Face Recognition For Attendance Using

Transfer Learning

Vasu Rawat Parwan Singh

Department of MCA Assistant professor
University Institute Of Computing University Institute Of computing
Chandigarh university Parwan.e15436@cumail.in
rawatvasu03@gmail.com

Abstract - Taking attendance plays a

significant role in monitoring availability If the traditional way of recording attendance wasn't
and performance of the relevant tedious in itself, institutes have to maintain the
stakeholder in an organization. In an physical copies of the data too. With the
academic environment, the importance of advancements in technology, we can now record
attendance is even more significant as attendance with the help of devices that can recognize
students’ efficiency and conduct is highly the unique features of the people registered with it
relevant. The traditional methodology such as - iris, fingerprint, face, voice, etc.
induced in attendance tracking requires
the use of various non-value-added Algorithms for computer vision and machine learning
activities leading to unnecessary are used to do this. We can also reduce human
consumption of time and resources which mistakes by producing findings with great precision.
have additional security challenges. In The approach outlined in this study offers a transfer
order to overcome such issues, the present learning-based solution for face recognition.
study introduces the use of deep learning
approach in association with recognition Transfer learning is the technique of using an already
techniques. Face recognition involving trained model and making changes in the uppermost
retinal specification in association with layers so that they solve the newly proposed problem.
transfer learning is an emerging technique
which helps users to use existing dataset II. LITERATURE SURVEY
and generate accurate prediction without
[1] In this paper, the author approaches face
compromising the security issues. The verification checks to match an individual based on
study shows the competence of acquiring its multiple known profiles present in the system. In
and managing attendance without human order for the Siamese network to learn the properties
intervention ensuring optimum level of of the provided photos, the author feeds two faces
security and reliability. into it as input using the CNN structure AlexNet.
The best learning procedure and maximum values
Keywords: Face Recognition, Deep are then determined by retraining the network
Learning, Transfer Learning using the ReLU activation function. The datasets
for spectral resolution and 2D face photos are
I. INTRODUCTION tested and improved in this paper using the
The traditional way of recording attendance is not AlexNet framework. The high dependability of
only time consuming but the probability of having each spectrum demonstrates the potency of deep
human errors is also very high. Taking attendance learning technologies for face identification. The
does not seem like a time-consuming task until one paper's output is a categorization and a result
actually starts recording it. prediction. The accuracy can be increased to up to
98% by using a variety of alternative photos.
 a facial recognition system that can recognise and
[2] The Haar cascade model, which matches classify a user's face in its entirety. Unauthorized
the student's face with faces that have previously users are prevented from accessing the security
been saved in the database, is suggested here as a system. It utilises the AlexNet Deep Learning
technique for creating a system that helps programme. This particular model has been
instructors keep track of student attendance. To trained to recognise faces, and it allows the user to
recognise the images in the classification system, submit the faces of several people at once. The
a classifier is made by merging a number of Haar- system enhanced its capabilities to thwart
like features. Applications for detection and cyberattacks and other security threats, and it
recognition use ANNs (Artificial Neural successfully demonstrated its high levels of safety
Networks) and local binary pattern histograms. and security.
This evaluation shows that the recognition can
detect up to ten faces simultaneously, is up to 80% [6] This system, which was created with the
accurate, and can be utilised by non-IT help of a pre-trained model and the Dlib library,
professionals. However, if the image resolution is accurately recognises the faces of the students in
poor, it is more likely to fail. the frame and consistently records attendance
while simultaneously maintaining an excel sheet
to store the data. After noise reduction, an HD
[3] The author of this study addressed the 1080p camera's facial features are recognised
problem of recognising persons wearing masks by using a histogram-oriented gradient (HOG)
using a simple yet effective approach. The three method. Institutions may also use this gadget as a
steps of feature extraction, identification, and totally reliable surveillance system. Here, HD
detection are how the author divides the problem. cameras are used to increase productivity. They
Labelled faces in the Wild (LFW) is used to train enable the system to recognise even little faces by
the system, which then uses data augmentation to enlarging the picture pixel and frame size.
create a masked clone of the original facial
recognition data set. Then, the two datasets are
combined once more for the training phase. The [7] An enhanced system for face detection and
recommended model states that feature extraction recognition will be the end result of the author's
in the LFW database used ResNet 50, ResNet with approach, the Transfer Learning Model for Face
Soft max classifier, and obtained 98.3 testing Recognition (TLMFR), which was presented in
accuracy, demonstrating ResNet 50's higher this article. To determine if two faces are similar
performance for simple face recognition when or not, transfer learning is employed. The channels
masks are applied. were shown using the GRAD CAM approach,
emphasising the areas of the images that are most
important for recognising an object as a certain
object. An assessment of two data structures with
[4] The ORL and CHUK datasets were used by both comparable and distinct faces showed that
the authors of this research to conduct model weights could correctly identify the
experiments and train the AlexNet convolutional important situations in both cases.
neural network. The author starts by discussing
the issues that came up during the training of the [8] This work uses the Haar Cascade model,
classes for facial recognition before suggesting a which combines Haar-like characteristics to create
feature transform framework to resolve the issues. a classifier that can identify pictures in the system.
188 facial pictures for the CHUK face database Here, the backpropagation method is employed
and 400 for the ORL face database were used in for the face recognition learning process of an
this investigation, which had a testing phase artificial neural network (ANN). Here, it was
accuracy of 99% for CHUK and 98% for ORL. determined that combining local binary patterns
But it still needs a huge dataset to be shown as the with histograms of characterised angles (HOG)
most efficient framework for face recognition. directed improved the safe execution of more
informative indexes. With a detection rate of 95%
[5] With the use of images of the user that are and identification rate of 85% over a 0-degree
kept in a database, the author of this study creates front face, which decreases over different angles
of the front face like 20 and) at 70 and 90 degrees, III. MATERIALS AND METHODS
the system built here is readily controlled by a Methodology
non-IT person. It can recognise up to 10 faces.

