IRJET - Face Recognition based Attendance System

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 03 | Mar 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 1899
Face Recognition based Attendance System
Mr. Vinay Sripathi
1
, Ms. Neha Savakhande
2
, Mr. Kiran Pote
3
, Mr. Parth Shinde
4
Guided By: Dr. Jayant Mahajan
5
1,2,3,4,5Department of Electronics and Telecommunication Engineering, Rajiv Gandhi Institute of Technology,
Mumbai, Maharashtra, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - In the field of image analysis and computer vision,
one of the most arduous tasks presently considered is Face
recognition. The biometric system which basically works on
the principle of face recognition is used for the identification
or verification of a person from a digitalized image preferably
used in surveillance, security and attendance purpose. Diving
into matter of concern, to preserve the record generated by
attendance of the student and co-operatingordinaryactivities
becomes a tedious task all together. The run-of-the-mill
method of calling the name of each student is time ingesting
and there is always a risk of proxy attendance. The defined
system is based on real-time multiple face recognition to
maintain the attendance record of students. The variation in
illumination and posing as well as focus issues and blurring
are important factor to be considered during face
identification process in the classroom using real-time
videography. This above-mentioned problem can be nullified
to some extent by the proposed system. The execution is
performed with the help of Histogram of Oriented Gradients
algorithm for detecting the faces, which is donebyitsexcellent
performance in differentiating feature descriptor of
individuals. The training data sets comprises of different
qualities of images. These data sets are fed to Convolutional
Neural Network, through which the result of recognition is
thoroughly obtained. Thus, making the attendance process
work uninterruptedly.
Key Words: Face Detection, Haar cascade classifier, Face
Recognition, HOG, Attendance System, CNN, Euclidean
distance, Face identification, Embedding.
1. INTRODUCTION
The process through which we can recognizea face,involves
certain steps that has to be mandatorily followed to get the
desired output. The steps are as follows, firstlyitbeginswith
collecting the strong data sets for the students in the class,
followed by training them with a good model, later on
detecting the faces with a good model at the time of actual
attendance, then feature extraction, classificationofthedata
acquired and recognition of faces, ending it by transferring
the faces recognized to a excel sheet throughCSV.During the
course of technological advancement in the world, plethora
of algorithms were developed and improvised especiallyfor
detection and identification. The implementation of the
overall project can subject to vary between real- time and
digital photo frame. However, the process of feature
extraction in both cases remains more orlessthesame.They
might include eyes, nose, lips, earsetc.when weskimaround
through various algorithms, the process of trainingthefaces
and storing in database and then matching it with the input
faces remains the same [1].
Albeit, face identification has diversified its popularity and
applications, it stands challenging task to execute. In real-
time facial recognition certainly, there are problems to be
dealt with such as illumination, focus, blur, pose etc. The
necessity of making the identification and recognition
buttress under a varying environmental situation, where the
captured images are likely to get deteriorated and also lose
some of the essential detailsduetoimproperlightconditions
such as, under and over exposure of the image or video
which is a major criterion for getting a functioning output.
Due to these mentioned problems, the images or video look
different at various time instant and leading it to a
misclassification of input image to the trained dataset
images. The mightier problem of all the problems that is
pose variation, it makes same images to look differentwhich
occurs due to change in viewing angles. Due tothis,itcreates
an obstacle while recognizing a face. In a face identification
model, detection plays an important and crucial role when
speaking overall. Various algorithms like histogram of
oriented gradients, Local binary pattern histogram, Fisher
faces eigenfaces. Our project inculcates histogram of
oriented gradients (HOG). Because we found working with
HOG more efficient and performance was, satisfactory HOG
takes the input image and juxtapose the pixels with other
pixels in terms of the darkness. Due to this, light problems
are solved to some extent and detection was accurate when
seen and observed in different conditions [2].
2. RELATED WORK
A) Haar Cascade
Haar classifier works through following steps:
1) Haar Features
As the training data is given to the model, theclassifierstarts
taking out features from each face. Haar features are
convolution kernels, whichconfirmsthata particularfeature
is present on an image, or not.
2) Integral Image
Haar Cascade classifier makes use of 24 x 24 base window
size, which produces more than 180,000 features. It is

