2. Table of Contents
Table of Contents 2
1. Introduction 3
2. Overview 3
3. Project Description 4
3.1 CoWIN System WorkFlow 4
3.2 What is Amazon Text Rekognition 5
3.3 Why Amazon Text Rekognition 5
3.4 Known Limitations 6
3.5 Key Technologies 6
3.6 Interoperability 6
3. 7 Amazon Rekognition Pricing 7
3.8 Platform Dependence 7
3.9 WorkFlow Diagram 7
3.10 Technical Implementation 8
3.10.1 Proof of Concept 8
3.10.2 Scalability 8
4. Problem Statement 2 9
4.1 What is Amazon Image Rekognition 10
4.2 Key Features 10
4.3 Known Limitations 11
4.4 Key Technologies 11
4.5 Interoperability 11
4.6 Platform Dependence 11
4.7 Data Flow Diagram 12
4.9 Proof of concept 14
4.10 API Response 15
5. Problem Statement 3 15
5.1 CoWin ‘How to’ 16
5.2 Virtual Training via BigBlueButton 17
5.2.1 Features 17
5.3 Introduction to COWIN ( Introduction wizard ) 19
5.4 CoWin Learning Resources 19
5.5 Cowin Assessment 20
5.6 Assessment Result 24
5.7 Support (Chat System) 25
6. Contact 25
3. 1. Introduction
Coronavirus pandemic has impacted more than 90M people all over the world, India has
reportedaround~10Mcasessofar.India hasalreadycomeupwithfollowing vaccinesafter
the trial -
1. CoviShield[1]
2. Covaxin[1]
Indian Electronic Vaccine Intelligence Network (eVIN) system is ready to plan, manage and
distribute vaccines to all states around India. The proposal is the full fledged plan.
2. Overview
This CoWIN system will be a subset of COVID India Portal which provides end to end
management of COVID19.
Below are the problem statements we plan to work on -
1. How to achieve High Adherence rate from front line Health workers (ANM, ASHA,
AWW and any other health workers deployed for line listing) for CoWIN
application
2. How to ensure portability across the country to account for travel/migration
between vaccination sessions and across geography.
3. How to have a comprehensive & innovative Learning Management System and
dynamic learning support for COVID management
[1] https://vaccine.icmr.org.in/covid-19-vaccine
4. 3. Project Description
3.1 CoWIN System WorkFlow
CoWINconsists ofseveralmodulesasanAPIinterfaceallowing eachother tocommunicate
with each other.
Problem Statement 1 - How to achieve High Adherence rate from front line Health workers
(ANM,ASHA,AWWandanyotherhealthworkersdeployedforlinelisting)forCoWIN
application
Solution - We plan to use Amazon Rekognition API[2]. Using the API, text can be easily
extracted from the physical forms and then can be stored on the database using User
Management APIprotected viaanAPI TokenAuthentication. We compared the inhouse
solutionbuiltwithTensorflow and OpenCVand the Rekognition APIseems to bescalable
really well and has much better response time ( proof of concept below )
[1] https://aws.amazon.com/rekognition/
5. Figure 3.1: Flow of Front line worker like ASHA, ANM, ASHA, AWW for line listing
3.2 What is Amazon Text Rekognition
Amazon Rekognition makes iteasy to add image and video analysis to your applications. We can just
provide animageorvideotothe AmazonRekognitionAPI,andthe servicecanidentifyobjects,people,
text, scenes,and activities. AmazonRekognition textdetectioncandetecttextinimages. Itcanthen
convert the detected text into machine-readable text.
DetectTextdetectstextin.jpegor.pngformatimages.BothimageandvideotextdetectionAPIssupport
mostfonts,includinghighlystylizedones.Afterdetectingtext,AmazonRekognitioncreatesa
representation of detected words and lines of text,shows the relationship between them, and tells you
where the text is on an image or frame of video.
3.3 Why Amazon Text Rekognition
1. Simple integration – Amazon Rekognition makes it easy to add image recognition
capabilities into your applications with a simple API.
2. Deep learning technology – Amazon Rekognition employs deep learning technology
to analyze images and videos. It’s continually trained on new data to improve and
expand its ability to accurately identify objects, scenes, and activities.
3. Scalable image analysis – With Rekognition, you can analyze millions of images.
6. 4. Integration with other AWS services – Rekognition works seamlessly with other
AWS services, such as AWS Lambda and Amazon S3.
5. Low cost – The service does not have minimumfees or upfront commitments. You’re
only charged for what you use.
6. Fully managed – With Amazon Recognition, you getconsistent response times
irrespective of the volume of requests you make.
7. Real-time analysis – Amazon Rekognition allows you torun real-time analysis on videos and
images.
