Khóa luận tốt nghiệp Hệ thống thông tin: Deploying a face recognition web application on Kubernetes

VIETNAM NATIONAL UNIVERSITY HO CHI MINH CITY UNIVERSITY OF INFORMATION TECHNOLOGY ADVANCED PROGRAM IN INFORMATION SYSTEMSVU NGOC HAO NGUYEN CONG PHUONG NAM GRADUATION THESIS DEPLOYING A

VIETNAM NATIONAL UNIVERSITY HO CHI MINH CITY UNIVERSITY OF INFORMATION TECHNOLOGY ADVANCED PROGRAM IN INFORMATION SYSTEMS

VU NGOC HAO NGUYEN CONG PHUONG NAM

GRADUATION THESIS

DEPLOYING A FACE RECOGNITION API ON KUBERNETES AND MOBILE APPLICATION FRAME

USING REACT-NATIVE

VIETNAM NATIONAL UNIVERSITY HO CHI MINH CITY UNIVERSITY OF INFORMATION TECHNOLOGY

ADVANCED PROGRAM IN INFORMATION SYSTEMS

ASSESSMENT COMMITTEE

The Assessment Committee is established under the Decision , date

" by Rector of the University of Information Technology.

1 - Chairman.

H1 - Secretary.

ĐẠI HỌC QUỐC GIA TP HO CHÍMINH CONG HOA XÃ HỘI CHỦ NGHĨA VIỆT NAM

TRƯỜNG ĐẠI HỌC a ra Ũ

CÔNG NGHỆ THÔNG TIN Độc Lập - Tự Do - Hạnh Phúc

TP HCM, ngay thang ndm

NHAN XET KHOA LUAN TOT NGHIEP

CUA CAN BO HUONG DAN

Tên khóa luân:

DEPLOYING A FACE RECOGNITION API ON KUBERNETES AND MOBILE

APPLICATION FRAME USING REACT-NATIVE

Nhóm SV thực hiện: Cán bộ hướng dẫn:

Vii Ngọc Hào - 17521299 TS Đỗ Trọng Hợp

Nguyễn Công Phương Nam - 17520778

Đánh giá Khóa luận

1 Vé cuôn báo cáo:

Số bảng số liệu: - Số hình vẽ:

Số tài liệu tha Sản phẩm:

Một số nhận xét về hình thức cuốn báo cáo:

4 Về thái độ làm việc của sinh viên:

Danh gid Chung? 0

Điểm từng sinh viên:

ĐẠI HỌC QUÓC GIA TP HÒ CHÍMINH CỘNG HÒA XÃ HỘI CHỦ NGHĨA VIỆT NAM

TRƯỜNG ĐẠI HỌC Ty „

CÔNG NGHỆ THÔNG TIN Độc Lập - Tự Do - Hạnh Phúc

TP HCM, ngày tháng năm

NHẬN XÉT KHÓA LUẬN TÓT NGHIỆP

CUA CÁN BO PHAN BIEN

DEPLOYING A FACE RECOGNITION API ON KUBERNETES AND MOBILE

APPLICATION FRAME USING REACT-NATIVE Nhóm SV thực hiện: Cán bộ phản biện:

Vũ Ngọc Hào - 17521299 TS Lê Kim Hùng

Nguyễn Công Phương Nam - 17520778

Đánh giá Khóa luận

1 Vé cuôn báo cáo:

SỐ trang: ccccecccccrrrrree Số chương:

-Số bảng số liệu: - -Số hình

vẽ: Số tài liệu tham khảo Sản phẩm: -+

3 Về chương trình ứng dụng:

4 Về thái độ làm việc của sinh viên:

Danh gid Chung? 0

Điểm từng sinh viên:

Thank You

The graduation thesis is the result of the author's work, research and study during his

time in the lecture hall of the University of Information Technology The author could not

have completed the graduation thesis well without the care and dedicated help of teachers, friends and family.

First of all, the team would like to express our sincere thanks to the staff of the

University of Information Technology - Vietnam National University, Ho Chi Minh City.

Ho Chi Minh and the Faculty of Information Systems have created a favorable environment and conditions, helping the author to have the basic knowledge as a basis to carry out the thesis.

In particular, the author would like to express deep gratitude and deepest gratitude to the instructor — Ph.D Do Trong Hop He directly guided, corrected and contributed many valuable comments to help the author successfully complete his graduation thesis.

Besides, the author would like to thank his family, relatives and friends of the same course

for encouraging, encouraging and supporting the author mentally during the completion of the graduation thesis.