difficult to compute pixel difference for all the facial
characteristics. By considering the four corner values of any
rectangle, all the pixels are added which results in integral
image.[12]
3) Adaboost
As 180,000 features are being generated in integral image
step. However, not all features are required to detect a face.
Therefore, to choose the best facial characteristics from the
whole set, adaboost a ML algorithm is implemented. This
decreases the features count from 180,000toapproximately
6000.
4) Cascade
Since it is difficult to keep all the 6000 extractions on a
window, the features are combined into different steps of
classifiers and applied one ata time. The window is removed
if it fails in primary stage. If the window passes, apply the
second step of facial characteristics and follow the process.
The window which passes throughallstepsisrecognizedasa
face region. [5]
B) Histogram of Oriented Gradients (HOG)
Histograms of Oriented Gradients (HOGs) are a feature
descriptor that has been widely and successfully used for
object detection. It represents objects as a single feature
vector as opposed to a set of feature vectors where each
represents a segment of the image.
In face identification steps, first we must find all of the faces
in an image. This detection is accomplished bymakinguseof
HOG within the acquired photo. Before applying HOG, the
input picture is converted from RGB into gray scale
photography for simplicity. Itiscomputed bysliding window
detector over an image, wherea HOGdescriptoriscomputed
for each position. By using an 8 x 8-pixel detection window
or cell, gradient vector or edge orientations at each pixel are
calculated. This generates 64 (8x 8) gradient vectors which
are then represented as a histogram. Each cell is then split
into angular bins, where each bin corresponds to a gradient
direction (e.g. x, y). This effectively reduces 64 vectors to
just 9 values. As it stores gradients magnitudes, it is
relatively immune to deformations. Histogram counts are
normalized so to curtail the loss due to illumination. This is
accomplished by accumulating measure of local histogram
gradient over the higher linked region and to support those,
the outcomes are used to normalizeall cellsintheblock.This
imperturbable positioning of histogram characterizes the
eventual hog descriptor. By withdrawing the descriptors
through only primary key points within the scale space of
the image using a rotation normalization, the scale and
rotation invariance can be found.[4]
Change in x and y direction can be calculated as:
Gx (y, x) = Y (y, x+1) – Y (y, x-1); Gy (y, x) = Y (y+1, x) – Y (y-1,
x)
Y (y, x) – pixel intensity
Gx (y, x) – change in x- direction
Gy (y, x) – change in y- direction
The values of change in x and y direction helps in finding a
gradient vector. We can compute the magnitudeandangleof
a vector as shown below:
Consider gradient vector as:
Magnitude = = 50.91
Angle = arctan = 0.785 rads
= 45 degrees
Working of HOG is shown below:
Fig 1: Block diagram for HOG.
C) CNN
A convolutional neural system (CNN) is a profound, feed-
forward counterfeit neural system in which the neural
system saves the various leveled structure by learninginner
component portrayals and summing up the highlights in the
basic picture issues like item acknowledgment and other
computer vision issues. It isn't confined to pictures; it
additionally accomplishes best in class brings about normal
language preparing issues and discourse acknowledgment.
Fig 2: Layer Structure.

