International Journal of Scientific Research in Computer Science, Engineering and Information Technology

ISSN : 2456-3307 (www.ijsrcseit.com)

doi : https://doi.org/10.32628/CSEIT2173128

Gender and Age Detection using Deep Learning

Utkarsha Kumbhar*, Prof. A. S. Shingare
Department of Computer Engineering, Vishwakarma Institute of Technology, Pune, Maharashtra, India

ABSTRACT

Article Info For the past few years, gender and age detection has been an active area of study
Volume 7, Issue 3 and researchers have been putting a lot of effort to contribute quality research in
Page Number: 604-610 this area. Starting from preprocessing of data to building a model which gives
high precision results is tedious task for researchers. There is a immense
Publication Issue : dormant field of study as it can be used in monitoring, surveillance, human-
May-June-2021 computer interaction and security. However, there is still a lack of the
performance of existing methods on real live images. Many difficult tasks such as
Article History computer vision, speech recognition, and natural language processing are easily
Accepted : 10 June 2021 solved with deep learning. Therefore, the approach of deep learning remarkably
Published : 15 June 2021 growing and this also takes place in image classification. Therefore, to analyses
and focuses on comparative study of different algorithms for gender and age
recognition system to give elevated degree of precision is required.
Keywords : CNN, Adience dataset, Feature extraction, neural network, deep
learning

I. INTRODUCTION In the current scenario, developing systems based on

age and gender for various platforms such as mobile,
In every sector, age and gender have played a applications that use social networks such as
significant role. It appears that a great deal of effort Facebook, Instagram, Snapchats, and various
has been done in every department. Unreliable data matrimonial and dating websites such as shadi.com,
was gathered from the media, healthcare, and FMCG happen, tinder, bumble, and others is a critical task
(rapid moving consumer goods) industries. Due to for human security concerns.
today's broad market of consumers and job hopefuls,
diversity appears to be more vital than ever. The important task is to use this data intelligently
Employees have been pushed to reassess standard and to build helpful solutions for human needs. An
staffing and company methods as a result of the intelligent algorithm must be designed to overcome
competitive economy. A varied workplace made up of this problem.
men and women of all ages is extremely important in
today's world for a variety of reasons. Deep learning is an artificial intelligence (AI)
function that mimics the human brain's processing of
data and pattern creation in order to make decisions.

Copyright: © the author(s), publisher and licensee Technoscience Academy. This is an open-access article distributed under the 604
terms of the Creative Commons Attribution Non-Commercial License, which permits unrestricted non-commercial use,
distribution, and reproduction in any medium, provided the original work is properly cited
 Utkarsha Kumbhar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, May-June - 2021, 7 (3) : 604-610

Deep learning is an artificial intelligence subset of The deep learning technology has substantially
machine learning that uses neural networks to learn increased the quality of picture categorization and
unsupervised from unstructured or unlabelled data. object recognition over the last seven years. The
Deep neural learning or deep neural network are essential components of a gender recognition system
other terms for the same thing. Deep learning gives before applying deep learning methods are feature
numerous ways and algorithms to solve the age and descriptor and classification approach. Many
gender detection dilemma. Choosing which techniques have been used to extract discriminative
algorithms to use is a complex job. Data scarcity is features from facial images, which can be broadly
also a major issue in determining gender and age. classified into geometric and appearance-based
methods. Geometric features such as the distance
Deep learning and machine learning can help solve a between eyes, the length of the eyes and ears, the
variety of problems. In this paper, a convolution length and width of the face, and so on have been
neural network is utilised to extract features from the used in the former category. On the other hand, in
Audience dataset, and a few classifiers are utilised to appearance-based approaches, the entire image is
improve accuracy. taken into account rather than just a portion of it.

