Ethnicity Classification Based on Facial Images using Deep Learning Approach

Full text

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
217 | P a g e
www.ijacsa.thesai.org
Ethnicity Classification Based on Facial Images using
Deep Learning Approach
Abdul-aziz Kalkatawi, Usman Saeed
Dept. of Computer Science and Artificial Intelligence-College of Computer Science and Engineering,
University of Jeddah, Jeddah, Saudi Arabia
Abstract—Race and ethnicity are terminologies used to
describe and categorize humans into groups based on biological
and sociological criteria. One of these criteria is the physical
appearance such as facial traits which are explicitly represented
by a person’s facial structure. The field of computer science has
mostly been concerned with the automatic detection of human
ethnicity using computer vision-based techniques, where it can be
challenging due to the ambiguity and complexity on how an
ethnic class can be implicitly inferred from the facial traits in
terms of quantitative and conceptual models. The current
techniques for ethnicity recognition in the field of computer
vision are based on encoded facial feature descriptors or
Convolutional Neural Network (CNN) based feature extractors.
However, deep learning techniques developed for image-based
classification can provide a better end to end solution for
ethnicity recognition. This paper is a first attempt to utilize a
deep learning-based technique called vision transformer to
recognize the ethnicity of a person using real world facial images.
The implementation of Multi-Axis Vision Transformer achieves
77.2% classification accuracy for the ethnic groups of Asian,
Black, Indian, Latino Hispanic, Middle Eastern, and White.
Keywords—Vision transformer; deep learning; ethnicity; race;
classification; recognition
I. INTRODUCTION
The terms race and ethnicity are often used interchangeably
which leads to misconception in some circumstances. The
word race is used to categorize humans into groups
biologically based on physical appearance traits inherited from
the ancestors [1], whereas the term ethnicity is used to
categorize humans into groups ethnographically based on
geographic regions, language, cultural tradition, and shared
ancestry which could refer to the similar physical appearance
traits inherited but not inclusively [2].
Racial categories were first proposed in 1779s by Johann
Friedrich Blumenbach, these categories were Ethiopian-black
race, Caucasian-white race, Mongolian-yellow race, American-
red race, and Malayan-brown race [3]. A commonly adopted
racial categorization is proposed by the U.S. Census Bureau
where they categorize race into White, Black or African
American, American Indian or Alaska Native, Asian, Native
Hawaiian and Pacific Islander [4]. The White race represents
the ethnic groups originating in Europe, the Middle East, and
North Africa. The Black race represents the ethnic groups
originating in South Africa, Nigeria, Ghana, Kenya, etc. The
American Indian or Alaska Native race represents the ethnic
groups originating in North and South America also including
Central America. The Asian race represents the ethnic groups
originating in East or Southeast Asia, and the Indian
subcontinent. The Native Hawaiian and Pacific Islander race
represent the ethnic groups originating in Hawaii, Guam,
Samoa, and other Pacific Islands [5].
The human face conveys a set of semantic traits; these traits
can be used to conclude several attributes for a person such as
identity, gender, age, race or ethnicity, and expressions [6]. The
human face is the area from the upper edge of the forehead to
the chin and from the left ear to the right ear. The structure of
the facial area is represented in three main regions which are
superior, middle, and inferior. The superior region describes
the shape of the forehead, eyebrows, and eyes. The middle
region describes the shape of the nose, cheeks, and ears. The
inferior region describes the shape of the lips, chin, and jawline
[7]. Thereby, the shape of the facial structure provides
discriminant appearance traits from one person to another's. In
facial recognition systems based on computer vision techniques
the shape of the facial structure is referred to as facial features.
The complexity of facial recognition systems lies in the process
of transformation from visual facial features to a quantitative
representation of the data.
Majority of the proposed methods are based on facial
features descriptors where pre-defined procedures are
performed to capture and analyze facial images to construct a
geometry map of facial traits such as the shapes of the mouth,
nose, eyes and facial landmarks or image texture such as skin
color. Then the extracted features are encoded into a feature
vector to be used in a classifier [8]. However, recent methods
are mainly based on the automation of feature extraction using
deep learning such as convolutional neural network (CNN)
models, which have achieved better accuracy and
generalization results when trained with a sufficient amount of
representative data [8].
