Frontier of Face Recognition Technology...Frontal face image Mask Sunglasses Profile By enhancing...
Transcript of Frontier of Face Recognition Technology...Frontal face image Mask Sunglasses Profile By enhancing...
Frontier of Face Recognition Technology
October 24 , 2019Hitoshi Imaoka, NEC Fellow
NEC Corporation
2 © NEC Corporation 2019
NEC Fellow
Hitoshi Imaoka
1997 Joined NEC corporation, human brain system
2002 Started face recognition research, and
commercialization of face recognition product “NeoFace”
2009~ Achieved the No.1 accuracy 5 times in face
recognition benchmark tests
2015~ Started new research areas
- medical imaging: endoscopic cancer detection
- remote gaze detection
- otoacoustic recognition
2019 the youngest NEC Fellow in company history
The “face” of face recognition
3 © NEC Corporation 2019
What is the hurdle in face recognition?
4 © NEC Corporation 2019
Accuracy
Social consensus
5 © NEC Corporation 2019
Test image
B
C
A
Question:Which of these three pictures is me?
Why is face recognition
is so difficult?
6 © NEC Corporation 2019
Test image
Different
Same
Different
Same
Different
Same
B
C
A
Even in this sample, a lot of problems include
- long term aging change
- facial view, expression, similar face etc.
However, why can people understand?
7 © NEC Corporation 2019
Why is face recognition so difficult?
Hair style
Eyebrows
Eye close and open
Wearing glassesNose has
little information
Mouth open and close, smile
Other variations• view and
illumination• aging change• facial expression
• makeup• identical twins• plastic surgery etc.
Beard
Most facial parts can be changed
8 © NEC Corporation 2019
Do you know “Bio-IDiom” ?
9 © NEC Corporation 2019
Fingerprint /palm print
Face
Oto-acoustic
Voice
Fingervein
Iris
10 © NEC Corporation 2019
NEC’s world No.1 biometric authentication
All are the results of contests by the National Institute of Standards and technology (NIST).
Finger printauthentication
FpVTE (2003,2012)SFSE (2004)
MINEX(2016,2016)ELFT (2007)
PFT/PFTII (2009,2013)
8 times
No.1
Iris recognition
IREX IXIris Exchange IX
(2018)
Also Iris recognition
No.1
FRVT (2019)
MBGC (2009)
MBE (2010)
FRVT (2013)
FIVE (2017)
face recognition
Achieved world No.1 5 times
11 © NEC Corporation 2019
Multi-modal authentication
Combination of various biometrics such as fingerprint, face, and iris.
1970 1980 1990 2000
History of NEC’s Biometric Authentication technology
• Wider range of application in our daily lives through enhancement of technology and combination of different modalities.
A pioneer with over 50 years of experience in R&D
1982: Metropolitan Police Department (Japan)
1971: Started researchfor fingerprint recognition
Early 1990’s: Criminal AFIS for states across US
Late 1990’s: Global deployment
1984: San Francisco Police Department
2008: Developed finger hybrid recognition technology
2003: Ranked No.1 in US Government benchmark testing
1989: Started researchfor face recognition
1996 Started 3D face recognition R&D
1999 Commercialization of face recognition
2003 Global deployment
2009, 2010, and 2013:Ranked No.1 in US Government benchmark testing
2017: Ranked No.1 for the 4th consecutive time in US Government benchmark testing
FingerprintRecognition
Face Recognition
2010
12 © NEC Corporation 2019
Over 1000 systems in over 70 countries globally
Systems for Immigration Control, National IDs, Entertainment, etc.
Global Deployment of NEC’s Solutions
13 © NEC Corporation 2019
Integrated platform
Expansion of market value through biometrics technology
Authentication Security
Enterprise
Mobile
City Group
Public Safety
Individual
Society
QR code payments
User authentication
when opening
accounts
Police, local governments,
shopping areas
Crime investigation
and surveillance
Apartment buildings,
hospitals
Safety and security
of residentsLarge-scale events
Sporting events
Boarding
procedures
Airports
Entire regions, including airports,
tourist sites, hotels, etc.