[9] The proposed article uses a one-shot

learning approach model technique in this
instance, which only needs one photograph of
each student to train the system to recognise their
faces and track their attendance in an Excel sheet
in line with that data. The authors created this
method utilising the face-recognition deep neural
network, which was built using Dlib and has an
accuracy of roughly 99.38% on the LFW
(Labelled Faces in the Wild) dataset. Because the
authors' method only recognises one student at a
time, it takes longer to mark attendance than the
manual or traditional methods because it cannot
register attendance for the entire class in the
allocated time.

[10] The authors of this study have used CNN-

PCA as a hybrid feature extraction method
(Convolutional Neural Network Principal
Component Analysis). This method combination
aims to produce a more accurate feature extraction
strategy. The recommended research employs
CNN-PCA and obtains an accuracy of 90%-98%,
proving that CNN-PCA can be more efficient than
PCA alone. This system can reach an accuracy of
90%-96% when utilising PCA for feature
extraction.

Fig. 0 Process of Transfer Learning

WHAT IS TRANSFER LEARNING?

1. Obtain the pre-trained model Getting a pre-

trained model that best fits your situation is the first
step. These pre-trained models may be found on
Hugging Face, Keras, TensorFlow Hub, etc. Each
model that you choose to download comes with pre-
trained weights.

2. Create a base model

First, using one of the architectures, such as ResNet
or Xception, we instantiate the underlying model.
You will need to utilise the architecture to train your
model from start if you don't have the weights. The
base model already has layers; all that is needed to
make it suitable with our situation is to take away
the last output layer and replace it with another one.
3. Freeze Layers Experimental Procedure
The pre-trained model's layers must be frozen if you
The top layer receives each image as a
don't want the model's weights to change while it is
2048dimensional vector. On top of this
being trained. You must do this in order to preserve
representation is a SoftMax layer that has been
all of the prior learning that your model has
trained. It will be essential to learn the (N +
undergone. It will be the same as starting the model
2048*N) model parameters that match to the
from scratch if this is not done.
learned biases and weights if the SoftMax layer
has N labels.
4. Add new trainable layers
Old features will be converted into predictions on By default, this script will employ the Inception
the new dataset by adding new trainable layers. v3 model architecture, which has great accuracy
Because the pretrained model is loaded without the but is also rather big and sluggish. It is advised
final output layer, this step becomes crucial. that we begin with this to confirm that we have
acquired high-quality training data.
5. Train the new layers on the available
dataset The ultimate output of the pre-trained model
will probably differ from what you anticipate. For STEPS OF THE PROPOSED SYSTEM
instance, your model could only have two classes
compared to the over 1000 classes in pre-trained 1. Capture Images
models that were trained on the ImageNet dataset. In 2. Training the model
this situation, we train the model using a new output 3. Marking Attendance
layer.

We may add as many dense layers as we like, but

the most crucial one is the final dense layer, which
should have units equal to the number of outputs
predicted by your model.

6. Improve the model via fine-tuning This is

Fig. 1 UI of the proposed system
an optional step, but if taken, we can improve the
output accuracy of our model. The sole purpose of
fine-tuning is to enhance the model's performance. It
is accomplished by unfreezing the model's layers, or
a portion of it, and retraining the entire model at a
slow learning rate. In addition to preventing
overfitting, this also improves system performance.
Because the model is fairly huge and the data set is
quite tiny, a slow learning rate is required. It leads to
overfitting, which is why the learning rate is low.
Fig. 2 After Training Model

Dataset and Model

It has been shown that the picture recognition model
Inception v3 can reach greater than 78.1% accuracy
on the ImageNet dataset.