3.4 Known Limitations
● Maximum image size stored as an Amazon S3 object is limited to 15 MB.
3.5 Key Technologies
● Amazon Rekognition
● Amazon Lambda
● Python
● HTML/CSS
3.6 Interoperability
The solutioniscompletelyinteroperableandcanbeusedwithanyother techstack,
Considering the Cowin existing tech stack is built in Node.js
The solutionwould be anAPIcommunicating withfrontend as astandalone module and can
be integrated independently
7. 3. 7 Amazon Rekognition Pricing
Service Cost Comments
Amazon S3 $0.08 3MB per pic x 1K images x $0.023/ GB - month
Amazon Rekognition $6.01 5API(withimages)processedperiterationx1.2Kiterations=$61K
facemetadatastoredx$0.01per1000facemetadatastored=$0.01
AWS Lambda $0 20invocationsperimagex1Kimages=(~25,200GBseconds)-free
tier eligible
Table 3.7 - Pricing per 1k transactions
3.8 Platform Dependence
The solution works on any operating system without any dependency and scales really well.
3.9 WorkFlow Diagram
8. 3.10 Technical Implementation
Objective - Ability to auto filldatafrom provided documents from the end users ( i.e Aadhar
Card ) withminimalerrors and efforts from nontech savvyfront line workers.
TheusermanagementserviceAPIwillprocesstheformimagessentviatheCOWIN
applicationand thenwould besentover toAmazonRekognitionAPIextracting allthetextwe
would like to store and map to the current schema.
3.10.1 Proof of Concept
import boto3 // import the boto3 package
d
ef detect_text(photo, bucket):
client=boto3.client('rekognition') // register the 'rekognition' API instance
response=client.detect_text(Image={'S3Object':{'Bucket':bucket,'Name':photo}})
textDetections = response['TextDetections']
print ('Detected textn ------- ')
for text in textDetections:
print ('Detected text:' + text['DetectedText'])
print ('Confidence: ' + "{:.2f}".format(text['Confidence']) + "%")
print ('Parent Id:{}'.format(text['ParentId']))
return len(textDetections)
3.10.2 Scalability
Amazon Rekognition is a highly scalable solution.
1. Billions of documents or images can be indexed within a couple of days.
2. Serverless Architecture via AWS lambda[3].
3. Identifying text response within seconds.
[2] https://aws.amazon.com/lambda/
if 'ParentId' in text:
print()
print ('Id: {}'.format(text['Id']))
print ('Type:' + text['Type'])
9. 4. Problem Statement 2
How to ensure portability across the country to account for travel/migration between
vaccination sessions and across geography.
Solution - We plan to capture the faces of the vaccinators and then match them against
existingrecordstomakesuretheyaren’tvaccinatedbefore.WeplantouseAmazon
Rekognition. Amazon Rekognition is a cloud-based Software as a service (SaaS) computer
vision platform that was launched in 2016.
Amazon Rekognition makes iteasy toadd image and video analysis to your applications
using proven, highly scalable, deep learning technology that requires no machine learning
expertise to use.
Considering its acontactless method we decided tochoose itover anyother biometric
de-duplication method.
Figure 4: Process flow diagram
10. 4.1 What is Amazon Image Rekognition
Amazon Rekognition Image is a deep learning powered image recognition service that
detects objects, scenes, and faces; extracts text; recognizes celebrities; and identifies
inappropriatecontentinimages.Italsoallowsyoutosearchandcomparefaces.Rekognition
Image isbasedonthesame proven, highlyscalable,deep learning technologydeveloped by
Amazon’scomputer visionscientiststoanalyze billionsofimages dailyforPrime Photos. The
servicereturns aconfidencescoreforeverything itidentifiessothatyoucanmake informed
decisions about how you want touse the results. Inaddition, alldetectedfaces arereturned
withboundingboxcoordinates,whichisarectangularframethatfullyencompassestheface
that can be used to locate the position of the face in the image.
4.2 Key Features
4.2.1 OBJECT AND SCENE DETECTION
RekognitionImageidentifiesthousandsofobjectssuchasvehicles,pets,andfurniture.
Rekognition also detects scenes within an image, such as a sunset or beach. This makes it easy
for you to search, filter, and curate large image libraries.
4.2.2 FACIAL RECOGNITION
Rekognition Image enables you to find similar faces in a large collection of images. You can
createanindexoffacesdetected inyourimages. RekognitionImage’sfastandaccuratesearch
returns faces that best match your reference face.
4.2.3 FACIAL ANALYSIS
WithRekognitionImage,youcanlocatefaceswithinimagesandanalyzefaceattributes,suchas
whether or not the face is smiling or the eyes are open. When analyzing animage, Rekognition
Image will return the position and arectangular frame for each detected face.