II. RELATED WORK For the encoding of facial images, Dyadic Wavelet
Transform (DyWT) and Local Binary Pattern (LBP)
There are numerous systems on the market that can were previously utilised. LBP is a state-of-the-art
determine gender and age using various technologies, texture descriptor, and DyWT depicts a facial picture
approaches, and techniques. Deep Learning-based at multiple scales. DyWT decomposes an image into
techniques are an increasingly prominent set of distinct sub-bands at different scales, making analysis
approaches used by software engineering academics simple. The DWT transform has been utilised for face
to automate development chores (DL). The description, however because it is not translation
popularity of these techniques is due to their invariant, it does not have the best potential for
automated feature engineering capabilities, which feature extraction. DyWT is a superior choice for face
helps in the modelling of software artefacts. However, description since it is translation invariant. LBP, on
because deep learning techniques are being adopted the other hand, is superior at capturing local detail.
at such a quick rate, it is difficult to distil the current In a new method, DyWT and LBP are combined. For
research landscape's successes, failures, and prospects. the challenge of gender recognition, we propose a
Deep learning (DL) is a machine learning field with a description of facial images. Deep learning provides a
lot of potential. technique to dissecting an image into subcategories
rather than dissecting it into subcategories.
The state-of-the-art studies in several scientific areas, Convolutional neural networks (CNNs)
including as computer vision, object recognition, revolutionised the field of computer vision. It not
speech recognition, and natural language processing, only improves image classification accuracy with
have led to the artificial intelligence waking up from time, but it also helps with generic feature extraction
its deep slumber in the recent decade. Many such as scene categorization, object detection, image
researchers are currently using DL approaches to try retrieval, image caption and semantic segmentation.
to solve a variety of challenges in various domains. In image processing tasks, convolutional neural
networks (CNNs) are one of the most powerful kinds

Volume 7, Issue 3, May-June-2021 | http://ijsrcseit.com 603

 Utkarsha Kumbhar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, May-June - 2021, 7 (3) : 604-610

of deep neural networks. It's a powerful tool that's recognition, and language translation are some of the
widely utilised in computer vision applications. tasks performed through deep learning.
Convolution layers, subsampling layers, and full
connection layers are the three types of layers in a Deep learning is a subset of machine learning that
convolution neural network. analyses data using hierarchical neural networks.
Within these hierarchical neural networks,
III. PRAPOSED METHOD AND MATERIALS comparable to the human brain, neuron codes are
linked together. The hierarchical nature of deep
In proposed system, Adience dataset [15] is used for learning, unlike other standard linear programmes in
solving the problem. Adience benchmark, which is machines, allows it to take a nonlinear approach,
used to categorise people based on their age and processing data over a number of layers, each of
gender. Images from smart-phone devices are which will integrate subsequent tiers of additional
automatically uploaded to Flickr in the Adience set. information.
Because these photographs were posted without
deliberate filtering, as is common on media webpages There are three different types of neural networks:
(e.g., photographs from the LFW collection [25]) or ANN, RNN, and CNN. Each has its own importance
social websites (the Group Photos set [14]), viewing in a number of situations. Feature engineering is a
circumstances in these photographs are a little critical aspect in building a network-based model.
different. Feature engineering consists of two steps: feature
As a result, Adience photos record extreme extraction and feature selection. A convolutional
differences in head attitude, lighting quality, and neural network (CNN) is employed in the proposed
more. There are around 26K photos in the Adience system to extract a feature from the dataset. CNN
collection, with 2,284 subjects. The collection is models are employed in a wide range of applications
divided by age groups. A typical five-fold, subject- and domains, but they're especially common in image
exclusive cross-validation methodology [15] is used to and video processing.
test for both age and gender classification. These
images are used more than recent alignment
approaches to emphasise the performance
improvement ascribed to network architecture rather
than better pre-processing. Deep Learning is used to
correctly determine a person's gender and age from a
single photograph of their face. The anticipated
gender and age ranges are (0-2), (4-6), (8-12) (15-20)
(25-32) (38-43) (48-53) (60-100).
Deep learning, also known as deep neural networks
or neural learning, is a type of artificial intelligence
(AI) that aims to mimic the brain's functions. It's a
type of machine learning that uses functions to make
decisions in a nonlinear way. Deep learning occurs
when decisions are made without supervision on
unstructured data. Recognition of objects, speech Figure 1. Proposed system neural architecture.

Volume 7, Issue 3, May-June-2021 | http://ijsrcseit.com 604

 Utkarsha Kumbhar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, May-June - 2021, 7 (3) : 604-610

