IRJET- Face Recognition using Landmark Estimation and Convolution Neural Network

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 08 | Aug 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 1279
Face Recognition using Landmark Estimation and Convolution Neural
Network
Neelam Patel1, Namrata Sharma2
1PG Student, Dept. of Computer Science and Engineering, Sushila Devi Bansal College of Engineering, Indore
2Asst. Professor, Dept. of Information Technology, Sushila Devi Bansal College of Engineering, Indore
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - Face recognition is a particular situation in which objectsarerecognized. Ithasgainedspecialattentionin recentyears
due to a broad range of applications such as robot-human interaction, gesture control, surveillance, security, and monitoring of
people. We tried to make it usable in real time and we used the LFW dataset to test the system, which is a set of random images
available for a person on the internet. Completely different from normal datasets collected under controlled circumstances.
Additionally, because CNN needs a lot of information, we have used a big dataset; we have accomplished an increase in both
training and testing. We checked the official LFW information page showing the outcomes of nearly all techniques and only a few
techniques crossed the 90% mark. Up to 95 percent of the maximum precision we have attained is seen, but it can be time
consuming compared to others. There is still a great deal of room for enhancement and itistobeachievedusingvariousalgorithm
permutations.
Key Words: Face Recognition, Convolutional Neural Network, Landmark Estimation, LFW Dataset, Recognition Accuracy
1. INTRODUCTION
1.1 Motivation
Our main motive for this project was our interest in implementing distinct algorithms and techniques of pre-processing to fix
important issues with facial recognition issue that has huge and meaningful uses of many domains. There are countless and
substantial applications of this studies, including facial recognition, facial disfigurementmedical diagnosisand evena person's
monitoring, surveillance and mobile apps. We were significantly influenced by the notion that we were working to solve an
open and crucial issue with all these possible apps. [1]
We specifically chose the Facial Key Point Detection Kaggle technique, which enabled us to test a broad range of methods and
neural net models to fix the location issue which is otherwise simple. Recognition of facial main points is a difficult issue given
differences in both faces and photo situations. [2] Facial characteristics vary by size, place and expression, while the
illumination and visual angle vary with the picture circumstances. In conjunction with the need for very precise coordinated
projections (e.g. the precise corner of an eye), these abundant differences lead us to think that the job is excellent to work on.
[3]
CNN and HOG detectors have been developed primarily in the form of pythons for the formation of convolutional neural
networks.[4][5] The real-time python interpreter programhasbeenusedfortestingfacedetection,overlapremoval,detection,
facial normalization and face recognition over many months, as distinct modules and as a consistent scheme. OpenCV was
developed for computer effectiveness and focused on real-time apps. A large number of structures and functions have been
used to manipulate the matrix and efficiently process images in the OpenCV library. The library ofIntel PerformancePrimitive
(IPP) has been licensed for the acceleration of the image in real time.
2. Tools Used
2.1 OpenCV
(Open Source Computer Vision Library) is released under a BSD license and hence it’s free for both academic and commercial
use. It has C++, Python and Java interfaces and supports Windows, Linux, Mac OS, iOS and Android. OpenCV was designed for
computational efficiency and with a strong focus on real-time applications. Written in optimized C/C++, the library can take
advantage of multi-core processing. Enabled with OpenCL, itcantakeadvantageofthe hardwareaccelerationoftheunderlying
heterogeneous compute platform. [6]
Adopted all around the world, OpenCV has more than 47 thousand people of user community and estimated number of
downloads exceeding 14 million. Usage ranges from interactive art, to mines inspection, stitching mapsonthe web orthrough
advanced robotics. [7]