The lack of exploitation of deep learning techniques other
than CNN motivated the study of deep learning techniques that
can model the facial features for ethnicity recognition. This
paper employs the deep learning model Multi-axis Vision
Transformer (MaxViT) proposed by Google research team [9]
for the purpose of image-based classification. The objective is
to test the capability of MaxViT to recognize cognate facial
features that implicitly represent the discriminative appearance
traits which distinguish one ethnic group from other using
facial images. The proposed model is trained on a database
created by merging three different ethnicity datasets namely
FairFace [10], UTKFace [11], and Arab face dataset [12]. The
main contribution is that the proposed model achieves better
generalization capabilities compared to other models with an

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
218 | P a g e
www.ijacsa.thesai.org
accuracy of 77.2% for classifying six ethnic groups i.e., Asian,
Black, Indian, Latino Hispanic, Middle Eastern, and White.
The utilization of deep learning techniques such as MaxViT
would significantly improve the current state of the art for
ethnicity recognition with implication for various fields such as
human computer interaction and video surveillance.
II. RELATED WORK
A. Databases
One of the crucial factors for the advancement in the scope
of race or ethnicity recognition in computer vision is the
availability of a large and diverse dataset that provides reliable
annotated facial images based on racial or ethnic categories.
However, the research area of ethnicity recognition is still
lacking in this factor, as no dataset that represents all the racial
or ethnic groups is available. One of the most recently
proposed datasets is called FairFace [10] consisting of 97,698
images for seven ethnic groups (Black, East Asian, Indian,
Latino Hispanic, Middle Eastern, Southeast Asian, and White)
labeled by age, gender, and ethnicity. Another dataset proposed
in the field of ethnicity recognition by Zhifei Zhang et al. [11]
called the UTKFace dataset consists of 20,000 images for five
ethnic groups (Asian, Black, Indian, White, Others) labeled by
age, gender, and ethnicity. The dataset proposed by Ziwei Liu
et al. [13] called Labelled Faces in the Wild (LFW) consists of
13,233 images for three ethnic groups (Asian, Black, White)
labeled by gender, and ethnicity. MORPH dataset proposed by
Karl Ricanek et al. [14] consists of 55,134 images for five
ethnic groups (African, European, Asian, Hispanic, Others).
BUPT-BALANCEDFACE dataset proposed by Mei Wang et
al. [15] consists of more than one million images for four
ethnic groups (Asian, African, Indian, and Caucasian). BUPT-
GLOBALFACE dataset proposed by Mei Wang et al. [16]
consists of two million images for four ethnic groups (Asian,
African, Indian, and Caucasian). Mivia Ethnicity Recognition
(VMER) dataset composed from VGG-Face2 dataset [8,17]
and consisting of more than three million images for the ethnic
groups (African American, East Asian, Caucasian Latin, and
Asian Indian). There are many other facial datasets such as
Diversity in Faces (Dif) [18], IMDB-WIKI dataset [19], and
Cross-Age Reference Coding (CARC) dataset [20], these
datasets are not optimally oriented toward ethnicity
recognition.
B. Conventional Feature Extraction
This section summarizes the methods that have been
commonly used for facial features extraction using computer-
vision techniques for race or ethnicity recognition.
A study conducted by L. Farkas [21] which is based on the
relations between well-defined facial landmarks in terms of the
Euclidean distance between two points, the angle formed by a
point and two other points, and the perpendicular distance from
a point to the straight line between two other points. This study
shows that these relations can be used to distinguish the
differences in facial features of different ethnic groups.