For both age and gender classification, we used our a variety of instances to evaluate. If the new dataset is
proposed network design throughout our studies. greater than the old one, the weight should be
Only three convolutional layers and two fully- adjusted; this is the fine-tuning strategy.
connected layers with a modest number of neurons
make up the network. This is in compared to the
much bigger structures used in [16] and [17]. Our
decision to use a smaller network was influenced by
both our goal to avoid overfitting and the nature of
the problem. Figure 1 depicts our network
architecture in its entirety. More information is
available in the text. one of the issues we're trying to
resolve: On the Adience set, age is classified into
eight categories, but gender is only divided into two.
In contrast to, for example, this is in contrast to the
ten thousand identity classes used to train the facial
recognition network in [48]. The network processes
all three colour channels directly. Images are initially
rescaled to 256 256 pixels and then sent to the Figure 2. small data with exactly similar to original
network with a crop of 227 227 pixels. Following that, dataset.
the three convolutional layers are defined as follows.
In the figure.2, the transfer learning function is used
on small data that are identical to the original dataset.
1. In the first convolutional layer, 96 filters with a
When the targeted data is small, new fully connected
size of 3707 pixels are applied to the input,
layer is generated by keeping previous weights
followed by a rectified linear operator (ReLU), a
constants.
max pooling layer with a maximum value of 3 3
regions with two-pixel strides, and a local
IV. RESULTS AND DISCUSSION
response normalisation layer [28].
2. The second convolutional layer, which contains
In the proposed system, CNN is employed to extract
256 filters of size 96 5 5 pixels, processes the
features, and various classifiers are employed to
preceding layer's 96 28 28 output. The same
achieve high accuracy.
hyper parameters as before are used for ReLU, a
max pooling layer, and a local response
The proposed system is organised into three primary
normalisation layer.
parts. Images are captured utilising a live webcam as
3. Finally, the third and final convolutional layer
well as images from a saved location in the first two
applies a set of 384 filters of size 256 3 3 pixels to
modules. With the help of the OpenCV module, it is
the 256 14 14 blobs, followed by ReLU and a
determined whether the image is male or female and
max pooling layer.
what age range it belongs to after receiving it as an
input.
In the CNN feature extraction strategy, just the top
layer of the network is trained, and the rest of the
network remains constant. In this method, there are

Volume 7, Issue 3, May-June-2021 | http://ijsrcseit.com 605

 Utkarsha Kumbhar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, May-June - 2021, 7 (3) : 604-610

There are various classification algorithms in

machine learning. SVM, MLP, Random Forest,
decision tree Extratree, adaboost, gradient boosting
classifiers are used to calculate accuracy for the
proposed system.

Figure 3. Image captured from web cam

Because of elements such as cosmetics, lighting,

obstacles, and facial expressions, determining an
exact age from a single image is extremely difficult.
As a result, rather than utilising regression to solve an
Figure 5. gender and age detection
issue, it is preferable to use classification.
Classification is a task that necessitates the
application of machine learning algorithms to learn TABLE I
how to assign a class label to problem domain ACCURACY OF DIFFERENT CLASSIFIERS
instances. The term "man" or "female" was used here.
CLASSIFIER ACCURACY
In terms of modelling, classification necessitates a
training dataset with a large number of examples of SVM 97.79%
inputs and outputs from which to learn. DECISION 97.00%
The results of classification predictive modelling TREE
algorithms are examined. Classification accuracy is a ADABOOST 97.16%
common metric for evaluating a model's performance GRADIENT 97.003%
based on projected class labels. Although BOOSTING
classification accuracy isn't ideal, it's a solid place to RANDOM 97.31%
start for a lot of classification problems. FOREST
MLP 97.03%

Figure 4. Image getting from saved location

Volume 7, Issue 3, May-June-2021 | http://ijsrcseit.com 606

 Utkarsha Kumbhar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, May-June - 2021, 7 (3) : 604-610

98 the-art adaptive video analytics platforms, each

97.8
97.6
camera can be turned into an alarm sensor, sending
97.4 push notifications to a central monitoring station that
97.2
97 can help cut expenses and enhance operations.
96.8
96.6
VII. REFERENCES

[1]. Arora, Shefali, and M. P. S. Bhatia, “A Robust

Approach for Gender Recognition Using Deep
Accuracy(in%) 0-20 30-40 50-60 70-80 90-100 Learning,” In 2018 9th International
Conference on Computing, Communication
Figure 6. Graph for accuracy of different classifiers
and Networking Technologies (ICCCNT), pp.
1-6, 2018.
Table 1 and figure 6 presents the results after [2]. Levi, Gil, and Tal Hassner, “Age and gender
applying different classifiers respectively. For the classification using convolutional neural
feature extraction CNN (figure 2) is used whereas networks.” In Proceedings of the IEEE
accuracy is calculated by classifiers. conference on computer vision and pattern
recognition workshops, pp. 34-42, 2015.
V. CONCLUSION [3]. Wang, Xiaofeng, Azliza Mohd Ali, and Plamen
Angelov, “Gender and age classification of
We presented the accuracy of many classifiers in this human faces for automatic detection of
paper. CNN is primarily employed as a learning anomalous human behaviour.” In 2017 3rd
function for feature extraction, which allows it to IEEE International Conference on Cybernetics
determine age and gender. We use 5 hidden layers (CYBCONF), pp. 1-6, 2017.
and 5-fold cross validation to produce a more accurate [4]. Haseena, S., S. Bharathi, I. Padmapriya, and R.
result. Lekhaa, “Deep Learning Based Approach for
Combination of feature extractions like frequency, Gender Classification.” In 2018 Second
and color with different classifiers will add accuracy International Conference on Electronics,
to these applications. Communication and Aerospace Technology
(ICECA), pp. 1396-1399, 2018.
VI.FUTURE SCOPE [5]. Ramdhani, B., Djamal, E.C. and Ilyas, R.,
“Convolutional Neural Networks Models for
The implementation can be used as exorbitant model Facial Expression Recognition. In 2018
and covers different field such as retail management International Symposium on Advanced
sector, cosmetic industry and security area Intelligent Informatics (SAIN), pp. 96-101,
prominently. Face recognition is utilised in a variety August 2018.
of ways in retail, from security to ads. From sending [6]. Tiagrajah V. Janahiraman1 and Prasantth
personalised adverts to identifiable customers to Subramaniam (2019) Gender Classification
gender identification. Also, Security, operational Based on Asian Faces using Deep Learning.
problems, and safety are all important considerations 2019 IEEE 9th International Conference on
in retail venues, both indoors and out. With state-of-

