Real-Time Face-Age-Gender Detection System

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 568
Real-Time Face-Age-Gender Detection System
Lakshitaa Sehgal1, Anshal Aggarwal2
1University Institute of Engineering and Technology Chandigarh, Panjab University, Chandigarh, India
2University Institute of Engineering and Technology Chandigarh, Panjab University, Chandigarh, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - It is crucial to detect a person's gender and age
group since keeping track of who enters the building is
important from a security and other perspective nowadays.
A strict check and knowing the maximum information about
people entering a bank, ATM booth, state/country border,
data centers, hotels, restaurant, hospital, police station,
railway station, airport, etc. is vital. So, automated real-time
facial detection with gender and age group recognition
becomes the foremost concern. This paper proposes a real-
time “face-gender-age detection system” using Raspberry Pi
to capture real-time facial images of the people which are
then passed into the proposed model for gender and age
group prediction using the Convolutional Neural Network
(CNN). Using the proposed model, this facial image and
basic information is gathered and saved into the database,
so they can be easily retrieved and referred to in the future.
UTKFace and AGE, GENDER AND ETHNICITY (FACE DATA)
CSV are the datasets on which the model has been tested.
Additionally, noise is added to the images in the above
dataset while testing the model to see how well it performs
on noisy images.
Key Words: Human Age and Gender Detection,
Convolutional Neural Networks, Deep Learning, Facial
Detection, Computer Vision
1.INTRODUCTION
In recent times where crime is increasing day by day, it
becomes important to keep a check and store basic
information about the person entering the building,
specially public buildings like banks, cyber security
sensitive areas, country and state border crossing points,
police stations, highly security sensitive government and
private buildings hospitals, restaurants, multiplexes, shops
etc. Basic information which could be recognized through
the appearance of a person is age and gender recognition.
So here, to keep a track for security reasons, we have
proposed a system which would capture an image of a
person and store the image along with detected age and
gender in the database of the computer along with the
model using Convolutional Neural Network (CNN) for age
and gender prediction.
The Raspberry Pi 4[1] and a CCTV camera with NVR/DVR
or webcam with the attached system memory are used to
capture the real-time videos and storing them, which is
then passed through a face detection model, whose output
is used as the input for the proposed age and gender
detection model. The output of both the face detection
model and age and gender detection model is stored into
the database system of the computer.
To perform training on the datasets, a deep learning model
has been proposed. Different experiments have also been
performed on various datasets. Along with training and
testing on the original dataset, images have also been
trained and tested after adding salt and pepper noise to
the images. The proposed model has been trained and
tested on the following two datasets:
1) UTKFace Dataset[2]
2) AGE, GENDER AND ETHNICITY (FACE DATA) CSV
Dataset[3]
Our Main contributions are:
1) Proposed fully developed system to store basic
information (face, age and gender) of a person
into the database.
2) Trained and tested images (both with and without
adding noise) on a CNN model to detect the age
and gender of the person.
3) Performed fine tuning to enhance the robustness
of the model.
2. PROPOSED MODEL
This section presents the proposed system and the deep
learning model using convolutional neural networks for
detecting age and gender of a person.

2.1 SYSTEM SETUP
Fig -1: System Setup
The Raspberry Pi 4[1] and a CCTV camera with NVR for IP
cameras and DVR for analog cameras or webcam with the
attached system memory are used to capture the real-time
videos. The hardware setup captures videos, which are
then converted to images. A facial recognition model is
then run on these images, and its output is then passed
into our proposed model. The outcomes of both models
are then stored in the database.
2.2 PREPROCESSING
During preprocessing of videos, the OpenCV library
extracts images at 72 frames per second. The images are
then cropped off using MTCNN[4] for face detection, as it
is more accurate than the Haar Cascade classifiers[5] in
OpenCV. The final shape of the image after facial detection
is 48x48x3 before passing into the proposed model.
2.3 NETWORK ARCHITECTURE
Computer vision tasks, such as image classification and
detection, utilize deep, feed-forward artificial neural
networks such as convolutional neural networks (CNNs).
A CNN is similar to a traditional neural network, but with
deeper layers. The neurons of the CNN are arranged in a
volumetric manner, including height, width, and depth.
There are three dimensions to the input image(RGB
channels): width, height, and depth which is [48x48x 3] in
this model. An output of neurons connected to local
regions in the input will be computed by the convolutional
layer. A layer's parameters consist of a set of learnable
filters (or kernels) that are convolved across the width and
height of the input volume extending through its depth,
computing the dot product of the input and the filter
entries. Thus, the network learns filters that trigger when
it detects some particular type of feature at some
particular spatial position in the input by generating a 2-
dimensional activation map of the filter. A function called
the Rectified Linear Unit (ReLU)[6] layer will execute
elementwise activation, represented below.
(𝑥) = max(0, 𝑥)
Fig -2: Network Architecture
The maximal activation in a region is output by the pooling
layer. The spatial dimensions, such as width and height,
are down sampled in this way. The completely connected
layer, which is identical to the final layer of a neural
network, is the output layer.
This model has been trained using the sigmoid activation
function[7] for age determination and ReLU activation
function[6] for gender determination. Here, the binary
cross-entropy loss function along with the Adam optimizer
has been used.
3. EXPERIMENTS PERFORMED
This section elucidates the experimentation performed on
the datasets, Model Training and Hyper-Parameter tuning
and obtained results.
3.1 DATASET DESCRIPTION
Two datasets have been used for training and testing our
proposed model. Following are the two datasets:
1) UTKFace[2]
2) AGE, GENDER AND ETHNICITY (FACE DATA)
CSV[3]