During one semester of project implementation, the author has applied the accumulated background knowledge, combined with learning and researching new knowledge to complete a project report best way However, in the implementation process, the author cannot avoid shortcomings Therefore, the author is looking forward to receiving

suggestions from the teachers in order to improve the knowledge that the author has studied

and to be a preparation for the author to carry out other topics in the future.

Thank you very much, teachers!

Ho Chi Minh City, February 2023

Vu Ngoc Hao

UNIVERSITY OF INFORMATION TECHNOLOGY

THESIS TITLE: DEPLOYING A FACE RECOGNITION API ON KUBERNETES AND

MOBILE APPLICATION FRAME USING REACT-NATIVE.

Advisor: Ph.D D6 Trong Hop

- Building a frame of mobile application face recognition based on a large user face registration.

- Basic functionality of the application:

o Provide a demo application for the face recognition.

©_ Provide 2 options:

® Mobile application for demo.

e Register and recognition api.

Scope:

Based on Face Detection, Face Recognition, Minio for high performance search and saved image Based on frameworks Tensorflow, React Native, Fastapi, py-torch, torch-lighting.

- Selecting the suitable machine learning model.

© Goal: Selecting the right model that has fast convergence rates and light-weight.

o Content: Based on Triplet Loss, Arcface, Cosface for evaluating and selecting models.

co Expected results: Model selected and weights saved of the model.

- Selecting the inputs is crucial.

© Goal: Aim to high convergence for the model.

© Content: Based on test train, famous face recognition paper.

co Expected results: model extracts facial features and encodes it into vector.

- Al-Backend server will be deployed, application will be built on physical mobile device.

° Goal: Deploy AI - Backend server as a restful-api server, and application can run on a

physical mobile device.

o Content:

® Design AI- Backend server and deploy using Docker.

® Design Mobile applications that interact with AI-Backend servers, register,

recognition face api

co Expected results: Successfully building a mobile application for demonstration.

- Result:

© Provide an implementable apis and tracking mobile application frame.

Methodology:

- Parallel research and implementation.

- Distributed system design.

Result:

- Building a successful mobile application on a physical mobile device for demonstration.

- This system can deal with a large face register and high availability.

Research timelines:

- Research about Tensorflow modeling, Fastapi, py-torch, torch-lighting and integrate them

together (2 weeks).

- Research about flutter, building an example react native application (1,5 weeks).

- Selecting machine learning model (1,5 weeks).

Research training input types for suggestion models, developing models for training (3 weeks) Training in data enrichment from Ifw (3 weeks).

Deploy backend server using docker (1 weeks).

Prepare a document for the thesis report (4 week).

References Document:

o Machine learning model:

* face recognition: https://arxiv.org/abs/2103.14983

* triplet loss for face embedding: https://arxiv.org/abs/1503.03832

Approved by the advisor

Face detection Face recognition Minlo storage RESTfull API

Nginx proxy manager:

Research about React Native

Research about Ngrok Research about Expo

Research about Expo Camera

Research about Expo FaceDetector

Research about Expo ImageManipulator Research about Canvas

Overview 3.1.2 Workflow

3.3.1 Overview

3.3.2 Workflow

3.4.1 OV€TVÏ€W HH2 HH HH HH 1 1 11g 3.4.2 WorkflowW co th HH H111 re Chapter 4 IMPLEMENTATION 2t 2t 2t 2t tt thi 4.1 Face recognition API cú HH” HH HH gi 4.2 DepÏOyI€IIK - 1t HH” 22121121 1.121 HH gà ngư 4.3 L0ggÏng: tt 1 1 HH 1 01 1101011 1e 4.4 Swagger

4.4.1 APT face reC0gTnÏfÏOI nh ngHnH ng Ho TH ng tr 4.4.2 Face searching:

4.4.3 Getting all faC€s cọ HH1 1g 4.4.4 Deletting a FaCe: co co nh HH 1 re 4.4.5 Deleting a S€T- St th HH1 re 4.5 Minio console UI for manager S1 3t >Ev tr g rngg gtrrư 4.6 User ÍnferÍaCe 22t 2t t2 221122 2211211211111211 11 1 11111 1 1.10111011011011 ke 4.6.1 Structure

4.6.2 Demo fun€fiOIS - ¿5 522 2 2E 3E t2 211 1111.21.11.11 111.11 Chapter 5 CONCLUSION AND FUTURE VISION ¿5-55-5222 St E2 2E 1E 212210112111 11.111 11g tri 5,1 AdVanAg6 nh HH HH HT TH TH HH1 HT HH0 T1 T111 11k rkrk 5.2 Disadvanfage - cách 0H 101.11 01001011 11100 Š.3 Future YisÏ0n tt tinh 11 1 H01 101.111