Convolution means to extricate highlights from the info
picture, and thus it safeguards the spatial connection
between pixels by learning picture highlights utilizing little
squares of information. Rotational invariance, translation
invariance, and scale invariance can be normal.Forinstance,
a pivoted feline picture or rescaled feline picture can be
effortlessly recognized by a CNN as a result of the
convolution step. You slide the channel (square network)
over your unique picture, and at each given position, you
figure component insightful duplication (between the
frameworks of the channel and the first picture) and addthe
increase yields to get the last whole number that shapes the
components of the yield lattice.[2]
Fig 3: Subsampling
As appeared in Figure, channels have input loads and create
a yield neuron. Suppose you characterize a convolutional
layer with six channels and responsivefieldsthatare2pixels
wide and 2 pixels high and utilize a default walk width of 1,
and the default cushioning is set to 0. Each channel gets
contribution from 2×2 pixels, segment of picture. As it were,
that is 4 pixels one after another. Henceforth, you can say it
will require 4 + 1 (inclination) input loads. The information
volume is 5×5×3 (width × stature × number of channel),
there are six channels of size 2×2 with walk 1 and cushion 0.
Thus, the quantity of parameters right noweachchannel has
2*2*3 + 1 = 13 parameters (included +1 for inclination).
Since there are six channels, you get 13*6 = 78 parameters.
Fig 4: Input Volume
The pooling layers diminishthepastlayers'enactment maps.
It is trailed by at least one convolutional layerandmergesall
the highlights that were found out in the past layers'
enactment maps. This decreases the overfitting of the
preparation information and sums up the highlights spoken
to by the system. The responsive field size is quite often set
to 2×2 and utilize a walk of 1 or 2 (or higher) to guarantee
there is no cover. You will utilize a maximum activity for
each open field with the goal that the initiation is the most
extreme info esteem. Here, each four numbers guide to only
one number. Along these lines, the quantity of pixels goes
down to one-fourth of the first right now.[14]
Fig 5: Maxpooling-reducing the number of pixels
A completely associated layer is a feed-forward fake neural
system layer. These layers have a nonlinear initiation
capacity to yield class forecast probabilities. They are
utilized close to the end after all the highlights are
recognized and separated by convolutional layers and have
been solidified by the pooling layers in the system. Here, the
covered up and yield layers are the completely associated
layers.
D) Euclidean Distance
Euclidean distance is one ofthemostuseddistancemetrics.It
is calculated using Minkowski’s Distance formula by setting
p’s value to 2. Thus, the distance ‘d’ formula is shown below:
Euclidean distance formula, which looks similar to the
“Pythagoras Theorem”, can be used to calculate the distance
between two data points in a plane.
Fig 6: Euclidean Distance

It is shown in Figure, by using this formula of distance,
Euclidean space becomes a metric space. The Euclidean
distance between points x and y is the length of the line
segment connecting them [xy]. In Cartesian coordinates, if x
= (x1, x2, … , xn) and y = (y1, y2, … , yn) are two points in
Euclidean n-space, then the distance from x to y, or from y to
x is given by: 𝑑 (x, y). In three-dimensional Euclidean space,
the distance is:
𝑑 (x, y) = (x1 − y1 )2 + (x2 − y2 )2 + (x3 − y3 )2
The Euclidean distance between landmarks is used by most
authors as a morphometric measure. Once facial feature
pointsare obtained from a facial image or a two-dimensional
face, they selectsomesignificantdistancesbetweenthemand
compute the corresponding Euclidean distances. Then these
distances are used to compare faces for face recognition
systems. The Euclidean distances computed between all
possible pairs the facial feature points constitute a vector of
number of elements equal to the number of distances. This
vector gives the human face features of 2D image andusedas
input of classification algorithm for face recognition system.
3. PROPOSED SYSTEM
The proposed system is to develop a face recognition-based
attendance system with satisfactory rate of recognition from
a classroom. First and the most important step in proposed
system is to collect database of classroom students.
Databased collection is to be done using Haar cascade
classifiers to find face when webcam is turned on to collect
database (images of face) of a student.
Fig 7: Detected Face
The above image is the result of Haar Cascade, where face is
detected with the help of webcam and depending upon the
value of n, represented in the program the number of snap
shots is taken and stored as datasets.
Captured database images are stored in respective folder
with respective name of student. After collecting database of
all students, the encodings are to be extracted from those
images using CNN model.
Fig 8: Triplet Training
Extracted encodings are then stored with respective
student’s name in a pickle file.
At the third step, webcam is turned on to capture the
classroom. Faces are to be detectedusingHOG(Histogram of
Gradient) method. Detectedfaces’encodingsareextractedin
real time and are to be compared with those stored
previously with appropriate student names. Based on
maximum counts gathered after comparing the extracted
encodings and stored encodings, names identified in real
time. Internally compare_faces function is used to compute
the Euclidean distance between face in image andall facesin
the dataset. If the current image is matched with the 85%
threshold with the existing dataset, respective name is
identified and marked in the attendance report being
generated. Recognized names are then to be stored in a CSV
file with time & date. Main aim of the proposed system is to
develop a recognition algorithm that will identify maximum
no. of faces correctly in real time with low false acceptance
rate. The below block diagram refers to the complete
proposed system.
Fig 9: Block Diagram of Proposed System