Therefore, the use of geometric facial features to classify ethnic
groups is applicable. On the other hand, Xiaoguang et al. [22]
used appearance-based approaches that extract facial features
based on the pixel intensity values in a black-and-white image
of the face. This method achieved high accuracy when
implemented to classify between two ethnic groups Non-Asian,
and Asian, however, this method may be insufficient to classify
between more specific ethnic groups because it can vary
significantly based on images quality factors such as
resolution, viewing angles, and illumination. Another
appearance-based approach proposed by G. Zhang et al. [23]
which is based on the invariant of monotonic transformation in
the grayscale images using Local Binary Pattern histograms to
describe the texture and shape variations. S. Hosoi et al. [24]
extracted ethnic facial features using Gabor Wavelet
Transformation besides Retina sampling. Their proposed
method achieved a high accuracy relative to the number of
ethnic groups concluded in the experiment. An approach
proposed by N. Narang et al. [25] to extract facial features
from images by locating eye centers using manual annotation
and affine transformation to construct a geometric
representation of face images. Kazimov T. et al. [26] proposed
a method to define ethnic features based on the Euclidean
distance between 30 geometric landmarks. H. Ding et al. [27]
proposed an approach based on 3D face models where ethnic
features are extended using Oriented Gradient Maps. M. A.
Uddin et al. [28] proposed an integrated approach to classify
the ethnicity of Caucasian, African, and Asian based on texture
and shape features using a histogram of oriented gradients and
Gabor filter to extract features from a grayscale image, and
then combining both feature vectors into one.
C. Deep Learning-based Feature Extraction
This section summarizes the deep learning approaches that
have been proposed previously for feature extraction for
ethnicity recognition.
Marwa Obayya et al. [29] used a fusion of three pre-trained
CNN models as feature extractors, namely VGG16, Inception
v3, and capsule networks. And a bidirectional long short-term
memory model as a classifier, the model is trained using
VMER dataset and achieves an accuracy of 70% for classifying
four ethnic groups of African American, East Asian, Caucasian
Latin, and Asian Indian. Gurram Sunitha, K. et al. [30] used a
pre-trained Xception CNN model as a feature extractor and
kernel extreme learning machine model is used as classifier.
The model is trained using the BUPT-GLOBALFACE dataset
and achieves an accuracy of 97% for classifying four ethnic
groups of Asian, African, Caucasians, and Indian. Norah A.
Al-Humaidan et al. [12] used a pre-trained ResNet50 CNN
model as a feature extractor and a fully connected layer for
classification. The model is trained on a sub-ethnic group of
Arabs dataset consisting of 5,598 images of Gulf Cooperation
Council (GCC) countries people, 1,665 images of Levant
people, and 1,555 images of Egyptian people. The model
achieved 76% classification accuracy. Heng Zhao et al. [31]
proposed ethnicity recognition framework by utilizing a CNN
model, Content-Based Image Retrieval model (CBIR), and
Support Vector Machines (SVM) classifier. A VGG-16 CNN
model is used for feature extraction and a Bag-of-Words model
is used as CBIR, a combination of CNN feature and ranking
feature are used to train SVM model for classification. The
model is trained using a dataset consisting of 1,000 images of
Bangladeshi people, 1,520 images of Chinese people, and
1,078 images of Indian people. The model achieved 95%

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
219 | P a g e
www.ijacsa.thesai.org
classification accuracy. Hu Han et al. [32] used a modified
AlexNet CNN model with batch normalization layers for
feature extraction and two fully connected layers for
classification. The model is trained on MORPH-II dataset
achieving 96% classifying accuracy for three ethnic groups of
Black, White, and Other. Anwar Inzamam et al. [33] used a
pre-trained VGG-Face CNN model for feature extraction and a
SVM model as a classifier. The model is trained on ten
different databases, using ten-fold cross-validation where nine
databases are used for training and one for testing. The model
achieved 98% average classification accuracy over all
databases for three ethnic groups of Asian, White, and Black.
Amr Ahmed et al. [34] used a Feed-Forward based CNN
model and max pooling layer for classification. The model is
trained using Face Recognition Grand Challenge dataset
achieving 93% classifying accuracy for three ethnic groups of
Asian, White, and Other.
A summary of related work based on deep learning
approach is shown in Table I, describing the model used, the
number of ethnicity groups classified by the model and
accuracy achieved by the model compared to the proposed
model.
