifier. In
this paper, an approach is presented that uses k-means to classify African women images. African women
rely on beauty standards recommendations, personal preference, or the newest trends in hairstyles to
decide on the appropriate hairstyle for them. In this paper, an approach is presented that uses K-means
clustering to classify African women's images. In order to identify potential facial clusters, Haarcascade is
used for feature-based training, and K-means clustering is applied for image classification.

KEYWORDS
Face detection, k-means, African, hairstyle

1. INTRODUCTION
Hairstyles have always been an expression of beauty and personality for African women. In many
cultures, people see a woman’s hair as her crown of glory and consider it a lordly symbol that
symbolizes her femininity. In the early 15th century, for example, hair served as a carrier of
messages in West African countries such as Yoruba (Nigeria) and Mandingo (Sierra Leone)
communities; it communicated marital status, ethnic identity, wealth, and status in the community
while also identifying geographical origin [1]. Historical background and the influence of
mainstream media, that promotes beauty stereotypes that esteem European standards, have
affected how African women perceive their natural hair [2]. As highlighted by Bankhead &
Johnson [1], it is therefore very important to educate young girls and women about accepting
their natural hair and to motivate African women to view positive images of themselves.
According to psychologists, having beautiful hair and the feeling of having such an appearance
directly affects the psychology of women in a positive or negative way .In addition , a good
hairstyle can highlight facial features and can help improve facial appearance by 70% as it has
the tendency to enhance facial features [10].However, as we alter our hairstyle to embrace our
authentic selves, we have to remember that every person is different. It is necessary to match the
characteristics of each individual with a suitable hairstyle [3]. Many hairstyle recommendation
systems exist, as shown in Table 1. However, most of them have been created using datasets that
do not accommodate the of African women. In the existing datasets, mainly Caucasian faces are
present, while our main focus is on the classification of African women's faces.

David C. Wyld et al. (Eds): SCDD, CIOS, EDUR, CSEAI -2023

pp. 55-65, 2023. IJCI – 2023 DOI:10.5121/ijci.2023.120305
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023
Table 1. Related Work

Paper Face Shape classification Target

method

Alzahrani, Al-Nuaimy & Al-Bander (2021) Inception V3 Men and Women

Chen,Zhang,Huang, Chen&Luo (2021) YOLO v4 Men and Women

Weerasinghe & Haarcascade Classifier Men and women

Vidanagama (2020)

Liu, Ji, Hu & Zhan (2019) VGGNet-16 Men and women

Pasupa, Sunhem & Loo (2019) Support Vector Machines Women

Rajapaksha & Kumara (2018) Dlib machine learning Men and Women
library (C++)

Additionally, most of these studies use the five classic face shapes (oval, heart, round, oblong and
square) to classify facial features. In this paper, we will explore the gathered images using
clustering methods to see if these five face classes are a fit for our dataset. We will then classify
each of the faces into one of the clusters obtained, as the basis for an African hairstyle
recommendation system.

This study aims to accomplish the following major objectives:

● Collect images of African women to construct a dataset;

● Extract facial features of collected images using Haar Cascade method;
● Using K-Means clustering to perform image classification to identify the potential facial
clusters(groupings);
● Develop and evaluate a multi-layer perceptron for assigning each image to a cluster.

2. MATERIAL AND METHODS

Figure 1. Block diagram of clustering images

56
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023

2.1. Data Collection

The image datasets are obtained from Figaro1K dataset and include images gathered from the
web for scientific research conducted by non-profits. There are six classes of hair in the Figaro1K
database, which are: straight, wavy, curly, kinky, braids, dreadlocks, and short men, each
containing 120 images. African women typically have afro-textured hair or kinky hair, so I
selected images based on that hair type.

2.2. Facial Detection and Extraction

Human faces generally share a similar geometrical configuration while sharing the following
features: eyes, mouth, nose, and chin. The Viola-Jones face detection algorithm goes through an
image to detect these features. The method uses Haar-like features to extract the face and
construct a bounding box around it. These features are filters run through an image to determine
the location of the face [5].

Figure 2. Block diagram of clustering images

2.2.1. Calculating Haar Values

When scanning the image, the Haar value of each region is calculated as follows [6]:

𝑆𝑢𝑚 𝑜𝑓 𝑑𝑎𝑟𝑘 𝑝𝑖𝑥𝑒𝑙𝑠 𝑆𝑢𝑚 𝑜𝑓 𝑙𝑖𝑔ℎ𝑡 𝑝𝑖𝑥𝑒𝑙𝑠 𝑣𝑎𝑙𝑢𝑒𝑠

𝐻𝑎𝑎𝑟𝑉𝑎𝑙𝑢𝑒 = −
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑎𝑟𝑘 𝑝𝑖𝑥𝑒𝑙𝑠 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑙𝑖𝑔ℎ𝑡 𝑝𝑖𝑥𝑒𝑙𝑠

57
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023

To make the calculations faster, images can be converted to integral images. To obtain an integral
image, the value of each pixel is replaced by the sum of all pixel values above it, all pixel values
to its left and the value of the pixel itself [5]. Using the integral image, the haar value is computed
using only four values for each region, as shown on Figure below.