Fig. 3 Marking Attendance

 | | | |
| +-------------------+ |
+----------------------------------------------------+
Scaling
Fig. 4 Updation of Database

IV. RESULTS AND DISCUSSION

If the bottleneck layer values are only calculated
once for each picture during pre-processing, and
then frequently retrieved from the cache during
training, this can considerably speed up the
process. This is due to the possibility that we may
read the same image more than once (if there are
no distortions applied during training).
When training with distortions like cropping,
scaling, or flipping, we are unable to use cached
bottleneck values since we must recalculate the
whole model for each image. Instead, to obtain the
bottleneck results for each, we select random
photos from the specified category and subject
them to the distortion graph and complete graph.
When training, applying easy distortions to the
pictures, such as cropping, scaling, and flips, can
assist the results. These correspond to the kinds of
variables we anticipate in the actual world and can
Scaling features a bounding box that is always in
the centre and whose size varies at random within a
predetermined range, similar to cropping. For
instance, if the bounding box is the same size as the
input and the scale% is 0, no scaling is performed.
In the event that it is 50%, the bounding box will, at
random, be either 50% of the width and height or
100% of the size.

V. CONCLUSION
The most accurate model to use is "Inception v3,"
however it is also the slowest. Choose a
MobileNet Model if you want something quicker
or smaller. For example, the command
"MobileNet 1.0 224" will choose a model that is
17 MB in size and accepts input images of 224
pixels; in contrast, the command "MobileNet 0.25
128 quantized" will choose a network that is much
less accurate but smaller and faster, is 920 KB in
size, and accepts images of 128x128 pixels.
 aid in better preparing the model to handle real- world data.
References:
Cropping
[1] S. Singh, S. Sharma, A. Jain, P. Singh, and A.
To crop an image, a bounding box is randomly placed Kudake, ‘Transfer Learning: Convolutional Neural
anywhere inside the entire image. The size of the box
Network-AlexNet Achieving Face
in relation to the input image is controlled by the
cropping option. If it's 0, the box's size is set to match Recognition’, Oct. 2022, pp. 1–4.
the input, and cropping is not done. The crop box will doi: 10.1109/asiancon55314.2022.9908650.
be half as wide and tall as the input if the value is 50%.
In a diagram, [2] it appears as follows: K. al Hajji, A. Cenadi, and F. Ahmad, ‘Machine
Learning based Human Face Recognition for
< width >
Attendance System’, in HORA 2022 - 4th
+----------------------------------------------------+
International Congress on HumanComputer
| width - crop% | Interaction, Optimization and Robotic
| < > | Applications, Proceedings, 2022. doi:

| +-------------------+ | [3] 10.1109/HORA55278.2022.9799883.

O. A. Muhi, M. Farhat, and M. Frikha, ‘Transfer
| | | | Learning for Robust Masked Face Recognition’,
| | | | in International Conference on Advanced

Technologies for Signal and Image Processing,

ATSIP 2022, 2022. doi: University Classroom’, in 2nd IEEE International
10.1109/ATSIP55956.2022.9805960. Conference on Artificial Intelligence, ICAI 2022,

[4] G. Lokku, G. H. Reddy, and M. N. G. Prasad, ‘A 2022, pp. 165–168. doi:

Robust Face Recognition model using 10.1109/ICAI55435.2022.9773650.

Deep Transfer Metric Learning built on AlexNet [7] R. Ragsania and S. Vijayalakshmi, ‘An Efficient
Convolutional Neural Network’, in Face Recognition System using Deep Transfer
ICCISc 2021 - 2021 International Conference on Learning’, in 2022 2nd International Conference
Communication, Control and Information on Advance Computing and Innovative
Sciences, Proceedings, Jun. 2021. doi: Technologies in Engineering, ICACITE 2022,
10.1109/ICCISc52257.2021.9484935. 2022, pp. 2199–2203. doi:

[5] G. Nalinipriya, R. Meenakshi, R. Mythili, B. 10.1109/ICACITE53722.2022.9823797.

Mathiyazhagi, E. Harini, and S. B. Harini,

‘A Face Recognition Security Model Using [8] K. al Hajji, A. Cenadi, and F. Ahmad, “Machine

Transfer Learning Technique’, in 8th Learning based Human Face Recognition for

International Conference on Smart Structures and Attendance System,” in HORA 2022 - 4th

Systems, ICSSS 2022, 2022. doi: International Congress on Human-Computer

10.1109/ICSSS54381.2022.9782238. Interaction, Optimization and Robotic

Applications, Proceedings, 2022. doi:
[6] H. Sultan, M. H. Zafar, S. Anwer, A. Waris, H.
10.1109/HORA55278.2022.9799883.
Ijaz, and M. Sarwar, ‘Real Time Face
[9] R. L. Priya, S. Nanda, M. Dandekar, A. Devnani,
Recognition Based Attendance System For
and R. Kadakoti, “Face-Recognition Based
Attendance System using One-Shot Learning,” in
 2021 12th International Conference on Computing
Communication and Networking Technologies,
ICCCNT 2021, 2021. doi:
10.1109/ICCCNT51525.2021.9579899.
[10] STMIK AKAKOM Yogyakarta, Institute of
Electrical and Electronics Engineers. Indonesia
Section, and Institute of Electrical and Electronics
Engineers, 2nd ISRITI 2019 proceeding : the 2nd
International Seminar on Research of Information
Technology and Intelligent Systems 2019 : “The
future & challenges of extended intelligence” :
Yogyakarta, Indonesia, 05-06 December 2019.