Regional
revitalization
through hospitality
Identity verification
when entering/
exiting
Major financial
institutions (demonstration experiments)
Hands-free
payments using
face recognition
Train stations, malls,
vending machines
Push advertising
Fast food retailers
Personalized
customer loyalty
programs
Citizen ID
Nations/governments
MarketingPayment Hospitality
Major retailers
Unmanned stores
Safe and secure
operations
14 © NEC Corporation 2019
NEC's Face Recognition Activities
15 © NEC Corporation 2019
Policies for NEC's face recognition technology development efforts
Reliability
Provide customers with reliable
technology by the world's highest
accuracy
TransparencyClarify the level of
technology through third-party evaluations
16 © NEC Corporation 2019
Advantages of NEC's face recognition algorithm
Weighting variable w
Featu
re
extra
ctio
n
Loss function L
Input x∂L(x,w)
∂w
Feature extraction
space
Verify features
MATCH
Deep learning
……
highly accurate
fast search speed
small face
- min eye distance 10 pixel
robustness of facial variations
wearing glasses, masks etc.
profile view
Original loss function and network structure
17 © NEC Corporation 2019
enrolled image
Enhanced robustness against various changes
Mask Sunglasses ProfileFrontal face image
By enhancing the recognition accuracy for partially occluded faces
ex. wearing a mask or sunglasses, or are turned to one side
Images for verification
18 © NEC Corporation 2019
Demonstration Video: Security Gate Solution
NEC's face recognition ranked No.1 in latest NIST (US) benchmark test
Patrick Grother, Mei Ngan, Kayee Hanaoka, Face Recognition Vendor Test (FRVT), NIST Interagency
Report 8271(Sep 11,2019)Https(Link)://(Link)www.nist.gov/site
Results shown from NIST do not constitute an endorsement of any particular system, product, service, or
company by NIST."
※ Subsequent face images are derived from this report.
20 © NEC Corporation 2019
The purpose is to evaluate face recognition accuracy and search speed, using large-
scale data up to 12 million people enrollment
Overview of Face Recognition Vendor Test (FRVT) 2018
What is Face Recognition Vendor Test (FRVT) 2018?
Examples of
evaluation images※
Evaluator: U.S. National Institute of Standards and Technology (NIST)
Organization established to enhance technological innovation and industrial competitiveness
Completely blind test
Evaluations are
rigorous and fair
21 © NEC Corporation 2019
Face Recognition Vendor Test (FRVT) 2018* Participants
49 Leading organizations all over the world participate
Country/regionNumber of
organizationsParticipating organizations
U.S. 11Aware, Camvi, Ever AI, Frog wing, Incode, Micro Focus, Microsoft, Noblis, Rank One, Real
Networks, Remark AI, Shaman, Vigilant Solutions
China 11Anke, Dahua, Hikvision, megvii, Newland, Sensetime, SIAT CASIA, Tong Yi Trans,
Yisheng, Yitu
Europe 9Cogent Gemalto (France), Cognitec (Germany), Dermalog (Germany), Eyedea (Czech
Republic), Idemia (France), Innovatiscs (Slovakia), NeuroTechnology (Lithuania), NFS
(U.K.), Quantasoft (Czech Republic), Visidon (Finland)
Russia 8 3DiVi, Innovation Sys., NTechLab, Smilart, Synesis, Tevian, VisionLabs, Vocord
Japan 4 Ayonix, GLORY, NEC, Toshiba
Other 6Alchera (S. Korea), Gorilla (Taiwan), HB innovation (S. Korea), Imagus (Australia),
Lookman (India), Mvision, Nodeflux (Indonesia), Tiger IT (Bangladesh)
48 companies, 1 research institution, and no universities
Participation increased since the previous NIST FIVE benchmark test (up from 16 organizations to 49)
Names of participants and evaluation results are included in the report
FRVT2018 Overall evaluation
Comparison of Authentication Accuracy
Comparison of Authentication Accuracy and Search Speed
23 © NEC Corporation 2019
Evaluation Results: Comparison of recognition accuracy
NEC is the 1st rank of identification rate, whose error rate is less than 1%.