Figure 3. Integral image

2.2.2. AdaBoost Learning Algorithm

The majority of the Haar features obtained after are irrelevant and random. To select a good
subset of features from a huge set of features, the AdaBoost algorithm is applied [16]. AdaBoost
classifies whether a weighted combination of the Haar features contains a face or not, and so
reduces the number of potential facial regions [5][7]. The classifiers have excellent feature
selection mechanisms and fast learning rates, which makes them superior to other machine
learning algorithms.

2.2.3. Cascade Filter

The AdaBoost algorithm returns 6,000 features which are still non-face regions in majority [19].
As a next step, the algorithm therefore implements cascading classifiers. These latter are a series
of decision trees applied to the image regions obtained from the previous step. Only image
regions that are classified as facial regions by all cascading classifiers are ultimately retained [7].
This minimizes the algorithm’s false negative rate.

Figure 4. Cascade Classifier

58
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023

2.3. Image Processing

Image processing is the first step when supplying image data to any algorithm. The following
steps were involved in this stage image processing:

 Resizing: All extracted faces were reduced to the same size: 224 by 224 pixels.
 Colorscale: The images were then converted to grayscale (2-D).
 Scaling: All pixels should be normalized between 0 and 255.
 Flattening: in preparation for the next steps, the images were converted from 2-
dimensional arrays of 224 by 224 pixels into 1-dimensional arrays of 50176 pixel.

2.4. Feature Extraction

Principal component analysis involves extracting features from an image, visualizing complex
data, and capturing variation in the data. Principal Component Analysis aims to reduce the
dimensions of a dataset by selecting uncorrelated features that capture the most information about
the data [8]. The algorithm for Principal Component Analysis is as follows [11].

Step 1:Reduce continuous features to the same scale

(𝑥𝑖 − μ)
𝑧𝑖 =
σ

Step 2: Generate the covariance matrix

Covariance expresses the relationship between two variables and is calculated as ffollows:

∑(𝑋𝑖 − 𝑋)(𝑋𝑖 − 𝑌 ′ )
𝐶𝑂𝑉 (𝑋, 𝑌) =
𝑛

where x’ = mean of variable X; y’ = mean of variable Y; n = number of data points. The

covariance is calculated between every pair of variables and the results are saved in a p x p
matrix (p = number of variables in the dataset).

Step 3: Calculate eigenvectors and eigenvalues

For a square matrix A, the eigenvector (v) and eigenvalue (λ) are obtained by solving the
following equation:
𝐴𝑣 = 𝜆𝑣
Step 4: Create a feature vector to decide which principal components to keep.

Step 5: Recast the data along the principal components’ axes.

2.5. Cluster Analysis

2.5.1. K-Means Clustering

The K-Means algorithm is the most commonly used partitioning-based clustering algorithm.
Below is a brief outline of the K-Means algorithm [17]:

 Randomly select k points and make them the initial centroids.

59
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023

 Choose an element from the collection, compare it with each centroid, and then assign
that element to the cluster that corresponds to that centroid.
 Recalculate the centroids after each data point is assigned to a cluster.
 Steps 2 and 3 should be repeated until no data point moves from its previous cluster to
some other cluster.

A variety of methods can be applied to estimate the optimal number of clusters k based
on various parameters. This study measures the quality of clusters by using the Elbow Method
and the silhouette score to determine the optimal number of clusters k.

2.5.1.1. Elbow Method

The elbow method involves comparing the number of clusters that will form an elbow on the
curve to determine the optimal number of clusters. A cluster value is considered best if its value
forms an elbow on the curve. To compare cluster values, calculate SSE (Sum of Square
Error). Consequently, if there are more clusters K, then the SSE (Sum of Squared errors) value
will be smaller since the number of clusters is greater [18]. The downside of the method is that
when k is increased, the SSE value will approach zero, which makes it impossible to determine if
a cluster is good or bad [13]. To overcome this drawback, this method suggests that k should be
calculated when the maximum error value of the graph reduces significantly [14].

