deep learning facial r ecognition system - panasonic...2019/07/16 · a facial recognition system...
TRANSCRIPT
WHITE PAPER
Deep Learning Facial R
Panasonic Video Surveillance S
WHITE PAPER
Deep Learning Facial R
Panasonic Video Surveillance S
WHITE PAPER
Deep Learning Facial R
Panasonic Video Surveillance S
Deep Learning Facial R
Panasonic Video Surveillance S
Deep Learning Facial Recognition System
Panasonic Video Surveillance System
ecognition System
ystems
ecognition System
ecognition System
Table of contents
1. Introduction .......................................................................................................... 3
2. Deep Learning Facial Recognition Technology ..................................................... 3
3. Panasonic's Facial Recognition System ................................................................. 4
3.1. High precision .................................................................................................... 5
3.2. Cost Reduction .................................................................................................. 6
3.3. System expandability .......................................................................................... 6
3.4. Data protection………………………………………………………………………………………………………….7
4. Conclusion ............................................................................................................. 7
1. Introduction
Increasing attention is now being given to the monitoring of airports, train stations,
shopping malls, and other important facilities where large numbers of people
gather. In the security industry, there is a growing need for monitoring systems that
can help to stop crimes before they happen. Rather than using high resolution
cameras to passively capture people’s faces and behaviors, there is a growing need
to identify important and suspicious people to provide an active security response.
In order to solve problems such as these, Panasonic has newly developed FacePRO,
a facial recognition system which utilizes deep learning technology. With inbuilt
analytics changing traditional monitoring, the system is better equipped to provide
safer and more secure environments.
2. Deep Learning Facial Recognition Technology
With the Internet of Things (IoT) and Artificial Intelligence (AI) becoming popular
concepts in today’s society, industries are utilizing the ability to link together a
variety of information to make workflows and people’s lives easier. To achieve this,
one of the technologies being applied in each field is the new technology known as
“deep learning.”
In a joint effort with the National University of Singapore, Panasonic has engaged
in research and development that has led to the creation of an original algorithm
combining the machine learning method known as “deep learning” with a similarity
calculation method that suppresses errors. This algorithm has now been put to
practical use in our facial recognition system. As a result, we have been able to
improve the matching failure problems that have plagued conventional facial
recognition technology in cases where faces are angled, changed by aging, or
partially hidden by sunglasses, etc. This success in raising facial recognition accuracy
makes it possible to expand the situations in which facial recognition systems can
be used.
*1 In April 2017, the product achieved the highest leve
(IJB-A Face
NIST (National Institute of Standards and Technology) of the United
in the world.
3. Panasonic
3.1 High precision
The facial recognition software employs a face matching performance engine
that utilizes
University of Singapore. It has been rated at the highest level in the NIST (National
Institute of Standards and Technology) IJB
accurately face matching in s
conventional techniques such as faces that are angled (up to 45 degrees to the left
or right or 30 degrees up or down), changed by aging, or partially hidden by
sunglasses.
NIST (IJB
In April 2017, the product achieved the highest leve
A Face Verification Challenge Performance Report/IJB
NIST (National Institute of Standards and Technology) of the United
in the world.
Panasonic’s Facial Recognition
High precision
The facial recognition software employs a face matching performance engine
utilizes deep learning technology jointly developed with the National
University of Singapore. It has been rated at the highest level in the NIST (National
Institute of Standards and Technology) IJB
accurately face matching in s
conventional techniques such as faces that are angled (up to 45 degrees to the left
or right or 30 degrees up or down), changed by aging, or partially hidden by
sunglasses.
Registeredimage
Matching image
(IJB-A) competition : Competition between corporate and universitie
In April 2017, the product achieved the highest leve
Verification Challenge Performance Report/IJB
NIST (National Institute of Standards and Technology) of the United
s Facial Recognition
High precision
The facial recognition software employs a face matching performance engine
deep learning technology jointly developed with the National
University of Singapore. It has been rated at the highest level in the NIST (National
Institute of Standards and Technology) IJB
accurately face matching in situations that were difficult to handle with
conventional techniques such as faces that are angled (up to 45 degrees to the left
or right or 30 degrees up or down), changed by aging, or partially hidden by
Registered
A) competition : Competition between corporate and universitie
Feature quantification using deep learning
In April 2017, the product achieved the highest level of facial
Verification Challenge Performance Report/IJB
NIST (National Institute of Standards and Technology) of the United
s Facial Recognition
The facial recognition software employs a face matching performance engine
deep learning technology jointly developed with the National
University of Singapore. It has been rated at the highest level in the NIST (National
Institute of Standards and Technology) IJB-A face challenge
ituations that were difficult to handle with
conventional techniques such as faces that are angled (up to 45 degrees to the left
or right or 30 degrees up or down), changed by aging, or partially hidden by
A) competition : Competition between corporate and universitie
(unlike FRVT or FIVE)
Feature quantificationusing deep learning
ial recognition performance in
Verification Challenge Performance Report/IJB -A Face Identification Challenge Performance Report) of
NIST (National Institute of Standards and Technology) of the United States, one of the most authoritative institutes
s Facial Recognition System
The facial recognition software employs a face matching performance engine
deep learning technology jointly developed with the National
University of Singapore. It has been rated at the highest level in the NIST (National
A face challenge
ituations that were difficult to handle with
conventional techniques such as faces that are angled (up to 45 degrees to the left
or right or 30 degrees up or down), changed by aging, or partially hidden by
Feature quantities
(registered image)
Feature quantities