11. 4.2.4 FACE COMPARISON
RekognitionImageletsyou measurethelikelihood thatfacesin two imagesareof thesame
person.WithRekognition,youcanusethesimilarityscoretoverifyauseragainstareference
photo in near real time
4.3 Known Limitations
● The Maximum number of faces you can store in asingle face collection is 20 million.
But there can be multiple collections you can look-up into.
4.4 Key Technologies
● Amazon Image Rekognition
● Amazon Lambda
● Python
● HTML/CSS
4.5 Interoperability
The solutioniscompletelyinteroperableandcanbeusedwithanyother techstack,
Considering the Cowin existing tech stack is built in Node.js
The solutionwould be anAPIcommunicating withfrontend as astandalone module and can
be integrated independently
4.6 Platform Dependence
The solution works on any operating system without any dependency and scales really well.
13. 1. Using a contactless camera a photo will be clicked of the person willing to take the
vaccine.
2. The photo then would be passed to the vaccine API to tell us whether the user has
been already vaccinated or not using the facededuplication module
Figure 4.8.2: face search API using AWS facial rekognition
14. 4.9 Proof of concept
We ran a POC ( proof of concept ) to index and analyze 20,000 images and it worked really
well. We were able to compare images within ~0.5ms of response time.
Source code
import boto3
BUCKET = "amazon-rekognition"
KEY = "search.jpg"
COLLECTION = "my-collection-id"
def search_faces_by_image(bucket, key, collection_id,
threshold=80, region="eu-west-1"):
rekognition = boto3.client("rekognition", region)
response = rekognition.search_faces_by_image(
Image={
"S3Object": {
"Bucket": bucket,
"Name": key,
}
},
CollectionId=collection_id,
FaceMatchThreshold=threshold,
)
return response['FaceMatches']
for record in search_faces_by_image(BUCKET, KEY, COLLECTION):
face = record['Face']
print "Matched Face ({}%)".format(record['Similarity'])
print " FaceId : {}".format(face['FaceId'])
print " ImageId : {}".format(face['ExternalImageId'])
15. """
Output:
Matched Face (96.6647949219%)
FaceId : dc090f86-48a4-5f09-905f-44e97fb1d455
ImageId : test.jpg
"""
4.10API Response
Matched With 96.99021911621094% Similarity
To FaceId :32276a6d-f838-558a-bc1b-6f6d6e8b79cf
Which is ImageId : 51b4f021-b8ab-5945-95ed-1c6c02db5b54
With 99.99979400634766 Confidence
5. Problem Statement 3
How to have a comprehensive & innovative Learning Management System and dynamic
learning support for COVID management
Solution -Weplantointroduce3modulestomakesureCOWINfrontlineworkerscango
throughlearning‘Howto’andcantakeassessmentstoensuretheyunderstandwhattheyare
doing.
We are suggesting having BigBlueButton[4] as an interactive training platform to allow
frontend line workers to go through training from anywhere.
16. Also we plan tohave a two way chatallowing them to askquestions whenever they want
directly from the Cowin application.
5.1CoWin ‘How to’
The application will have a CoWIN training section in the navigation allowing them to go
through an introduction wizard anytime. Introduction wizard will take them through every step
they need to follow while the vaccination process.
Figure 5.1 : COWIN application training menu
17. 5.2Virtual Training via BigBlueButton
5.2.1 Features
1. Interactive training sessions with Multi-user whiteboard
BigBlueButton allows you to host the virtual training session with multi user whiteboard
allowing any participant to write down or interact with the whiteboard.
18. 2. Open source self hosted solution
BigBlueButton is an open source solution. The data including participants, documents and
the communication is done through in-house private servers keeping our 100% safe & secure
compared to commercial solutions like zoom.
3. Public / Private Chat
BigBlueButtongives youanabilitytoallowparticipants tochatwitheachother orthehost.
Theycanaskquestions.Notescanbeexchangedbythepresentatortothetrainees.
4. Polls
The hostcanaskquestionsand participantscanreplytothemusing thepollingfeature.
Making the training sessions more interactive.
5. Breakout Rooms
BreakoutRooms willallowhosts togroup usersintodifferentroomsfortheirrolespecific
content training and demonstration.
6. Mobile Support
The platform works on any device, allowing participants to join from any device without any
mobile application, training sessions can be joined directly from their browsers via a unique
meeting URL.
19. 5.3 Introduction to COWIN ( Introduction wizard )
Figure 5.3 : Introduction wizard
5.4 CoWin Learning Resources
The CoWIN learning module will consist of several learning materials including images, videos
and documents allowing them to read whenever they want straight from the CoWIN app.
20. Figure 5.4 : COWIN Learning resources
5.5 Cowin Assessment
Workers can take assessments for better understanding about application.
1) MCQ Quizzes