Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)

IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

Convolutional Neural Network based Automated

Attendance System by using Facial Recognition
Domain
Shailender Kumar Dhruv Kathpalia Dipen Singh M andeep Vats
Department of Computer Department of Computer Department of Computer Department of Computer
Science & Engineering Science & Engineering Science & Engineering Science & Engineering
Delhi Technological University Delhi Technological University Delhi Technological University Delhi Technological University
Delhi, India Delhi, India Delhi, India Delhi, India
shailenderkumar@dce.ac.in kathpaliadhruv@gmail.com dipen.singh.19@gmail.com mandeepvats98@gmail.com

Abstract - This project ai ms to recognize faces in an single image. Before facial recognition can be carried out, we
image, vi deo, or vi a li ve camera using a deep learni ng- must first ensure the presence of a face in the frame. This can
based Convolutional Neural Network model that is fast as be done by performing face detection. In this step, the model
well as accurate. Face recog nition is a process of detects the face and separates it from the image for
identifying faces in an i mage and has practical identification, eliminating redundant data that is not required
applications in a variety of domai ns, including information for facial recognition. This reduces the number of pixels on
security, biometrics, access control, l aw enforcement, which the model has to work on and hence increasing the
smart cards, and surveillance system. Deep Learning uses overall efficiency.
numerous layers to discover interpretati ons of data at However, facial recognition [12] also faces some problems,
di fferent extraction levels. It has improved the l andscape making it very hard to perform. Various factors like pose
for performing research in facial recogniti on. The state-of- variation, facial hair, image illu mination, image background ,
the-art i mplementation has been bettered by the and facial expressions affect the image, and the outcome can
introducti on of deep learning in face recognition and has differ based on these characteristics. In situations where the
stimulated success in practical applications. Convolutional face is not visible or hidden from the camera, the face might
neural networks, a kind of deep neural network model has not even be detected. Thus, the image used as input to the
been proven to achieve success in the face recognition model could be in d ifferent conditions as opposed to the image,
domain. For real-ti me systems, sampling must be done which is to be examined.
before using CNNs. On the other hand, complete i mages We are attempting to instill this technique in universities,
(all the pixel values) are passed as the input to so that common issues like h igher time-consumption during
Convolutional Neural Networks . The following steps: attendance, students marking pro xy attendance and mass bunks
feature selection, feature extracti on, and training are during lectures can be prevented. Marking attendance in
performed in each step. This might lead to the assumption, classes is an overwhelming task for the professors as it is not
where convolutional neural network implementation has a only time consuming, but also the students tend to mark pro xy
chance to get complicated and time-consuming. attendance, which leads to inaccurate records of attendance.
Manual attendance is certainly tough for the professors , as it
Keywords – Face detection, face recognition, deep makes it difficult to maintain a record of the students. The
learning, convolutional neural networks conventional ways often have their difficu lties. The majority of
these methods lack dependability. It leads to an increasing need
for better methods of attendance. This research stresses on
using facial recognition as a technique for marking attendance.
I. INT RODUCT ION Real-time automated attendance monitoring without wasting
Face recognition is a unique technique for performing teacher’s precious time is the main objective of this project.
authentication biometrically. It has broad applications in areas Not only does this method save time, but it is also more
of finance, security, and military. Face recognition has gained a reliable than traditional methods .
lot of interest in the last few years, which has led several In the next section, we discuss some of the facial
researchers to work for developing new techniques and recognition algorithms, assessing their pros and cons. In
improve the existing ones. Its wide range of applications Section III, we present our proposed model and briefly exp lain
appeals to researchers and keeps them driven. Face recognition how it works. The experimental results obtained after testing
can be performed on a real-time v ideo by considering it as a the model are given in Section IV. Section V concludes the
sequence of frames where each frame is considered to be a paper with a summary.