2.2 TensorFlow
TensorFlow is an open source software libraryfornumerical computationusingdata flowgraphs.Nodesinthegraphrepresent
mathematical operations, while the graph edges representthemultidimensional data arrays(tensors)communicatedbetween
them. The flexible architecture allows you to deploy computation to one or more CPUs or GPUs in a desktop, server, or mobile
device with a single API. TensorFlow wasoriginallydeveloped byresearchersand engineers workingontheGoogleBrainTeam
within Google's Machine Intelligence research organization for the purposes of conducting machine learning and deep neural
networks research, but the system is general enough to be applicable in a wide variety of other domains as well. [8]
2.3 Python V3.7
Python is a great object-oriented, interpreted, and interactive programming language. It is often compared (favorably of
course) to Lisp, Tcl, Perl, Ruby, C#, Visual Basic, Visual Fox Pro, SchemeorJava...butit’s actuallysimpleandeasytowork on.[9]
Python combines remarkable power with very clear syntax. It has modules, classes, exceptions, very high level dynamic data
types, and dynamic typing. There are interfaces to many system calls and libraries, as well as to various windowing systems.
New built-in modules are easily written in C or C++ (or other languages, depending on the chosen implementation). Python is
also usable as an extension language for applications written in other languages that need easy-to-usescriptingorautomation
interfaces. [10]
3. Proposed System
The general architecture of any recognition system involves these basic steps:
Fig-1: General Architecture of Recognition System
3.1 Method 1: Using HOG and SVM
In our system we started with basic recognition system and used LFW dataset because it is equivalent to real time data and is
not taken in constrained condition. Most basic dimensionality reduction technique PCAisbeingusedtoreducecalculationand
operation time.
Input Dataset: LFW
Pre-Process: Eigen-Face (PCA)
Classifier: Support Vector Machine
A strong classifier SVM is being used to get higher recognition accuracy. Results are being presented in next section..
3.2 Method 2: Further Improvement using CNN
Because of the big variation between face poses, the intuition behind this is; face recognition is hard to tackle. If the CNN were
to receive the pictures as they are, it would have to cope with individuals in all ways,backgroundnoise,distinctfacial positions
in the picture, etc. Therefore, first attempting to decrease this variability by centering faces in the picture makes sense.

Therefore, the first stage would be to have all faces in the center of the picture and turn them to look at the front. Doing so
makes analyzing them much easier. In our scheme, we made use of the landmark estimation for face frontalization as; it
supplied a reliable application that worked quickly enough to meet our needs.
Fig-2: Proposed Architecture
4. Result Analysis
We have applied sklearn pipeline and preprocessing with MinMaxScaler. Training is done witha fullyconnectedlayerand100
hidden units with tensorflow backend.
Fig-3: Image from LFW dataset correctly recognized method I
The above results show that the people in dataset are predicted successfully. Few options are difficult to identify and in turn
reduces the overall accuracy. The system performance is shown in table below. The overall accuracy of the system was 85%.

Fig-4 Model Performances and loss in Detection and Recognition
We further applied augmentation; usual techniques of information increase are altering the illumination, rotating, scaling, or
image translation. We've only been interested in the first one fromthese.Theothertwo,aswearealreadylocating, centralizing
and resizing the faces, created no sense in our situation. On the contrary, even with the standardization step applied to the
feature vector, we felt it might be interesting to use modifications in illumination to train our CNN better.
The data provides keypoints as shown on the image below.
Fig-5: Keypoint Detected
Table-1: Performance in terms of different chosen Parameters
With regard to the CNN itself, we attempted to modify the setup of their layer and two dropout rates whichare0.5and0.7. We
can compare their performances. It was evident in this situation that the initial CNN setup was better conducted. We believed
we could avoid overfitting this way, given that we had less information to train, but empirical evidence showed that it did not
yield any advantage. Experiments and results hand, as they behaved differently with each CNN setup, we did not reach any
conclusion about the highest dropout rate.