TABLE I. A SUMMARY OF RELATED WORK BASED ON DEEP LEARNING APPROACH
Author
Method
Ethnicity groups
Accuracy
Marwa Obayya et al. [29]
Fusion of VGG16, Inception
v3, and capsule network CNN models
African American, East Asian, Caucasian
Latin, and Asian Indian
70%
Gurram Sunitha, K. et al. [30]
Xception CNN model
Asian, African, Caucasians, and Indian
97%
Norah A. Al-Humaidan et al. [12]
ResNet50 CNN model
GCC people, Levant, and Egyptian
76%
Heng Zhao et al. [31]
VGG-16 CNN model
Bangladeshi, Chinese, and Indian
95%
Hu Han et al. [32]
AlexNet CNN model
Black, White, and Other
96%
Anwar Inzamam et al. [33]
VGG-Face CNN model
Asian, White, and Black
98%
Ahmed et al. [34]
Feed-Forward based CNN model
Asian, White, and Other
93%
Proposed model
MaxVit vison transformer model
Asian, Black, Indian, Latino
Hispanic, Middle Eastern, and White
77%
III. PROPOSED METHOD
This section describes the proposed approach for ethnicity
recognition using computer-vision techniques based on deep
learning. There are two main limitations of the existing
techniques described in the literature review section. First,
most of the proposed techniques are limited to classifying up to
four ethnic groups. Secondly, the proposed techniques are
limited to the utilization of CNN models for the purpose of
feature extraction. Thus, this paper proposes a model for
classifying six ethnic groups i.e., Asian, Black, Indian, Latino
Hispanic, Middle Eastern, and White by employing the
MaxViT model which is a hybrid Vision Transformer based
model capable of feature extraction and classification.
A. Multi-Axis Vision Transformer (MaxViT)
Initially transformers were proposed for the task of natural
language processing [35], the prime feature of transformers is
the self-attention mechanism which is the ability to capture
semantic relations between data segments in a sequence.
However, recently in the scope of computer-vison transformers
have attracted considerable interest in the research community
and several approaches have been proposed for image
classification, segmentation, object detection, and generation.
Thereby, Zhengzhong Tu et al. [9] proposed a Self-Attention
mechanism named multi-axis self-attention (Max-SA) which
can capture both local and global semantic relations between
data segments. This is accomplished by decomposing the self-
attention mechanism into window attention for local interaction
and grid attention for global interaction. The Max-SA
mechanism is a stand-alone attention module which can be
adopted in any network architecture. Thus, The Max-SA
module is the backbone structure of MaxViT model (see Fig.
1) coupled with Inverted Residual Block (MBConv) [36]. The
model is available on Google Colab notebook (
https://colab.research.google.com/drive/1UvseIP7zvFiysagSp4
zfvt9f9ErHu-lo?usp=sharing ).
MaxViT module uses the relative positional multi-head
attention mechanism. The basic concept of an attention
mechanism is to estimate the relevance of one data token to
other data tokens in a sequence. In self-attention layer there are
three trainable weight matrices 󰇛󰇜from which
three variables are generated by performing dot-product
multiplication of the initial input variable 󰇛󰇜 with learnable
matrices represented as 󰇛      󰇜,
from which attention layer output is represented as in Eq. (1),
where  is input size [37].
󰇛 󰇜  󰇡
󰇢 
As for multi-head attention which is an extension of self-
attention where the input is first partitioned into several
segments and each segment is processed in parallel by a
separate attention layer from which the output of each layer is
considered as an attention head. Hence, multiple attention
heads are aggregated as the final output allowing the model to
capture various feature aspects of the input. As for the relative
positional self-attention, an additional bias is concatenated with
the output of the attention layer which incorporates positional
importance of data tokens in a sequence.
The MaxViT module is composed of three main blocks.
First, the MBConv with Squeeze-and-Excitation (SE) [38]
block, window attention block, and grid attention block. The
MBConv with SE is utilized to enhance the model efficiency

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
220 | P a g e
www.ijacsa.thesai.org
and generalization, where MBConv are used to scale the model
depth wise allowing it to capture complex features, and SE are
used as channels wise self-attention mechanism that capture
interdependencies between channels. The window attention
block transforms the input feature map into non-overlapping
windows to represent a confined attention by reshaping it
󰇡

    󰇢 where P is the window size. The grid
attention block transforms the input feature map into uniform
grid to represent a sparse attention by reshaping it 󰇡   


󰇢 where G is the grid size. Each of the attention blocks
outputs are reshaped back to the initial input shape and passed
through a multi-Layer perceptron block is illustrated in Fig. 1.