978-1-7281-4876-2/20/$31.00 ©2020 IEEE 654

Authorized licensed use limited to: University of Exeter. Downloaded on June 21,2020 at 15:05:08 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

Figure 1: Flow-chart showing the working of the facial recognition aspect of our paper

II. RELAT ED W ORKS

Haar cascades [10] use different features that are trained
Dr. Priya Gupta et al. [1], obtained an accuracy of 97.05% first with the help of a training set, which consists of both
using their proposed method. They performed feature positive as well as negative images. After we have trained the
extraction with the help of Haar Cascades, which were, in turn, classifier, it can be used to detect if the object is present in any
fed forward to the network instead of raw pixel values. The test image or not. These have been immensely useful in tasks
complexity was greatly reduced as it led to a decrease in the related to facial detection as they make the process much faster
redundant input features. It uses Deep Neural Networks , which as it has considerably fewer co mputations as compared to other
makes the model very efficient in terms of using fewer methods. Different filters can be used for features such as eyes,
resources and making it faster. nose, mouth, etc. as shown in figure 1.
R. Rahim et al. [2], used the Fisher Linear
Discriminant(FLD), wh ich was discovered by Robert Fisher in
1936. It is a popular pattern recognition method, which has
applications in the face and object recognition. It increases the
accuracy of the classification by forming inter-class, and intra-
class scatters. The algorithm can recognize the faces even with
changes in some attributes of the faces, like expressions or
wearing glasses.
Faizan Ahmed et al. [3], shows a comparative study on the
different methods for performing facial detection and
recognition. They achieved an accuracy of 96.7% using the
Adaboost classifier [4] along with Haar features and an
accuracy of 90.88% using Support Vector Machine(SVM ) [5]
classifier for Face detection application. For Facial
Recognition, they used the following methods:

Figure 2: Haar Features applied to a Face

978-1-7281-4876-2/20/$31.00 ©2020 IEEE 655

Authorized licensed use limited to: University of Exeter. Downloaded on June 21,2020 at 15:05:08 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

i. Principal Co mponent Analysis(PCA) [6] method  A video camera placed in the classroom would
which uses the concept of Eigenfaces, with an continually record the class and pass the input stream
accuracy of 71.15%. The co mputations required in
this method is much lower as compared to the other to the attendance system.
methods as it only considers the 2D face recognition  At regular intervals, individual frames are analyzed by
problem. Thus the complexity is reduced by a the model.
substantial amount.
ii. Linear Discriminant Analysis(LDA) [7] method uses  All the faces of the students for a particular frame
Fisherfaces, with an accuracy of 77.90%. It mainly
would be detected and, in turn, recognized by our
focuses on reducing the number of features being
model.
applied on each face.
iii. Local Binary Pattern(LBP) [8] has an accuracy of  The same process would be followed for each frame
82.94%. for real-time video analysis and facial recognition.
iv. Gabor Classifier [9], which takes into consideration  At the end of a class, after all the frames are done, the
the local features, has an accuracy of 92.35%. It is not model would mark the students as absent or present
designed specifically for face recognition, but its based on the model's results.
filters can recognize various prominent features in an  The student must be recognized in at least 60 percent
object. of the frames to be marked present.
 The database would update the information on
Convolutional Neural Networks [11] are a type of Neural marking the attendance of all the students for that
particular date.
Network that is mostly used in the field of image classification,
 The database also stores the attendance record for all
particularly Face Recognition. Convolutional Neural Networks dates. Therefore it will also make it easier to manage
take an input image and tweak the weights of the network- the attendance records.
based of the input image so that it can differentiate it fro m
other images. This allows the network to learn and identify the Pre-processing
important characteristics (that are essential for recognizing
Our model is based on deep learning-based facial
different faces) on its own. The need for human supervision is
recognition. We have used the face_recognition library that is a
thus min imized- it can automatically differentiate the images pre-trained network and uses it to generate 128-d vectors from
into separate classes. Convolutional Neural Networks also the training dataset. Pre-processing involves the following
reduce the need for pre-processing required to train the model, steps:
thus it utilizes less computation power. Due to these
advantages, deep learning algorithms like convolutional neural Step 1: Face Detection
networks have become the standard in facial recognition.
 Convert the given image to grayscale.
III. PROPOSED W ORK  Apply Haar features to each image by dividing the
image into smaller squares and detect the presence of
We have proposed a method for an automated attendance different features such as edges, corners , etc.
system using facial recognition. The system should be able to  We obtain a basic structure of the image, representing
detect faces in each frame of a real-time video. Further, after the obtained features.
recognizing the detected faces, it should be able to mark the  We can compare this structure to a previously
attendance of students whose faces are recognized by the extracted pattern of a face. This helps us in identifying
system. The cru x of the system is that it marks the attendance the different faces present in the image.
of only those students who have attended more than sixty
Step 2: Resolving the issue of projecting faces
percent of the total class; the rest of the students are marked
absent.  Although we have isolated faces, the computer
mistakes a face looking in different directions as
The proposed system requires a video camera in the class to
different faces.
be an initial requirement. The proposed system is designed to  To resolve this issue, we need to alter the positioning
work on video footage of students. The underlying idea is to of a face in an image.
extract features of students’ faces from the footage, and  We take some points that lie on a face and using these
compare these features with those which are extracted fro m the points, we manage to detect a face, its boundary, the
training images used for training the model. If these features positions of eyes, nose, and lips from an input image.
match, the student is marked present for that single frame.