UTKFace:
The UTKFace dataset[2] is a huge face dataset consisting
of people varying from age 0 to 116 years old. This dataset
contains over 20,000 face images with labeling of age,
gender, and ethnicity. The images cover large variations in
pose, facial expression, illumination, occlusion, resolution,
etc.
AGE, GENDER AND ETHNICITY (FACE DATA) CSV:
This dataset comprises data in the form of CSV format.
This includes a CSV of facial images that are labeled on the
basis of age, gender and ethnicity. This includes 27305
rows and 5 columns.
3.2 NOISE ADDITION
Salt and pepper noise is added to the images to test the
robustness and performance of the proposed model in the
presence of noise.
4. RESULTS AND OBSERVATIONS
Below are the results obtained after training and testing
the proposed model on both the datasets.
UTKFace Dataset
The proposed model, when trained on UTKFace[2]
without addition of noise, gave accuracy of 64% on age
estimation and 79% on gender estimation on 10 epochs.
Then after hyperparameter tuning the best outcome came
out to be 79% on age estimation and 93 % on gender
estimation on 60 Epochs, 64 Batch Size and 0.001
Learning Rate. After addition of salt and pepper noise the
best outcome came out to be 72% on age estimation and
86% on gender estimation on 70 Epochs, 64 Batch Size
and 0.0001 Learning Rate.
Chart -1: Testing Accuracy (Age and Gender Estimation)
vs Epochs for UTKFace (Without Addition of Noises in
Images)
vs Epochs for UTKFace (With Addition of Noises in
Images)
AGE, GENDER AND ETHNICITY (FACE DATA) CSV
The proposed model, when trained on AGE, GENDER AND
ETHNICITY (FACE DATA) CSV[3] without addition of
noise, gave testing accuracy of 67% on age estimation and
79.2 % gender estimation on 10 epochs. Then after
hyperparameter tuning the best testing outcome came out
to be 80.4% for age estimation and 94.2% on gender
estimation on 90 Epochs, 64 Batch Size and 0.0001
Learning Rate. After addition of salt and pepper noise the
results achieved the best testing accuracy of 70.3% on age
estimation and 85.6% on gender estimation on 110
Epochs, 64 Batch Size and 0.0001 Learning Rate.
vs Epochs (Without Addition of Noises in Images)

vs Epochs (With Addition of Noises in Images)
4. CONCLUSIONS
This paper puts forth an end to end face-gender-age
detection system and a deep learning model using
Convolutional Neural Network (CNN) which is trained and
tested on various different datasets. It captures real time
images and then detects faces. The output of facial
detection is further passed into the proposed CNN model
which has been trained for estimating age and gender.
These results are then stored in the database of the
computer system for further use.
We hope that this system will be incorporated in real life
in the near future. So, the results obtained demonstrate
that the model works fine for both images without noises
and containing noises and can be put into real time use.
REFERENCES
[1] Maksimovic, Mirjana & Vujovic, Vladimir & Davidovi´c,
Nikola & Milosevic, Vladimir & Perisic, Branko. (2014).
Raspberry Pi as Internet of Things hardware:
Performances and Constraints.
[2] https://www.kaggle.com/jangedoo/utkface-new
[3] https://www.kaggle.com/nipunarora8/age-gender-
and-ethnicity-face-data-csv
[4] N. Zhang, J. Luo and W. Gao, ”Research on Face
Detection Technology Based on MTCNN,” 2020
International Conference on Computer Network,
Electronic and Automation (ICCNEA), 2020, pp. 154-
158, doi: 10.1109/ICCNEA50255.2020.00040.
[5] L. Cuimei, Q. Zhiliang, J. Nan and W. Jianhua, ”Human
face detection algorithm via Haar cascade classifier
combined with three additional classifiers,” 2017 13th
IEEE International Conference on Electronic
Measurement & Instruments (ICEMI), 2017, pp. 483-
487, doi: 10.1109/ICEMI.2017.8265863.
[6] arXiv:1803.08375 [cs.NE]
[7] Pratiwi, Heny & Windarto, Agus & Susliansyah, S. &
Aria, Ririn & Susilowati, Susi & Rahayu, Luci & Fitriani,
Yuni & Merdekawati, Agustiena & Rahadjeng, Indra.
(2020). Sigmoid Activation Function in Selecting the
Best Model of Artificial Neural Networks. Journal of
Physics: Conference Series. 1471. 012010.
10.1088/1742-6596/1471/1/012010.
[8] Alqadi, Ziad. (2018). Salt and Pepper Noise: Effects
and Removal. International Journal on Electrical
Engineering and Informatics. 2.

Real-Time Face-Age-Gender Detection System

More Related Content

Similar to Real-Time Face-Age-Gender Detection System (20)

More from IRJET Journal (20)

Recently uploaded (20)

Real-Time Face-Age-Gender Detection System