Fig. 1. MaxViT module attention mechanism.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
221 | P a g e
www.ijacsa.thesai.org
Fig. 2. Architecture of MaxVit model.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
222 | P a g e
www.ijacsa.thesai.org
The MaxViT model architecture is shown in Fig. 2. The
MaxViT model architecture can be described as follows. First,
the input layer which takes an input feature map of size (C, H,
W) where C depicts the feature map channels/depth, H the
feature map height, and W the feature map width. The stem
layer uses convolutional layers to extract low-level features
from the input and reduce its spatial dimensionality. Thus it
reduces the computational complexity of the model. The
MaxVit block which is composed of sequentially staked
MaxViT modules where each block outputs half the resolution
of the prior block with a doubled channels size. Finally, the
classifier which transforms the multi-dimensional output into
one-dimensional i.e., a feature vector. From then the feature
vector is passed to a fully connected layer which performs
linear transformation and outputs a prediction.
IV. EXPERIMENT AND RESULTS
This section describes the datasets used for model training
and testing, the experiment conducted, and the results obtained.
The experiments were implemented using PyTorch
(2.0.0+cu118) for Python (3.10.5) and executed on a computer
with Intel Core i7-6700 processor with 16 GB RAM, and RTX
3080 with 10-GB VRAM GPU.
A. Dataset
The experiments were conducted on a database created by
merging three datasets FairFace [10], UTKFace [11], and Arab
face dataset [12]. Dataset is described in Fig. 3 composed of
six classes with sample sizes of 15,937 for Asian, 18,589 for
Black, 18,074 for Indian, 14,988 for Latino Hispanic, 15,188
for Middle Eastern and 28,645 White, with a total of 111,421
samples split into 101,474 samples for training and 9,947
samples for testing. The ratio of training samples to testing
samples per class is shown in Fig. 4. Random samples from
each dataset are shown in Fig. 5.
B. Experiment
This section describes the configuration and
hyperparameters used for the proposed model. The objective of
this experiment is to employ the MaxVit transformer-based
model for ethnicity recognition using facial images. The
experiment utilizes transfer learning technique to reduce
computational complexity and training time by reusing the pre-
trained parameters of all the model layers excluding the
classifier head which were modified to an output size of 6
hence the initial model is trained on ImageNet dataset [39] for
object classification with an output size of 1000. Also, in the
experiments all of the pre-trained model layers parameters are
retrained.
Data Preprocessing: Typically, when utilizing transfer
learning data must follow the same preprocessing pipeline used
for training the initial model. Thus, for data preprocessing the
first step is to resize the image to 224 × 224 pixels with center
crop applied, and then a random horizontal flip of the image is
applied with probability of 80% as a data augmentation. Next,
image pixels values are converted from 0 to 255 to be between
0.0 and 1.0, where each of the color channels values
represented as (Red, Green, Blue) are normalized by a mean of
0.485, 0.456, 0.406 and standard deviation of 0.229, 0.224,
0.225, which can enhance the model‟s learning process by
standardizing the input. The specific values are often
determined empirically based on the dataset being used. In this
case, these values are used when trained on ImageNet dataset.
Fig. 3. Dataset overview.
Fig. 4. Data split ratio.
Fig. 5. Samples overview from each dataset.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
223 | P a g e
www.ijacsa.thesai.org
Model hyperparameters: The model takes an input tensor
shape of (B, C, H, W) where B stands for batch size and has
the value of 20 images, C for color channels which is 3 for
each image, H for the pixel height of each image that is 224,
and W for the pixel width of each image that is 224. Based on
the experiment an optimal batch size is between 16 and 20,
thus for the purpose of reducing training time and fully
utilizing hardware capacity the batch size is set as 20 images. A
head dimension of 32 is used to represent the output feature
map of the attention layers with partitioning size of 7 × 7 for
both window and grid attentions, for the purpose of reusing the
pre-trained parameters. Cross-entropy loss is used as a loss
function to measure the dissimilarity between predicted
probabilities and the actual targets. As for model parameters
optimization the Adadelta algorithm is implemented with a
learning rate of 0.1. Adadelta is an adaptive learning rate
technique [40] which dynamically and automatically adjusts
the learning rates on a per-parameter level, thus based on the
experiment with Adadelta optimizer the learning rate value
does no substantial impact on the learning process unless
extreme values are used.