Table-2: Accuracy of various methods on LFW Dataset
The overall accuracy in keypoint detection is 93.3%. CNN gives accurate results when we removed the not available (na)data.
When we applied this outputs to recognize faces with Bray Curtis distance for similarity measure the overall system accuracy
achieved was 97.4%
We have checked the LFW data official website which shows the results of almost all methods and only few methods have
crossed the mark of 90%. The maximum accuracy is seen up to 95% we have achieved better accuracy but it may be time
consuming compared to others.
To better know where we stand on the graphs on the base of performance, comparative analysis of our suggested model with
modern techniques. Comparative analysis of our proposed model withcontemporary methods,tobetterunderstandwherewe
are standing on the charts on basis of performances.
Fig-6: Performance graph of our methods with others
5. Conclusion and Future Work
5.1 Conclusion
The system acquired was widely tested and various combinations of parameters were tested. In the Wild Huang et al., 2007
dataset, we used the Large Face to evaluate the efficiency of our face verification step.
In addition, we have used large dataset because of CNN it requires lot of data; we performedaugmentationin bothtrainingand
testing. The facial recognition portion was tested using two distinct sizes of data sets and we achieved constant outcomes of
around 90% precision, achieving a maximum of 85%. These findings are better than we anticipated, allowing for some
instances of real-life use. There is space for enhancement; there is still room for improvement. There are many parameters in
our scheme related to the earlier point that we were not able to experiment as widely as we wished. Althoughweattemptedto

create trials as comprehensive as possible, owing to time limitations we were unable to do so. As such, some parameters still
exist that we do not fully comprehend their behavior.
5.2 Future Work
To be sure of the face recognition system works at complete potential, bymeansoffurthertestingweneedtodiscovertheideal
parameter setup. Finally, as has already been said, with regard to the further experiments are possible. We intend to continue
working on them so that more characteristics are provided by the internet instrument. And, on the other side, various
individuals will have to be recognized at once by the video identification portion.After that,theFRschemecanbeused inmany
other apps, so there is still a lot of job to do with this project.
REFERENCES
[1]Turk, Matthew; Pentland, Alex: Eigenfaces for recognition. In: Journal of cognitive neuroscience 3 (1991), Nr. 1, S. 71–86.
[2] Alex Krizhevsky, Geoffrey E H. Ilya Sutskever S. Ilya Sutskever:ImageNet Classification with Deep Convolutional Neural
Networks. (2012). http://dx.doi.org/{kriz,ilya,hinton}@cs.utoronto.ca
[3] Yang, M-H ; Kriegman, D J. ; Ahuja, N: Detecting faces in images: A survey. In: IEEE Transactions on pattern analysis and
machine intelligence 24 (2002), Nr. 1, S. 34–58
[4] Taigman, Yaniv ; Yang, Ming ; Ranzato, Marc’Aurelio ; Wolf, Lior:Deepface: Closing the gap to human-level performance in
face verification.In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2014, S. 1701–1708
[5] Fei-Fei Ly, Justin J. Andrej Karpathy K. Andrej Karpathy: CS231n: Convolutional Neural Networks for Visual Recognition.
urlhttp:// cs231n.stanford.edu/, 2016
[6] Alex Krizhevsky, Geoffrey E H. Ilya Sutskever S. Ilya Sutskever: ImageNet Classification with Deep Convolutional Neural
Networks. (2012). http://dx.doi.org/{kriz,ilya,hinton}@cs.utoronto.ca. – DOI kriz,ilya,hinton@cs.utoronto.ca
[7] Simonyan, Karen ; Zisserman, Andrew: Very Deep Convolutional Networks for Large-Scale Image Recognition. In: ICLR
(2015). http://dx. doi.org/{karen,az}@robots.ox.ac.uk. – DOI karen,az@robots.ox.ac.uk
[8]https://www.google.com/search?q=snapchat+features&rlz=1C1CHBD_enIN800IN800&tbm=isch&
source=iu&ictx=1&fir=W5kEw1ET6_zktM%253A%252C3nxhz23VnbhO3M%252C_&vet=1&usg=AI4_kReKxi1EHiGU_lN9kMaA
WWbQvzNrg&sa=X&ved=2ahUKEwjDsPKlvqLjAhUm6nMBHV5gCZ8Q9QEwAHoECAQQAw#imgrc=W5kEw1ET6_zktM
[9] http://www.kscst.iisc.ernet.in/spp/39_series/SPP39S/02_Exhibition_Projects/184_39S_BE_1465.pdf
[10] https://www.bayometric.com/face-recognition-applications/

IRJET- Face Recognition using Landmark Estimation and Convolution Neural Network

More Related Content

What's hot (19)

Similar to IRJET- Face Recognition using Landmark Estimation and Convolution Neural Network (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Face Recognition using Landmark Estimation and Convolution Neural Network