REFERENCES

List of figures

2.1 Workflow FAISS library work

2.2 The components of a Kubernetes cluster

2.3 Dataset sample English

2.4 Dataset sample challenge

2.5 Face checking process

2.6 Recognition process

2.7 Algorithm comparison image

2.8 Minlo provides storage space

2.9 Minlo architecture

2.10 Network architecture

1 Example Nginx proxy manager UI usage

2 Example nginx gateway status

3 Example nginx route config

Overview system architecture

3.2 Restful-API-core overview

3.3 Connection checking

3.4 Register workflow

3.5 Recognition workflow

3.6 Face tracking general workflow

3.7 Face processing detail workflow

3.8 Face processing detail workflow

3.9 Apply nginx in Kubernetes for management

4.1 Build process

4.2 Build process (cont)

4.3 API logging after start service manually

20

21 22

23 24

25 26

27

27 29

31

31

35

38 39 40

42

44

46 47

49

50 50

4.4 Swagger UI

4.5 Test register url

4.6 Test Search url

4.7 Test get all registered faces API

4.8 Test delete a registered face

4.9 Test delete a registered user

4.10 Minio manager storage

4.11 Minio manager storage (cont’s)

4.12 Function handle crops face

4.13 Function handle face identify

4.14 The main page return result

Chapter 1 PROBLEM STATEMENT

1.1 Overview

Nowadays, information technology is developing day by day The application of

information technology in business is very popular Today's businesses are too familiar with

the application of information technology to their activities from simple or complex jobs

The world of technology is still changing every day and is increasingly simplifying

everything, finding the fastest solution to any problem To keep up with the times, we must

constantly adapt to the world and develop ourselves The more businesses grow, the more

employees there are, gradually making the job of managing employees more and more

difficult To ensure personal interests and to control more easily Therefore, it is very

important to build an attendance program to suit the situation.

Current timekeepers mainly use fingerprints, or some places use face recognition

fixed in one place Using these technologies still has certain limitations

Therefore, we chose the research topic of face recognition that can be used on the

phone to replace fixed attendance in one place

“Deploying a face recognition api on Kubernetes and Mobile application frame

The reason this thesis does not use available libraries for processing for facerecognition is because of limitations in personalization, management and system

implementation:

© Cannot separate the database and the system API separately, difficult to

expand the system It also reduces the availability of the whole system

© It is not possible to increase or decrease local resources when a small part of

the system needs excessive resources

© Can't integrate faiss to find use gpu to speed up when data becomes big data.

1.2.2 Scope

With this graduate thesis, the main goal that I aim for when doing this is as follows:

o Provide an implementable apis and tracking mobile application frame

o Based on face detection, Face recognition, Minio for high performance search

and saved image

o Based on frameworks Tensorflow, react native, fastapi, py-torch,

torch-lighting

o Deployment environment: Docker-compose or Docker-swarm or K8s.

1.2.3 Objective

- Selecting the suitable machine learning model

o Goal: Selecting the right model that has fast convergence rates and

light-weight

o Content: Based on triplet loss, arcface, cosface for evaluating and selecting

models

o Expected results: Model selection and weights saved of the model

- Selecting the inputs is crucial

o Goal: Aim to high convergence for the model

o Content: Based on test train, famous face recognition paper

o Expected results: model extracts facial features and encodes it into vector

- AI-Backend server will be deployed, application will be built on physical mobile

device

o Goal: Deploy AT - Backend server as a restful-API server, and application can

run on a physical mobile device

o Content:

® Design AI- Backend server and deploy using Docker

e Design Mobile applications that interact with AI-Backend servers,

register, recognition face API

o Expected results: Successfully building a mobile application for

Chapter 2: Literature review and theoretical background

Chapter 3: Model architecture

02/11/2022 - 23/11/2022 Training data enrichment from Hào

lfw

24/11/2022 — 01/12/2022 Deploy backend sersver using Nam

docker

02/12/2022 — 28/12/2022 Prepare a document for the Hao

thesis report Nam

Chapter 2 LITERATURE REVIEW AND THEORETICAL BACKGROUND

2.1 FAISS

FAISS stands for Facebook AI Similarity Search [1] This is a library that makes it

easy for developers to find extended versions of various informational content It addresses

the limitations of conventional query search methods designed with hash-based search in

mind and allows many other search engines to expand further

With FAISS, developers will search this content set in a variety of methods that are

not useful or impossible to use with standard database (SQL) engines alone It involves

implementing a search around closest to database pages in the millions to several billions in

size to ensure a balance of storage performance and accuracy FAISS aims to ensure high

performance for almost any important content, and these sites are also a big plus when your

access is slow or expensive

In this article, we will choose a FAISS-CPU library in python to index data and find

the latest data In addition, we also developed the FAISS interface for ease of working with