C. Results
In the main experiment the model was trained for 15
epochs achieving the highest classification accuracy of 0.772
for classifying six ethnic groups of Asian, Black, Indian, Latino
Hispanic, Middle Eastern, and White. Additional experiments
are conducted for a comparison in which four CNN models are
trained using the same dataset. These models are a pre-trained
VGG-Face model based on Vgg-16 architecture [41] which is
developed for face recognition with over two million faces
images. A pre-trained VGG-Face2 model based on ResNet-50
architecture [17] which is also developed for face recognition
but with over three million faces images. A pre-trained
EfficientNet-V2 model [42] for object classification is also
based on MBConv. Additionally for the purpose of testing
MaxVit model scalability an experiment is conducted by
training the proposed model on three ethnic classes i.e., Black,
White and others which is a merged class of all the remaining
categories. For training the sample sizes used are 17,015
samples for Black, 17,099 samples for White, and 18,458
samples for the merged class. For evaluation the sample sizes
used are 1,300 samples for black, 1,300 samples for white and
1,500 samples for the merged class. The model achieved a
classification accuracy of 0.835. Lastly for comparison the
same experiment is conducted with three classes using the
AlexNet model [32] which achieved the classification accuracy
of 0.782. The results are shown in Fig. 6. Additionally, the
classification accuracy of top two predicted classes is shown in
Fig. 7 where the proposed MaxVit model achieves the highest
score of 91.3%. A comparison of model size in terms of
parameters size is shown in Fig. 8, where the proposed MaxVit
model being the smallest model in terms of parameters size.
Hence, in terms of performance smaller models require lower
computational capacity thus being more efficient in terms of
speed and size on disk. Confusion matrices are shown in Fig. 9
describing the models classification performance of 9,947
samples for six classes.
Fig. 6. Models classification accuracies.
Fig. 7. Top two predicted classes accuracy scores.
Fig. 8. Models parameters size.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
224 | P a g e
www.ijacsa.thesai.org
Fig. 9. Confusion matrices.
Based on observation, the model‟s misclassification for
both of Latino Hispanic, and Middle Eastern is noteworthy.
This due to the high overlapping diversity between the three
ethnic groups of White, Latino Hispanic, and Middle Eastern
which are merely considered multiracial groups [43]. Thus,
considerable efforts are required for the creation of a
representative dataset for such ethnic groups, which certainly
could improve the performance of ethnicity recognition
models. However, in the conducted experiments the proposed
MaxVit model achieves better generalization compared to
other models.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
225 | P a g e
www.ijacsa.thesai.org
V. CONCLUSION
This paper addresses the two common limitations of
research in the field of race recognition. First, a large database
has been created with six racial categories i.e., Asian, Black,
Indian, Latino Hispanic, Middle Eastern, and White. Second, it
has proposed the usage of a vision transformer named MaxVit
as an ethnicity recognition model using facial images. It
achieves a classification accuracy of 77.2% and better
generalization than other recent works.
The research area of race and ethnicity recognition is still
unsaturated, mainly from the aspect of racial or ethnic groups
diversity, as the number of pre-defined racial categories is
limited by the available datasets. This could be particularly
problematic for individuals who have mixed racial
backgrounds, thus multiracial classification is noteworthy as
racial facial traits are noticeably overlapped for most of
population individuals. Therefore, a considerable effort should
be devoted towards such aspects, which certainly contribute to
the advancement of the research area.
ACKNOWLEDGMENT
This work was funded by the University of Jeddah, Jeddah,
Saudi Arabia, under grant no. (UJ-21-DR-12). The authors,
therefore, acknowledge with thanks the University of Jeddah
technical and financial support.
REFERENCES
[1] “Definition of Race,” Merriam-Webster Dictionary, 2023.
[2] A. Morgan, P. Catherine, P. Heather, “Ethnicity,” Oxford Classical
Dictionary, Oxford University Press, 2015.
[3] J. Blumenbach, T. Bendyshe, “The anthropological treatises of Johann
Friedrich Blumenbach,” 1865.