Volume 7, Issue 3, May-June-2021 | http://ijsrcseit.com 607

 Utkarsha Kumbhar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, May-June - 2021, 7 (3) : 604-610

System Engineering and Technology (ICSET), 7 Facial Expression Recognition. In 2018

October 2019, Shah Alam, Malaysia International Symposium on Advanced
[7]. Ericsdotter C and Ericsson A M 2001 Gender Intelligent Informatics (SAIN), pp. 96-101,
differences in vowel duration in read Swedish: August 2018.
Preliminary results Working Papers - Lund [15]. E. Edinger, R. Enbar, and T. Hassner. Age and
University Department of Linguistics 34 – 37. gender estimation of unfiltered faces. Trans. on
[8]. L. Best-Rowden, H. Han, C. Otto, B. F. Klare, Inform. Forensics and Security, 9(12), 2014
A. K. Jain, Unconstrained face recognition: [16]. A. Krizhevsky, I. Sutskever, and G. E. Hinton.
Identifying a person of interest from a media Imagenet classification with deep convolutional
collection, Information Forensics and Security, neural networks. In Neural Inform. Process.
IEEE Transactions on 9 (2014) 2144–2157. Syst., pages 1097–1105, 2012.
[9]. J. Orozco, O. Rudovic, J. Gonz`alez, M. Pantic, [17]. K. Chatfield, K. Simonyan, A. Vedaldi, and A.
Hierarchical on-line appearance-based tracking Zisserman. Return of the devil in the details:
for 3d head pose, eyebrows, lips, eyelids and Delving deep into convolutional nets. arXiv
irises, Image and Vision Computing 32 (2014) preprint arXiv:1405.3531, 2014
14–26
[10]. Arora, Shefali, and M. P. S. Bhatia, “A Robust Cite this article as :
Approach for Gender Recognition Using Deep
Learning,” In 2018 9th International Utkarsha Kumbhar, Prof. A. S. Shingare, "Gender and
Conference on Computing, Communication Age Detection using Deep Learning", International
and Networking Technologies (ICCCNT), pp. Journal of Scientific Research in Computer Science,
1-6, 2018. Engineering and Information Technology
[11]. Levi, Gil, and Tal Hassner, “Age and gender (IJSRCSEIT), ISSN : 2456-3307, Volume 7, Issue 3,
classification using convolutional neural pp.604-610, May-June-2021. Available at
networks.” In Proceedings of the IEEE doi: https://doi.org/10.32628/CSEIT2173128
conference on computer vision and pattern Journal URL : https://ijsrcseit.com/CSEIT2173128
recognition workshops, pp. 34-42, 2015.
[12]. Wang, Xiaofeng, Azliza Mohd Ali, and Plamen
Angelov, “Gender and age classification of
human faces for automatic detection of
anomalous human behaviour.” In 2017 3rd
IEEE International Conference on Cybernetics
(CYBCONF), pp. 1-6, 2017.
[13]. Haseena, S., S. Bharathi, I. Padmapriya, and R.
Lekhaa, “Deep Learning Based Approach for
Gender Classification.” In 2018 Second
International Conference on Electronics,
Communication and Aerospace Technology
(ICECA), pp. 1396-1399, 2018.
[14]. Ramdhani, B., Djamal, E.C. and Ilyas, R.,
“Convolutional Neural Networks Models for

Volume 7, Issue 3, May-June-2021 | http://ijsrcseit.com 608