978-1-7281-4876-2/20/$31.00 ©2020 IEEE 656

Authorized licensed use limited to: University of Exeter. Downloaded on June 21,2020 at 15:05:08 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

 Next, the image is rotated such that the transformed i.e., to detect and recognize the faces in the input data, is shown
face in the rotated image is as close as possible to a in figure 3. The only step left is to mark the attendance of
perfectly centered face. students whose faces are classified in the previous step.
 Using this method, the computer does not categorize
the projections of a face as different faces.

Step 3: Encoding Faces

 The faces that the model detected cannot be compared

to each face in the database, as this method would
require a lot of time. Figure 3: Steps involved in Classification of a face
 Instead, we have to take an approach that uses only a
few metrics for each face. Metrics such as distance The facial recognition is performed on each frame of the
between the eyes or the shape of the nose aren’t the real-time video. So, each frame contributes to the whole
most accurate in differentiating between faces. attendance. Being present in only a single frame doesn’t mean
 The metrics that are used are measurements that the that the student would be marked present. This is due to the
network defines on its own, using Deep learning. algorithm not being 100% accurate, as we have to consider
 Here the CNN will be trained to create 128 vectors for situations where the students’ faces are not visible. It is almost
each face. impossible to receive such high accuracy. Thus, we try and
 Now, we take a trip let: 2 images of the same known tackle the problem in a way that ensures that the students that
person and one of a different person. The are present in that class are only marked as present and not the
measurements for each image are generated. ones that may be incorrectly recognized in a single frame. Th is
 The network learns by adjusting so that the first two also ensures that if a student’s face isn’t recognized in one or
images are closer, and the measurements of the third two frames due to any issue, they are still marked present.
image are farther. The following steps are taken to ensure the above:
 The neural network does this repeatedly for thousands  We create a variable count that stores the count of the
of images so that for the same face, two d ifferent number of frames that each student’s face has
images give nearly the same 128 measurements. appeared in. We initialize this variable to 0 for each
 Huge amounts of input data and processing power is student.
required to train a CNN using this process. However,  We also create a total count variable that stores the
once trained, the network can be used to generate total number of frames.
accurate measurements, even for new faces.  For each frame, we do the following:
 Hence, now the network (that is already trained) only  Detect the faces in the frame.
requires our images to calculate the 128  Recognize the faces by matching each of
measurements accurately. them to the encodings that we have fro m our
training set.
Step 4: Finding a Match  Increment the count corresponding to the
 First, we detect the faces and match each of them to student.
the encodings that we have from our training set.  If a face is unknown, we skip that face
 The Euclidean distance is calculated with each of the  These steps are followed until we reach the
faces in our database: last frame, i.e., the end of the class.
o The distance must be below some threshold  Now we co mpare the count for each of the students to
value to predict that the face matched. the total count. If the count for a student is greater
o If the distance is higher than the threshold, than 60% of the total count, then we would mark the
then we consider the face to be unknown. student as present in the database, else we would mark
him/her as absent.
Processing the Input Data  This ensures that the student will only be marked
present if he has attended at least 60% of the class.
The input data is the video stream fro m the video camera
fro m the classroom. As the video stream is a collection of Also, this helps in improving the accuracy of the system as
frames, we apply the face detection module to each of the even if a face is recognized incorrectly, its effect would almost
frames. Once all the faces in a frame are detected, they are then be negligible on the final result.
compared with the generated pickle file fro m the training
dataset. The faces that are matched are classified with the name We have proposed a method here in which the model has
of the student, whereas the faces that have no match are been trained and tested on our customized database that
classified as unknown. The initial step of processing the data, contains images of 16 students as subjects. To find an effective