[4] E. Jensen, “Measuring racial and ethnic diversity for the 2020 census,”
The United States Census Bureau, 2021.
[5] „‟Revisions to the standards for the classification of Federal data on race
and ethnicity,‟‟ Office of the Federal Register, National Archives and
Records Administration, Federal Register 62, no. 210, 58782-58790
1997.
[6] J. Calder, G. Rhodes, M. Johnson, and J. V. Haxby, “Oxford handbook
of face perception,” 2011.
[7] J.D. Nguyen, H. Duong, “Anatomy, Head and Neck, Face,” Treasure
Island (FL): StatPearls Publishing, 2023.
[8] A. Greco, G. Percannella, M. Vento, and V. Vigilante, “Benchmarking
deep network architectures for ethnicity recognition using a new large
face dataset,‟‟ Machine Vision and Applications, 31-67, 2020.
[9] Z. Tu, H. Talebi, H. Zhang, F. Yang, P. Milanfar, A. Bovik, and Y. Li,
“MaxViT: Multi-Axis Vision Transformer,‟‟ European Conference on
Computer Vision, 2022.
[10] K. Kimmo, and J. Joo, “FairFace: Face attribute dataset for balanced
race, gender, and age for bias measurement and mitigation,” Workshop
on Applications of Computer Vision, 2021.
[11] Z. Zhang, Y. Song, and H. Qi, “Age progression/regression by
conditional adversarial autoencoder,” In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pp.5810–
5818, 2017.
[12] N.A. Al-Humaidan, M. Prince, “A classification of arab ethnicity based
on face image using deep learning approach,‟‟ in IEEE Access, vol. 9,
pp.50755-50766, 2021.
[13] Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in
the wild,” In Proceedings of International Conference on Computer
Vision (ICCV), 2015.
[14] K. Ricanek, T. Tesafaye, “Morph: A longitudinal image database of
normal adult age-progression,” In 7th International Conference on
Automatic Face and Gesture Recognition (FGR06), pp.341–345, 2006.
[15] M. Wang, W. Deng, J. Hu, X. Tao, Y. Huang, „‟Racial faces in the wild:
reducing racial bias by information maximization adaptation network,‟‟
ICCV, 2019.
[16] M. Wang, Y. Zhang, W. Deng, “Meta balanced network for fair face
recognition,‟‟ TPAMI, 2021.
[17] Q. Cao, L. Shen, W. Xie, O.M. Parkhi, and A. Zisserman, „‟Vggface2: a
dataset for recognising faces across pose and age. In: IEEE International
Conference on Automatic Face & Gesture Recognition, pp.67–74, 2018.
[18] M. Merler, N. Ratha, R. S. Feris, and J. R. Smith, „‟Diversity in faces,‟‟
arXiv preprint arXiv:1901.10436, 2019.
[19] R. Rothe, R. Timofte, and L. Van Gool, „‟Deep expectation of real and
apparent age from a single image without facial landmarks,‟‟
International Journal of Computer Vision (IJCV), 2016.
[20] B. Chen, C. Chen, and W. H. Hsu, „‟Face recognition and retrieval using
cross-age reference coding with cross-age celebrity dataset,‟‟ IEEE
Transactions on Multimedia, 17(6):804-815, 2015.
[21] L. Farkas, “Anthropometry of the head and face,” Raven Press, 2nd ed.,
1994.
[22] X. Lu and A. K. Jain, “Ethnicity identification from face images,” Proc.
SPIE 5404, Biometric Technology for Human Identification, 2004.
[23] G. Zhang, and Y. Wang, “Multimodal 2D and 3D facial ethnicity
classification,” 2009 Fifth International Conference on Image and
Graphics, 2009.
[24] S. Hosoi, E. Takikawa and M. Kawade, “Ethnicity estimation with facial
images,” Sixth IEEE International Conference on Automatic Face and
Gesture Recognition, 2004.
[25] N. Narang, T. Bourlai, “Gender and ethnicity classification using deep
learning in heterogeneous face recognition,” 2016 International
Conference on Biometrics (ICB), 2016.
