ASSIGNMENT ON
Reviewing Research Paper
BIG DATA AND IOT
CSE 412
SUBMITTED TO
Nasima Islam Bithi
Lecturer
Department of CSE
Daffodil International University
SUBMITTED BY
SAKHWAT HOSAN
ID : 201-15-13731, Section : A
Department of CSE
Daffodil International University
Date of Submission
Fall 23
31 August 2023
A review On
Identifying Ethnics of People through Face Recognition:
A Deep CNN Approach
Introduction :
The field of image recognition has seen growth by letting the identification of individuals,
from images and video frames. Various techniques have been used in this area, such as
recognizing 2D patterns, matching the face and analyzing features. By combining different
feature extraction methods making a hybrid approaches involving 3D images have emerged.
Though accurately classifying ethnicities still poses challenges, Convolutional Neural
Networks (CNNs) have recently become a tool for image recognition. This paper introduces
models for recognition that aim to identify peoples ethnicities and origins by analyzing their
facial features using RGB images in real time.
Problem of Statement :
Despite advancements in face recognition, accurately identifying the ethnicities of individuals
with similar features across different ethnic groups remains a challenge. This is a challenge
because people from different backgrounds might share similar facial features, making it
tough for computers to correctly guess where someone comes from based on their face. This
research aims to find a way for computers to do this better by creating new models that focus
on facial features and using special types of pictures in real time.
Objectives :
The main objectives of this paper are to propose and evaluate two new face recognition
models – one incorporating regularization and another without – to accurately recognize the
ethnicities and origins of individuals based on their facial features. The models aim to
differentiate between detailed facial features using RGB images and real-time recognition.
Additionally, a new high-resolution dataset from various Asian regions is introduced, serving
as a resource for research on ethnicity-based face recognition.
The Proposed Method :
The paper presents two novel models built using convolutional neural networks (CNNs) –
one with dropout layers and another without – to address the challenge of ethnicity-based
face recognition. The models are trained to discern ethnicities by leveraging facial features,
and the impact of regularization through dropout layers is assessed. A distinct dataset
comprising images from different Asian regions is collected for model training and
evaluation.
Twelve layers Deep Learning Model
Outcome and Contribution :
The proposed models showcase the potential to accurately recognize individuals' ethnicities
and origins through their facial features. Notably, the first model, enriched with dropout
layers, achieves an impressive validation accuracy rate of 90.65%. In contrast, the second
model, lacking dropout layers, attains a validation accuracy of 76.70%. This substantial
difference underscores the pivotal role of regularization in enhancing model performance and
mitigating overfitting. Furthermore, their proposed approach excels when compared to state-
of-the-art methods. Specifically, their approach outperforms VGG and Inception V3,
showcasing a validation accuracy rate of 96.6% with a validation loss of 0.22. In contrast,
VGG and Inception V3 achieve validation accuracy rates of 91.48% and 61.92%, with
validation losses of 0.23 and 0.81, respectively. The introduction of a new high-resolution
dataset specifically collected for ethnicity-based recognition offers a valuable resource to the
research community. The study demonstrates the significance of regularization through
dropout layers in enhancing model performance, as evidenced by the improved accuracy of
the model with dropout layers compared to the one without, when evaluated on an unseen
dataset.