※ False negative identification rate at a false positive identification rate of 0.1% at 1.6 million registered people. Compared only by the highest identification accuracy algorithm in each organization. Including Research Institutes in Part of the "Vendor" Labeling
Identific
atio
nerro
r rate
(%)
100.0
90.0
80.0
70.0
60.0
50.0
40.0
30.0
20.0
10.0
0.0
NEC
Vendor
1
Vendor
2
Vendor
3
Vendor
4
Vendor
5
Vendor
6
Vendor
7
Vendor
8
Vendor
9
Vendor
11
Vendor
12
Vendor
13
Vendor
10
Vendor
14
Vendor
17
Vendor
16
Vendor
18
Vendor
19
Vendor
20
Vendor
21V
endor22
Vendor
23
Vendor
24
Vendor
25
Vendor
26
Vendor
27
Vendor
28
Vendor
29
Vendor
30
Vendor
31
Vendor
32
Vendor
33
Vendor
34
Vendor
35
Vendor
37
Vendor
36
Vendor
39
Vendor
38
Vendor
40
Vendor
41
Vendor
42
Vendor
43
Vendor
44
Vendor
45
Vendor
46
Vendor
47
Vendor
48
Vendor
15
Major differences among topvendorsNEC has about 1/3 the error rate of the runner-up
Effective for applications that require high reliability such as payments
NEC
Vendor
1
Vendor
2
Vendor
3
Vendor
4
Vendor
5
Vendor
6
5.0
4.0
3.0
2.0
1.0
0.00.4
1.21.4
2.02.3
2.9
3.8 3.9 4.0
4.6
Vendor
7
Vendor
8
Vendor
9
24 © NEC Corporation 2019
Evaluation results: Comparison of recognition accuracy and search speed
Identification error rate (%)
NEC
Vendor 19
Vendor 17
Vendor 6
Vendor 28
Vendor 46
Vendor 35
Vendor 38Vendor 39
Vendor 23
Vendor 21
Vendor 44
Vendor 16Vendor 12
Vendor 22
Vendor 3
Vendor 45
Vendor 11
Vendor 47
Vendor 36
Vendor 26
Vendor 15
Vendor 31
Vendor 24
Vendor 4
Vendor 37
Vendor 9
Vendor 48
Vendor 5
Vendor 25Vendor 27
Vendor 30
Vendor 43
Vendor 7
Vendor 13
Vendor 20
Vendor 14
Vendor 10
Vendor 33
Vendor 2
Vendor 42
Vendor 18
Vendor 1
Vendor 40
Vendor 29
Vendor 34
Vendor 41
Vendor 8
Vendor 32
0.1%1.0%10.0%100.0%
Compared only among the algorithm with the highest identification accuracy in
each organization. Including Research Institutes in Part of the "Vendor" Labeling
Searc
h s
peed (m
atc
hin
gs/s
ec)
1 million
10 million
100 million
1 billionComparing recognition Accuracy and Search Speed in NIST FRVT2018
NEC's high accuracy
algorithm is 230
million matchings/sec.
※ False negative identification rate at a false positive identification rate of 0.1% at 1.6 million registered people
NEC achieved overwhelming performance in terms of both recognition accuracy and search speed.