[26] T. Kazimov and S. Mahmudova, “About a method of recognition of race
and ethnicity of individuals based on portrait photographs,” Intelligent
Control and Automation, 5, pp.120-125, 2014.
[27] H. Ding, D. Huang, Y. Wang, and L. Chen, "Facial ethnicity
classification based on boosted local texture and shape descriptions,"
10th IEEE International Conference and Workshops on Automatic Face
and Gesture Recognition (FG), 2013.
[28] M. A. Uddin, and S. A. Chowdhury, "An integrated approach to classify
gender and ethnicity," 2016 International Conference on Innovations in
Science, Engineering and Technology (ICISET), 2016.
[29] M. Obayya, S. S. Alotaibi, S. Dhahb, R. Alabdan, M. Al-Duhayyim, M.
A. Hamza, M. Rizwanullah, and A. Motwakel, „‟Optimal deep transfer
learning based ethnicity recognition on face images,‟‟ Image and Vision
Computing, Volume 128, 2022.
[30] G. Sunitha, K. Geetha, S. Neelakandan, A. K. S. Pundir, S. Hemalatha,
and V. Kumar, „‟Intelligent deep learning based ethnicity recognition
and classification using facial images,‟‟ Image and Vision Computing,
Volume 121, 2022.
[31] H. Zhao, D. Manandhar and Kim-Hui Yap, “Hybrid supervised deep
learning for ethnicity classification using face images,” 2018 IEEE
International Symposium on Circuits and Systems (ISCAS), 2018.
[32] H. Hu, J. Anil K., W. Fang, S. Shiguang, and C. Xilin, „‟Heterogeneous
face attribute estimation: A deep multi-task learning approach,‟‟ IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 40, no.
11, pp.2597-2609, 2018.
[33] A. Inzamam, and N. Ul-Islam, “Learned features are better for ethnicity
classification,” Cybernetics and Information Technologies 17, 2017.
[34] A. Ahmed, K. Yu, W. Xu, Y. Gong, and E. Xing, „‟Training hierarchical
feed-forward visual recognition models using transfer learning from
pseudo-tasks,‟‟ In: Forsyth, D., Torr, P., Zisserman, A. (eds) Computer
Vision – ECCV 2008. ECCV, 2008.
[35] A. Vaswani, N. Shazeer, N. Parmar , J. Uszkoreit , L. Jones , A. N.
Gomez, L. Kaiser , and I. Polosukhin, „‟Attention is all you need,‟‟
Advances in neural information processing systems 30, 2017.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
226 | P a g e
www.ijacsa.thesai.org
[36] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.C. Chen,
„‟Mobilenetv2: Inverted residuals and linear bottlenecks,‟‟ In
Proceedings of the IEEE conference on computer vision and pattern
recognition, pp.4510-4520, 2018.
[37] S. Khan, M. Naseer, M. Hayat, S. W. Zamir, F. S. Khan, and M. Shah,
„‟Transformers in vision: A survey‟‟, In ACM Computing Surveys,
Association for Computing Machinery (ACM), 2022.
[38] J. Hu, L. Shen, and G. Sun, „‟Squeeze-and-excitation networks,‟‟ In
Proceedings of the IEEE conference on computer vision and pattern
recognition. pp.7132-7141, 2018.
[39] J. Deng, W. Dong, R. Socher, L. -J. Li, K. Li, and Li Fei-Fei,
"ImageNet: A large-scale hierarchical image database," 2009 IEEE
Conference on Computer Vision and Pattern Recognition, Miami, FL,
USA, 2009.
[40] M. D. Zeiler, “ADADELTA: An Adaptive Learning Rate Method,”
ArXiv, abs/1212.5701, 2012.
[41] P. Omkar M., A. Vedaldi, and A. Zisserman, “Deep face recognition.”
British Machine Vision Conference, 2015.
[42] M. Tan, and Q. V. Le, „‟ EfficientNetV2: Smaller models and faster
training.‟‟ ArXiv, abs/2104.00298, 2021.
[43] L. Charmaraman, M. Woo, A. Quach, and S. Erkut, „‟How have
researchers studied multiracial populations? A content and
methodological review of 20 years of research,‟‟ Cultural Diversity and
Ethnic Minority Psychology, 20(3), 2014.