978-1-7281-4876-2/20/$31.00 ©2020 IEEE 657

Authorized licensed use limited to: University of Exeter. Downloaded on June 21,2020 at 15:05:08 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

number of images per person required for our model to provide Further, the pickle file generated is used in the model to
the best result, we trained and tested the model with a d ifferent recognize the faces, instead of training the model every time
number of images per person. Starting fro m a single image of the model is given an input.
each person, we performed the testing with as many as 25 We analyzed our model both quantitatively as well as
images per person. The images of every subject differ in their quantitatively. We first assessed the number of faces the model
facial expressions, contrast, exposure, configuration, etc. We was able to detect, i.e., quantitative analysis, when given real-
implemented the model on a 64-bit system using Python 3.6.9. time video input, and then measured the accuracy of the model
For the pre-processing of input images, we have used the by calculating the number of faces that were recognized
OpenCV package using Haar cascade and its frontal face correctly, i.e., qualitative analysis.
feature.
Also, the result of our framework, when stored in the
database is sorted according to the dates on which the lecture
has taken place. Initially, all the students for a lecture are
marked absent and on recognizing their faces, and if they
follow certain criteria mentioned above, the model updates
their attendance in the database marking them present for the
lecture. A snapshot of a part of a database, showing the names
of only five students, is shown in Fig. 5. We have used
WampServer for our MySQL database.

Figure 5: A snapshot of the Attendance marked in the Database

Once we get the number of faces recognized in a frame, we

store the list of such students. For each student, we maintain a
count that contains the number of frames in which a student’s
face is present. And at the end of the session, if a student’s face
is present in 60 percent of the total number of frames, the
student is marked present. Failing to fulfill the required
criterion, the student is marked absent.
Table 1: Images used for Model Training and Testing

Test Images Total Students Accuracy

Case Per students marked present
Person present in 4 classes (%)
in the (average)
Training
set
Case 1 1 13 7.25 55.77

Case 2 3 13 9 69.23

Case 3 5 13 10 76.92

Figure 4: Flow Chart for Data Processing Case 4 10 13 10.75 82.69

IV. EXPERIMENT AL RESULT S Case 5 15 13 11.5 88.46

All the images are then compressed, using a compression
algorithm, to a suitable size that it contains all the important Case 6 20 13 12 92.31
features required for recognition of faces and does not take too
much time to extract the classifiers fro m the training images, Case 7 25 13 12.5 96.15
converting them to 128-d vector and creating a pickle file.