25 © NEC Corporation 2019
Evaluation Results: Impact of aging changes on accuracy
NEC algorithm is more accurate than other vendors for long-term aging changes
- Differences increase as aging increases
- Error rate 4 times lower than the runner-up
Maintains high recognitionaccuracyover long-term changes,such as passport authentication
※ False negative identification rate at a false positive identification rate of 0.1% at 3.1 million registered people
0
5
10
15
20
25
30
35
NEC
Identific
atio
nerro
r rate
(%)
※
0-2 3-4 5-6 7-8 9-10 13-12 13-14 15-18 (year)
Vendor E
Vendor F
Vendor D
Vendor B
Vendor A
Vendor C
2.7
2.11.71.51.30.9
0.7
1.32.3
1.1
3.44.6
5.87.1
9.0
11.5
* Face Image Source: Patrick Grother, Mei Ngan, Kayee Hanaoka, Face Recognition Vendor Test (FRVT), NIST Interagency Report 8271(Sep 11, 2019)
26 © NEC Corporation 2019
Evaluation Results: Impact of the number of registered people on accuracy
NEC’s technology can be applied to large-scale systems and maintain recognition accuracy
Similar looking
The difference increases as the numberof registered users increases
Vendor E
Vendor F
Vendor G
Vendor H
NEC
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
Number of registered people (million)
Identific
atio
nerro
r rate
(%) ※
0.64 1.6 3 6 12
0.40.4 0.4 0.5
1.8
1.3
1.10.9
0.8 0.5
※ False negative identification rate at a false positive identification rate of 0.3%
As the number of registered users increases,the number of similar looking people increases,making it difficult to recognize
Identification Error Rate is just 0.5%at 12 million registered people
* Face Image Source: Patrick Grother, Mei Ngan, Kayee Hanaoka, Face Recognition Vendor Test (FRVT), NIST Interagency Report 8271(Sep 11, 2019)
27 © NEC Corporation 2019
Face Recognition Vendor Test (FRVT) 2018 Summary
With the evolution of AI technology, many companies
have begun adopting face recognition
49 organizations from around the world participate
(48 companies, 1 research organization)
Large differences for the evaluation of ageing changes
As much as 4 times difference between 1st and 2nd
Error rate for 15-18 years: NEC 2. 7%, 2nd place: 11.5%
NEC continues to be in the top position with
identification accuracy Error rate 0.5% @ 12 million registered people
NEC is the top in search speed Search speed: 230 million matchings a second
NIST comments:
NEC algorithms
continue to have the
Most Accuratefollowing 2010 and 2013
NIST IR 8271 p.11Technical Summary
Quote from
※ Comparing the most accurate algorithms for each participating organization
Comparison between Computer and Human face recognition
29 © NEC Corporation 2019
Are you up for the challenge?
Face Recognitionby AI
Face Recognitionby Human
30 © NEC Corporation 2019
Global deployments ofNEC’s biometrics
31 © NEC Corporation 2019
Faster and securer immigration control with face recognition.
JFK International Airport, USA: Entry Control system
NEC press release https://www.nec.com/en/press/201605/global_20160505_01.html
Immigration Control
Enhanced security with face recognition at Automated Passport Control kiosks at JFK International Airport in New York.
NeoFace face recognition engine.
Reduce processing time, enhance customer service and safety of airport operations.
FaceRecognition
Matching of face data in the e-passportagainst the photo of the passengercaptured at the kiosk, to ensure thelegitimate owner of the passport isentering the country.
System delivered by Unisys and inoperation since January 2016.
Challenges
Solution
Results
32 © NEC Corporation 2019
Frictionless exit control system using face recognition.
USA: US Exit Control face recognition
NEC press release https://www.nec.com/en/press/201706/global_20170627_03.html
Immigration Control
Enhanced security for exit control using face recognition.
Enable frictionless exit process.
Walkthrough face recognition system NeoFace Express
Reduce processing time, enhance customer service and safety of airport operations.
FaceRecognition
Exit control trial in several US airportsincluding Dulles International Airport.
Checking ageist passport DB (US citizens)
and entry data (visitors) to confirm theidentities of those leaving the country.
Enhance security and enable detection ofillegal overstays of foreign visitors.
Challenges
Solution
ResultsUS Customs and Border Protection
33 © NEC Corporation 2019
Preventing illegal activities at airport customs control.
Brazil: Face recognition for 14 international airports
FaceRecognition
Detection of pre-registered individuals who have been registered for suspicious activities.
NeoFace Watch real-time face recognition system
Increase efficiently and rigor of customs control at airports.
Used to enhance efficiency andeffectiveness of customs operations in14 international airports in Brazil.
Screening of passengers as they walkpast the customs control area basedon a database of pre-registeredsuspects.