4. Experimental Results
The results were generated in 3 different cases for
understanding the impact of training the datasets, light
conditions, pose variations.
Fig 10: Case-1
Case-1 was the first result through the proposed model. The
arrangements of the model were done in a classroom in
presence of little sun light and moderate classroom tube
lights. Due to uneven lightening condition i.e. light coming
from the window dominated the light whichwasintheroom
resulting in undersaturation on the face and oversaturation
around the window thereby leading to drop in efficiency
30.77%.
Fig 11: Case-2
Case-2 was the second attempt which was performed in the
Laboratory. This time around we madesureofthelightening
conditions by balancing the light in front to the back side of
the image. Efficiency increased by 7.69% compared to the
case-1. But efficiency was not up to the mark due to the data
set not containing the various pose variations of the
students, meaning, the datasets in the database only
contained frontal faces of the students.
Fig 12: Case-3
Case-3 was accomplished in thelaboratorywithjustthetube
lights. This case led us to the 100% efficiency due to even
lightening condition and strong dataset generation. The
dataset contained the frontal faces as well as side faces,
resulting total faces for an individual about 50 images.
The below Table shows detailed result of all the cases.
Table 1: Result
There are many papers based on algorithms like HOG and
CNN combinations. For example [14]. The working of the
model remains same but interpretation with respect to
dataset and classification of images changes. By making use
of Euclidean Distance for image Classification the hardware
(laptop) is able to process the datasets as well as recognize
more than 5 students. When using SVM as a Classifier the
processing gets too complex for a decent laptop to perform
the process. So, the maximum possible students which can
be detected through the SVM classifier is 3 to 4 [14].
5. Conclusion
This system gives satisfactory results if conditions of
capturing images is kept ideal i.e. sufficient illumination and
enough head pose variation in dataset. ThesystemusesHOG
for face detection and CNN encodings to recognize multiple
present faces in the frame, but the accuracy and efficiency of
recognition process is hampered by factors like false
acceptance due to mismatched encodings, poor lighting and
occlusion. Therefore, the proposed system is useful in
classroom attendance but its shortcomings cannot be
ignored either.