We are using the FAISS library for N-dimensional vector data and storing it in ram

to simplify the search

2.2 Kubernetes

Kubernetes is a free, open source platform that allows extensive control over the

amount of products and data available in the storage area, including management and

automation functions [2] It has a large and fast-growing ecosystem Kubernetes services,

applications and tools are developed for free

The name Kubernetes comes from the Greek word, meaning rider or driver K8 is another

name for the eight letter calculation between the letter "K" and the letter "s” Google made

the official source code of the Kubernetes project available in 2014 Kubernetes combines

over 15 years of Google's experience in running this workload on a global scale with the

best ideas and methods from society

A dataset is a collection of data In other words, a data set fits the theme of a database

table or large dataset, where each of its columns represents a particular variable and rows

equate to a certain member of the dataset in question In Machine Learning projects, we

need a training dataset This is a dataset that can be used to train and practice a variety of

activities Data sets are classified into three types that can be used for machine learning:

training set, validation set, and test set

Currently on the internet there are quite a few datasets showing how to recognize

Fake News Covid, but is using such datasets really useful and effective? Therefore, using

the dataset for actual training is one of the biggest challenges worth considering

The dataset collected and used by their team

- https://www.kaggle.com/competitions/face-recognition2/data

= kaggle a Kế xử

4 Create

@ Home

© Comet | Face recognition

Use non-parametric density estimation

+ trainJaffe.csv - the training set

+ testJaffe.csv - the test set

+ sample.Jaffe.csv - a sample submission file in the correct format + trainJaffelabels,csv - file containing labels

Figure 2.3 Dataset sample English

- https://www.kaggle.com/datasets/atulanandjha/Ifwpeople

= kaggle Q search son GED

LFW - People (Face Recognition)

Create

† l a boo! +DataCard Code (26) Discussion (2) hàm) Newt Notetcor + Rownicad (24a) Tae

@ Home Welcome to Labeled Faces in the Wild, a database of face photographs designed for studying the problem of unconstrained License

face recognition The data set contains more than 13,000 images of faces collected from the web Each face has been labeled GNU Lesser General Public Lice.

© Competitions with the name of the person pictured 1680 of the people pictured have two or more distinct photos in the data set The only

constraint on these faces is that they were detected by the Viola-Jones face detector Expected update frequency

Keyword-based artificial intelligence (AI) computing technology (commonly known as

face recognition) is used to search for and recognize human faces in digital images Facial

recognition technology can be used in various aspects: from medical to biometrics, law

enforcement, entertainment and personal security to enhance control and surveillance of

people in the community real time

Face recognition has evolved from basic computer vision techniques to the advancement

of machine learning (ML) to highly sophisticated artificial neural networks (ANNs) and

many other technologies, and as a result, improved significant productivity improvement It

now plays an important role as the forerunner of other technologies - including face

monitoring, face analysis and facial recognition Facial recognition is closely related to how

this will work in the future

In face tracking, facial recognition that identifies the color of an image or video can be

added to determine age, gender and mood using facial expressions In a facial recognition

system that builds a map of each individual's facial features and stores them as faces, facial

recognition data is the basis that allows algorithms to determine the degree of Images or

videos are suitable for face modeling Once identified, the new fingerprint can be compared

with the stored fingerprint to determine if there is a match

In this paper, we are using the face recognition database in python to determine if the

face retrieved from the user interface is suitable or not

incomming face image

2.5 Face recognition

Facial recognition is a way of identifying or confirming an individual's identity

through the use of their face Facial recognition systems can be used to identify things in

images, videos, or in real time

Facial recognition is a type of biometric authentication Various forms of biometrictechnology also include voice recognition, internal fingerprint recognition, and iris or eyelid

recognition The technology is mostly used for security and legal compliance, but there are

also people who focus on some private uses

pecs)

v

incomming face image

v

2S»

“cheœk

number of exist face

The figure above is how the register for new faces works in the backend.