Challenges
Solution
Results
Enhanced Citizen Services
NEC press release https://www.nec.com/en/press/201507/global_20150716_02.html
34 © NEC Corporation 2019
Social services for 1.3 billion with multi-modal biometrics.
India: Unique Identification Numbers Program
FaceRecognition
FingerprintRecognition
IrisRecognition
Accurate citizen data registered with the government.Equal and efficient social services for citizens.
Multi-modal biometric authentication system.
Equal services for its 1.3 billion citizens.
ID theft prevention
Multi-modal system combining face,fingerprint and iris recognition toidentify individuals.
Prevent duplicate registrations andensure all citizens have equal accessto services such as food supply,employment and tax payment.
Challenges
Solution
Results
Enhanced Citizen Services
35 © NEC Corporation 2019
Other applications in Japan
36 © NEC Corporation 2019
Identity theft prevention
Japan: Concert Ticket Holder Identity Verification
NEC Online TV: https://www.nec.com/en/global/onlinetv/en/concert.html
Supports events of every size or scale. Reduces burden on audience by eliminating the need to bring photo IDs to the venue.
NeoFace face recognition engine for identity verification.
Risk of entry with resold tickets using borrowed or fake IDs.
Ticket resell prevention and smooth admission!FaceRecognition
Matching of faces captured using tablets at admission with the pre-registered face images of the ticket holders.
Supports events of any size and scale, including those with tens and thousands of audience through fan clubs.
Reduces admission time by up to 50% compared to visual inspection by concert staff, enabling smooth entry for the audience.
Challenges
Solution
Results
Mega Event
37 © NEC Corporation 2019
Terra-cotta soldier’s face recognition (TV program project)
Sculptures of the first emperor of China’s army
Buried over 8,000 soldier sculptures
Analyzed sculpture faces using face recognition software
All of their faces are unique
38 © NEC Corporation 2019
Save the memory project in the Great East Japan Earthquake
▌Great East Japan Earthquake and Tsunami
11th March 2011
Magnitude 9.0
20,000 dead and missing people
39 © NEC Corporation 2019
“Save the Memory Project” (collaboration of Ricoh and NEC)
▌Earthquake disaster reconstruction project
Rescue team collect albums and photographs
Volunteer washed and digitized photographs
Face recognition is used to search 150,000
photographs
Return photographs to the owner
Face recognition system assisted in returning
12% of the photographs to the owner
https://www.ricoh.com/-/Media/Ricoh/Sites/com/release/2015/pdf/0309_stmp_E.pdf
40 © NEC Corporation 2019
Applied biometrics technologies
41 © NEC Corporation 2019
Remote gaze detection
Understanding interest of person with remote camera
30-50 cm
Conventional methods
Ordinary camera:Capturing high resolution image
Special devices (IR, etc.):Detecting eyes for higher precision
Our technology
Key technology
Feature points detection of eyes,
which is developed for face recognition
Need to detect accurate eyes position (center of pupils, tail of eyes, etc.)in near distance
10m away
CamerasMultiple peoplein real time
Detects gazes accurately from low resolution images in remote distances
42 © NEC Corporation 2019
Demonstration Video : Gaze Detection Technology
43 © NEC Corporation 2019
5 persons 3 persons
Conventional method Our technology
High throughput even in crowd scene High accuracy in the scene with occlusion
Deriving the number of person in each patch and summing up them
Detecting each personand counting
Learn the number of persons in each patch image generated by simulation
Crowd behavior analysis
Recognize crowd accurately with existing surveillance camera
44 © NEC Corporation 2019
Demonstration Video: Crowd behavior analysis
45 © NEC Corporation 2019
Stress Estimation
Behavior Estimation
Understandingbehavior
SecuritySecret
computation
Social acceptabilitytechnology
Multi-modal
Personal verification
Multi-modalBiometrics
To the future application
Face recognition
Access control
Immigration Control
Terminal login
Warranty of safetyEnsuring fairness
Payments and Financial Transactions
Entertainment
Walk-through
City security
Healthcare
46 © NEC Corporation 2019
Summary
Security Privacy