6. Future scope
As the results state, the recognition efficiency drops as the
number of students increase in a frame. Also, false
acceptance rate is prone to increase if database is not
prepared carefully i.e. using head-pose variation,
illumination variation etc. The efficiency of CNN encodings
which are used to identify each face, can be improvedbyfine
tuning and training the Convolutional Neural Network
algorithm according to generated database and target
conditions of classroom. The proposed system is currently
capable enough to being used in actual classrooms or
laboratories or other places in organization which needs
daily attendance report to be maintained.Thiscanbefurther
improved in terms of accuracy of CNN, User interface,
connectivity and implementation of report in large systems
which maintain organizational report.
REFERENCES
[1] Manisha M. Kasar1, Debnath Bhattacharyya1 and Tai-
hoon Kim2 – “Face Recognition Using Neural Network: A
Review” International Journal of Security and Its
Applications Vol. 10, No. 3 (2016)
[2] Navneet Dalal and Bill Triggs - “Histograms of oriented
gradients for humandetection”ComputerVisionandPattern
Recognition” 2005. CVPR 2005. IEEE Computer Society
Conference on. Vol. 1. IEEE. 2005, pp. 886–893
[3] Radhika C. Damale, Bazeshree. V Pathak – “Face
Recognition Based Attendance System Using Machine
Learning Algorithms”2018SecondInternational Conference
on Intelligent Computing and Control Systems (ICICCS)
[4] Mohsen Ghorbani, Alireza Tavakoli Targhi, and
Mohammad Mahdi Dehshibi - “HOG and LBP: Towards a
robust face recognition system” Digital Information
Management (ICDIM), 2015 Tenth International Conference
on. IEEE. 2015, pp. 138–141
[5] Neha Savakhande1, Vinay Sripathi2, Kiran Pote3, Parth
Shinde4, Prof. Jayant Mahajan5 – “Face Recognition based
Attendance System: Review” International ResearchJournal
of Engineering and Technology (IRJET) Volume:07Issue: 02
| Feb 2020
[6] Vishakha Mehta, Sarika Khandelwal, Ashish Kumar
Kumawat – “A Survey on Face Recognition Algorithm” 2018
2nd International Conference on Trends in Electronics and
Informatics (ICOEI), 11-12 May 2018
[7] Vinay et al - “Effective Descriptors based Face
Recognition Technique for Robotic Surveillance Systems”
Procedia Computer Science 133 (2018), pp. 968–975
[8] Gang Zhang, Sikan Tang, and Jiaquan Li - “Face landmark
point tracking using LK pyramid optical ﬂow” Tenth
International Conference on Machine Vision (ICMV 2017).
Vol. 10696. International Society for Optics and Photonics.
2018, 106962B
[9] Yuqian Zhou,Ding Liu,Thomas Huang – “Survey of Face
Detection on Low-Quality Images” 2018 13th IEEE
International Conference on Automatic Face & Gesture
Recognition (FG 2018),15-19 May 2018
[10] Anushka Waingankar1, AkashUpadhyay2,RuchiShah3,
Nevil Pooniwala4, Prashant Kasambe5 – “Face Recognition
based Attendance Management System using Machine
Learning” International ResearchJournal ofEngineeringand
Technology (IRJET)
[11] Thanh Tan Nguyen Thi, Khanh Nguyen Trong – “An
Efficient Face Detection and Recognition” International
Journal of Innovative TechnologyandExploringEngineering
(IJITEE) ISSN: 2278-3075,Volume-7Issue-5,February,2018
[12] Li Cui-mei, Qi Zhiliang – “Human face detection
algorithm via Haar cascade classifier combined with three
additional classifiers”
[13] Mauzzam Siddiqui, Dr.Preetam Suman – “A Review on
Face Image Recognition & ClusteringbyUsingVariousTypes
of Approaches in Deep Learning” International Journal of
Computational Engineering Research (IJCER)
[14] R. Angeline1, Kavithvajen.K2, Toshita Balaji3, Malavika
Saji4, Sushmitha.S.R5– “CNN Integrated with HOG for
Efficient Face Recognition” International Journal of Recent
Technology and Engineering (IJRTE) Volume: 07 Issue: 06 |
March 2019
BIOGRAPHIES
VINAY SRIPATHI
Currently pursuing under
graduation final year in the branch
of EXTC at Rajiv Gandhi Institute of
Technology, Versova.
Team Leader of the Project Group.
NEHA SAVAKHANDE
Team Member of the Project Group.
PARTH SHINDE

KIRAN POTE
Dr. JAYANT MAHAJAN (Ph.D.)
Project Guide at Rajiv Gandhi
Institute of Technology, Versova.

IRJET - Face Recognition based Attendance System

More Related Content

What's hot (19)

Similar to IRJET - Face Recognition based Attendance System (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET - Face Recognition based Attendance System