2.5.1.2. Silhouette Score

The silhouette score was developed as a means of exploring and analyzing the distance
between the resulting clusters. In addition to that, it also serves as a measure of how well a
cluster is doing over a given period. In Silhouette, a score varies from -1 to 1 and -1 indicates
that the time-series was assigned to the incorrect cluster, 0 indicates that there are clusters that
overlap with samples near neighboring clusters and 1 indicates that it belonged to the correct
cluster [18]. For each sample, to calculate the Silhouette score, we use the following formula:

(𝐵𝑋 − 𝐴𝑋)
𝑆(𝑋) =
max(𝐴𝑋, 𝐵𝑋)

 Assuming that X object is in a cluster, calculate its average distance from each other
object in the cluster. Let AX be the value of this parameter.

 Prior to finding its average distance from X, find its average distance from all other
clusters containing objects in the same cluster. Calculate the minimum value based on all
the clusters. Let BX be the value.
 The silhouette value of the X object is:

(𝐵𝑋 − 𝐴𝑋)
𝑆 (𝑋 )  =  
max(𝐴𝑋, 𝐵𝑋)

3. RESULTS AND DISCUSSION

From the Figaro1K dataset, we extracted only 112 images that contain only African women from
the total of 672 images.

60
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023

Figure 5. Object detection by Haar Cascade Classifier

As shown in Figure 5, the Haar cascade classification results are displayed. The Haar cascade
classification was trained on images that were not identical regarding brightness of the faces and
lack of sharp features. This resulted in the faces not being spotted and caused the classifier to
struggle to classify the images. The algorithm's average recall is low due to many false positives
and negatives, which indicates a low accuracy rate.

Figure 6. Principal Component Analysis

As shown in figure 6, the first 80 Principal Components explained 99.99% of data variance.

61
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023

Figure 7. Elbow Method

In Figure 7, we see that there is no clear elbow. However, there seems to be a bend around k = 5.
Therefore, we looked at the silhouette score for k values around 5. When looking at Figure 8, we
see that k = 4 is a good choice since the Silhouette scores of all four clusters are higher than the
average score of the entire dataset (indicated by the red vertical line). Additionally, the sizes of
the clusters are more similar with k = 4 than with k = 3, 5 or 6.

Figure 8. Silhouette Score

Below are some of the figures obtained in each cluster.

62
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023

Figure 9. Cluster 0

Figure 10. Cluster 1

Figure 11. Cluster 2

Figure 12. Cluster 3

4. CONCLUSION
In this paper, we have examined the face of African women from the Figaro1K dataset. After
extracting facial features using Haar cascade classifiers, we performed feature selection using
Principal Component Analysis. To identify the potential facial groups or clusters, we used K-
Means for clustering and obtained four clusters. Using Haar cascade yielded many false negatives
and false positives and reduced the number of faces used for clustering. Future research will
therefore focus on extracting more faces using more advanced algorithms. of the new dataset will
then be used for creating a recommendation system that first assigns an image to a cluster then
recommends hairstyles based on visual similarity.