978-1-7281-4876-2/20/$31.00 ©2020 IEEE 658

Authorized licensed use limited to: University of Exeter. Downloaded on June 21,2020 at 15:05:08 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

96.15%. A lthough the accuracy is assumed to be very low for

fewer images, it is compensated by the extra step that ensures
On performing training and testing with a different number
that the student is marked present only if the nu mber of frames
of images per person, we noticed that on increasing the number their faces are identified is greater than the predefined
of images of a person, the number of important features in the
threshold of 60%. This results in accuracy that is much higher
pickle file from the compressed images also increases. This than expected. This model can also be applied to online
further increases the probability of the system comparing and
classes. During online lectures too, the conventional methods
deciding the detected face correctly with the features extracted would waste precious time. They cannot be considered very
from the pickle file.
reliable either, as anybody can log in as a student if they have
In our case for automatic attendance based on facial the login credentials. Instead of manually taking attendance,
recognition, the system admin can take multip le images of the which might be tedious for large groups of students, attendance
student at the time of registration, which would be used for the will be taken automatically in the background. Facial
training dataset for high accuracy. This makes the proposed recognition would ensure that attendance is reliable.
attendance system reliable for marking attendance of students
during lectures in the universities. The trend analysis of a
different number of images per person on the accuracy of the REFERENCES
system, as observed during the testing of the system is given in [1] P. Gupta, J. T ripathi, M. Sharma, and N. Saxena, "Deep Neural Network
Fig. 6. for Human Face Recognition," MECS, p. 9, 2018.
[2] R. Rahim, "Research of Face Recognition with Fisher Linear
Discriminant," IOP Conf. Ser., 2018.
[3] A. Najam, Z. Ahmed and F. Ahmed, "Image-based Face Detection and
Recognition," IJCSI, 2012.
[4] A. Tikare, S. Kadam, and K. T . Talele, "Efficient Face Detection using
Adaboost," in IJCA Proc on International Conference in Computational
Intelligence, 2012.
[5] B. McCane and I. Kukenys, "Support Vector Machines for Human Face
Detection," in Proceedings of the New Zealand Computer Science
Research Student Conference, 2008.
[6] M. A. T urk and A. P. Pentland, "Face recognition using Eigenfaces," in
Proceedings of the IEEE, 1991.
[7] J. Lu, N. Plataniotis, and A. N. Venetsanopoulos, "Face recognition
using LDA-based algorithms," in IEEE Neural Networks Transaction,
2003.
[8] T. Ahonen, A. Hadid, and M. Peitikainen, "Face recognition with local
binary patterns," in “In Proc. of European Conference of Computer
Vision, 2004.
[9] L. Wiskott, M. Fellous, N. Krger, and C. Malsburg, "Face recognition
by elastic bunch graph matching," in IEEE Trans, 1997.
[10] P. Viola and M. Jones, “Rapid Object Detection using a Boosted
Figure 6: Graph depicting dependency on the accuracy of the model on the Cascade of Simple Features,” IEEE Computer Society Conference on
size of the training set Computer Vision, 2001.
[11] Yamashita, R., Nishio, M., Do, R.K.G. et al. Convolutional neural
networks: an overview and application in radiology. Insights
V. CONCLUSION Imaging 9, 611–629 (2018).
Using Convolutional Neural Network for facial recognition [12] “A Summary of literature review: Face Recognition” by Kittikhun
helps in reducing time and the processing power used as Meethongjan & Dzulkifli Mohamad, August 2007, Postgraduate Annual
Research Seminar 2007
compared to other conventional methods. The model has great [13] King, Davis E. "Dlib-ml: A machine learning toolkit." Journal of
accuracy. For 25 images per subject, we achieve an accuracy of Machine Learning Research 10.Jul (2009): 1755-1758.

978-1-7281-4876-2/20/$31.00 ©2020 IEEE 659

Authorized licensed use limited to: University of Exeter. Downloaded on June 21,2020 at 15:05:08 UTC from IEEE Xplore. Restrictions apply.