There are many algorithms to apply to face recognition such as: ArcFace [3],

ArcOrthFace, ArcCentFace, MagFace, In this article , I am using the ArcFace algorithm

to apply to my face recognition system because based on this research

(https://www.researchgate.net/publication/363062634 Additive Orthant Loss for Deep_

Face_Recognition ) you can see the best accuracy is ArcFace

Loss Function Rank1 Accuracy Verification Accuracy

MinIO 1s an advanced, high-performance object storage system [4] It is built by

operating system, based on industry standards, and is 100% open source under the license

GNU AGPL v3

MinIO is different in that it was built from the ground up to the standards of

integrated/distributed cloud object storage Because MinIO is purposefully developed to

support every object in a single layer architecture achieving most of the essential features

without being involved The result is a single cloud-native object server that is fast, open,

and lightweight

While MinIO excels in many areas such as legacy object storage for secondary

storage, error correction, and caching, MinIO is unique in experiencing the effects of big

data workloads across the globe cloud, machine learning, and processing

My Object

Data My Bucket

Figure 2.9 Minio architecture

2.7 RESTfull API

A RESTful API is an architectural style for application program interfaces (APIs)

that use HTTP requests to exploit data usage The data api can be used for GET, PUT, POST

and DELETE many types of data, for the analysis, synthesis, compilation and deletion of

necessary information about the environment

The simplest way to use the REST API definition is to first know what an API is API

is an acronym for application programming language that is often used to link applications

together to perform a function built on the exchange of data and perform predetermined

tasks In essence, it allows one piece of software to interact with another The REST API

is one of the two most widely implemented forms of API, and the other is called the Simple

Objects Support (SOAP) API

REST is an enterprise support structure based on a question/answer model The REST

API has become familiar as part of the Web Service implementation approach Developers

using RESTful APIs can send emails and receive replies over many HTTP functions

2.8 Nginx proxy manager:

Authorization is also used in distributing data between several servers to seamlessly

share content from multiple websites with each other or to send processing requests to

application servers through methods outside of HTTP

Nginx is an open source web server with many features including reverse proxy,

caching, content allocation, data delivery, and more It starts with a web server built for

optimal performance and security With many HTTP server features, NGINX has also acted

as a proxy server for email (IMAP, POP3 and SMTP) as well as a reverse proxy and load

balancer on HTTP/2, TCP and UDP protocols.

Nginx uses an event-driven connection that allows multiple requests to be handled

without synchronization It is developed to use asynchronous event-driven connection

handling technology Therefore, the project process allows handling thousands of

connections (requests) through 1 thread These connection handling modules allow Nginx

to work extremely quickly and efficiently with small amounts of data In addition, you also

use Nginx which allows to handle more than 10,000 parallel connections with low resources

(CPU & Memory) under heavy requests

The Nginx reverse proxy acts as an intermediary server that processes your requests

and passes the data to the appropriate upstream supporting server and thereby sends the

information from the server to the system Reverse proxies provide the same features as an

abstraction layer above upstream servers

In this article, we are using 2 types of proxies The picture below depicts the

Internal service Internal service Internal service Internal service Internal service Internal service

Figure 2.10 Network architecture

Proxy types:

Nginx proxy manager: responsibility for forward traffic for each cluster in local

network Nginx proxy manager is a proxy with UI configuration and can gen cert using

certbot inside and cert is signed by letsencrypt for free

Ộ Nginx Proxy Manager Baristi000

Proxy Hosts luanvan|

Let's a

kubesphere-luanvan.hem ud.a torage-luanvan.hcm.u a lân m.unicloud.ai http://10.30.2.54:31080 Public

e Encrypt Online

Figure 2.11 Example Nginx proxy manager UI usage

Nginx cluster gateway is responsibility for routing traffic to each service inside k8s

cluster Nginx cluster gateway is a k8s gateway and configurable using gateway and route

Each route can be configured via k8s annotation

© Getting Started Getting Started ©) How can | use dock Administering secur Umee / Umee Crm- ©) Ubuntu 20.04 LTS (

© Piattorm Workbench 'É% KUBESPHERE 3 xám ~

Figure 2.12 Example nginx gateway status

Baie Routing Rules

+ Set Routing Rules

2.9 Research about React Native

React Native is a JavaScript framework that can build realistic mobile apps that are

optimally supported on iOS and Android [5] It is based on React and Facebook's JavaScript

library to improve the consumer experience, but instead of focusing on apps, it focuses on

certain android devices

Advantages of React Native:

- Save time when you need fast application development

- Works quite efficiently

- Stable development community

- Save cash.

- Small development team

- Reliable and sustainable application

- Build on different platforms with the least amount of native code

Risks and reductions:

- No native code required.

- Performance will be low compared to apps with native code

- The safety is not high because it is based on JavaScript

- Data management.

- Customization also doesn't really work with some modules

2.10 Research about Ngrok

Ngrok is a tunnel building tool that connects your local computer and the internet [6] It

helps some people on the internet connect to your local computer via Ngrok’'s custom

domain