63
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023

REFERENCES

[1] T. A. Johnson and T. Bankhead, “Hair It Is: Examining the Experiences of Black Women with
Natural Hair,” Open J. Soc. Sci., vol. 02, no. 01, pp. 86–100, 2014, doi: 10.4236/jss.2014.21010.
[2] C. Antunez, “Fros, Weaves, and Kinks: The Social and Political Significance of Hair for Black and
Coloured Women in Cape Town,” 2013.
[3] V.A. Cirolla, “Psychological Intervention,” no. October 2017.
[4] K. Pasupa, W. Sunhem, and C. K. Loo, “A hybrid approach to building face shape classifier for
hairstyle recommender system,” Expert Syst. Appl., vol. 120, no. March 2019, pp. 14–32, 2019, doi:
10.1016/j.eswa.2018.11.011.
[5] C. Rahmad, R. A. Asmara, D. R. H. Putra, I. Dharma, H. Darmono, and I. Muhiqqin, “Comparison
of Viola-Jones Haar Cascade Classifier and Histogram of Oriented Gradients (HOG) for face
detection,” IOP Conf. Ser. Mater. Sci. Eng., vol. 732, no. 1, pp. 0–8, 2020, doi: 10.1088/1757-
899X/732/1/012038.
[6] A. B. Shetty, Bhoomika, Deeksha, J. Rebeiro, and Ramyashree, “Facial recognition using Haar
cascade and LBP classifiers,” Glob. Transitions Proc., vol. 2, no. 2, pp. 330–335, 2021, doi:
10.1016/j.gltp.2021.08.044.
[7] R. Senthamizh Selvi, D. Sivakumar, J. S. Sandhya, S. Siva Sowmiya, S. Ramya, and S. Kanaga
Suba Raja, “Face recognition using Haar - Cascade classifier for criminal identification,” Int. J.
Recent Technol. Eng., vol. 7, no. 6, pp. 1871–1876, 2019.
[8] S.M. Holland,”Principal Components Analysis (PCA),” Department of Geology, University of
Georgia,Athens,GA (2019): 30602-2501.
[9] Y. Freund and R. E. Schapire, “A Decision-Theoretic Generalization of On-Line Learning and an
Application to Boosting,” J. Comput. Syst. Sci., vol. 55, no. 1, pp. 119–139, 1997, doi:
10.1006/jcss.1997.1504
[10] H. Stokman and T. Gevers, “Selection and fusion of color models for image feature detection,”
IEEE Trans. Pattern Anal. Mach.Intell. vol. 29, no. 3, pp. 371–381, Mar. 2007.
[11] S. Mishra et al., “Principal Component Analysis,” International Journal of Livestock Research, p. 1,
2017, doi: 10.5455/ijlr.20170415115235.
[12] C. Subbalakshmi, G. Rama Krishna, S. Krishna Mohan Rao, and P. Venketeswa Rao, “A method to
find optimum number of clusters based on fuzzy silhouette on dynamic data set,” Procedia Comput.
Sci., vol. 46, no. Icict 2014, pp. 346–353, 2015.
[13] J. T. Wei, S.-Y. Lin, Y.-Z. Yang, and H.-H. Wu, “Applying Data Mining and RFM Model to
Analyze Customers’ Values of a Veterinary Hospital,” in 2016 International Symposium on
Computer, Consumer and Control (IS3C), 2016, pp. 481–484.
[14] M. J. Garbade, “Understanding K-means Clustering in Machine Learning,” Towards Data Science,
2018. [Online]
[15] A. Picot-Lemasson, G. Decocq, F. Aghassian, and J. Leveque, “Influence of hairdressing on the
psychological mood of women.” International Journal of Cosmetic Science,vol.23, no.3, pp.161-
164,2001, doi:10.1046/j.1467-2494.2001.00072. x.
[16] C.Sun, J.Hu, and K.Lam, “Feature subset selection for efficient AdaBoost training.”2011 IEEE
International Conference on Multimedia and Expo, 2011, doi: 10.1109/icme.2011.6011905.
[17] Kaur,G.Deep, “A Review: Comparative Study of Image Clustering Algorithms ISSN:2278-
7844.”International Journal of Advanced and Innovative Research,vol.3, no.2,2019.
[18] D. M. Saputra, D. Saputra, and L. D. Oswari, “Effect of Distance Metrics in Determining K-Value
in K-Means Clustering Using Elbow and Silhouette Method.” Proceedings of the Sriwijaya
International Conference on Information Technology and Its Applications (SICONIAN 2019), 2020,
doi: 10.2991/aisr.k.200424.051.
[19] P. Viola and M. Jones, “Rapid object detection using a boosted cascade of Simple features,”
Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition. CVPR 2001.

64
 International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.3, June 2023

AUTHORS
The author, Phomolo N. Teffo, was born in Limpopo (South Africa) in 1993 and is
currently pursuing a master's degree at Tshwane University of Technology. She has a
bachelor's degree in intelligent industrial systems from the same university. Currently,
she is employed by the same university as a part-time lecturer.

Pius Adewale Owolawi studied electrical engineering at the Federal University of

Technology, Akure, Nigeria, in 2001, and at the University of Kwazulu Natal, South
Africa, in 2006 and 2010. Tshwane University of Technology, South Africa, currently
employs him as Department Head of Computer Systems Engineering. Among his
research interests are RF, green communication, radiowave propagation
(microwave/millimeter wave systems), satellite and free space optical communications,
the Internet of Things, embedded systems, machine learning, and data analytics. In
2012, he was awarded the Joint Holder of Best Paper Award for a paper he presented at the 2nd
International Conference on Applied and Theoretical Information Systems Research, which was held in
Taipei, Taiwan, in 2012, and in 2015 he was awarded the Vice Chancellor's Teaching Excellence Award.

The PhD candidate Moanda D.Pholo was born in Kinshasa (DRC) in 1987. In 2014,
she earned a Master's degree in Intelligent Industrial Systems from the same university,
as well as an MSc in electronics from the Ecole Supérieure d'Ingénieurs en
Electrotechnique et Electronique. Currently, she works as a data scientist in Pretoria,
South Africa. She has published several scientific articles, including one in the
GlobalSIP 2019 Proceedings. In Italy, Mrs Pholo recently received the "Best work in
the category of AI/big data" award at the 3rd ICEHTMC.

65