(matching image)
A) competition : Competition between corporate and universitie
(unlike FRVT or FIVE)
Feature quantification using deep learning
recognition performance in the world in a comparison test
A Face Identification Challenge Performance Report) of
States, one of the most authoritative institutes
System
The facial recognition software employs a face matching performance engine
deep learning technology jointly developed with the National
University of Singapore. It has been rated at the highest level in the NIST (National
A face challenge*1
, with the system
ituations that were difficult to handle with
conventional techniques such as faces that are angled (up to 45 degrees to the left
or right or 30 degrees up or down), changed by aging, or partially hidden by
Feature quantities
(registered image)
Feature quantities
(matching image)
Sim
ilarity
calcu
latio
n
A) competition : Competition between corporate and universities to compete latest technology
the world in a comparison test
A Face Identification Challenge Performance Report) of
States, one of the most authoritative institutes
The facial recognition software employs a face matching performance engine
deep learning technology jointly developed with the National
University of Singapore. It has been rated at the highest level in the NIST (National
, with the system
ituations that were difficult to handle with
conventional techniques such as faces that are angled (up to 45 degrees to the left
or right or 30 degrees up or down), changed by aging, or partially hidden by
Iden
tity ju
dgm
en
t
s to compete latest technology
the world in a comparison test
A Face Identification Challenge Performance Report) of
States, one of the most authoritative institutes
The facial recognition software employs a face matching performance engine
University of Singapore. It has been rated at the highest level in the NIST (National
, with the system
conventional techniques such as faces that are angled (up to 45 degrees to the left
or right or 30 degrees up or down), changed by aging, or partially hidden by
s to compete latest technology
*2 The Best Shot License Key must be installed on the
*3 From receiving the face thumbnail to matching
*4 The best case when the time range is narrowed down and the number of search result is 100 faces. Depending on the
search conditions, the search time may take several minutes.
By using the intelligent auto( iA) and Best Shot
i-PRO Extreme cameras, it is also possible to
performance and provide high recognition precision.
(iA function
This function enables cameras to automatically
optimize
The camera detects the moving objects, movement speed, faces, and light
intensity found in video that is us
and backlight, and it
optimal video of the subject.
(Best Shot function
This function automatically selects several images suitable for facial
recognition
front of the camera, and it sends only those selected images to the server.
This enables high
without putting a load on the network.
With these functions, high
with a maximum of 30,000 faces registered) and high
data for 5 million faces in 3 seconds
Angled F
Left/Right ±45°Up/down ±30°
The Best Shot License Key must be installed on the
From receiving the face thumbnail to matching
The best case when the time range is narrowed down and the number of search result is 100 faces. Depending on the
search conditions, the search time may take several minutes.
By using the intelligent auto( iA) and Best Shot
PRO Extreme cameras, it is also possible to
performance and provide high recognition precision.
iA function)
This function enables cameras to automatically
optimize the settings accordingly to improve the clarity of the video images.
The camera detects the moving objects, movement speed, faces, and light
intensity found in video that is us
and backlight, and it
optimal video of the subject.
(Best Shot function
This function automatically selects several images suitable for facial
recognition from the multiple face images captured when a person passes in
front of the camera, and it sends only those selected images to the server.
This enables high
without putting a load on the network.
With these functions, high
with a maximum of 30,000 faces registered) and high
data for 5 million faces in 3 seconds
Angled Face Left/Right ±45° Up/down ±30°
The Best Shot License Key must be installed on the
From receiving the face thumbnail to matching
The best case when the time range is narrowed down and the number of search result is 100 faces. Depending on the
search conditions, the search time may take several minutes.
By using the intelligent auto( iA) and Best Shot
PRO Extreme cameras, it is also possible to
performance and provide high recognition precision.
This function enables cameras to automatically
the settings accordingly to improve the clarity of the video images.
The camera detects the moving objects, movement speed, faces, and light
intensity found in video that is us
and backlight, and it optimizes
optimal video of the subject.
(Best Shot function*2
)
This function automatically selects several images suitable for facial
from the multiple face images captured when a person passes in
front of the camera, and it sends only those selected images to the server.
This enables high-quality images suitable for facial recognition to be sent
without putting a load on the network.
With these functions, high-speed matching (matching in approx. 1 second
with a maximum of 30,000 faces registered) and high
data for 5 million faces in 3 seconds
Sunglasses
The Best Shot License Key must be installed on the i-PRO Extreme camera
From receiving the face thumbnail to matching server
The best case when the time range is narrowed down and the number of search result is 100 faces. Depending on the
search conditions, the search time may take several minutes.
By using the intelligent auto( iA) and Best Shot
PRO Extreme cameras, it is also possible to
performance and provide high recognition precision.
This function enables cameras to automatically
the settings accordingly to improve the clarity of the video images.
The camera detects the moving objects, movement speed, faces, and light
intensity found in video that is usually hard to see due to subject movement
optimizes the settings in real
optimal video of the subject.
This function automatically selects several images suitable for facial
from the multiple face images captured when a person passes in
front of the camera, and it sends only those selected images to the server.
quality images suitable for facial recognition to be sent
without putting a load on the network.
speed matching (matching in approx. 1 second
with a maximum of 30,000 faces registered) and high
data for 5 million faces in 3 seconds*4
) can be
Sunglasses Surgical M
PRO Extreme camera
The best case when the time range is narrowed down and the number of search result is 100 faces. Depending on the
search conditions, the search time may take several minutes.
By using the intelligent auto( iA) and Best Shot*2
function
PRO Extreme cameras, it is also possible to maximize
performance and provide high recognition precision.
This function enables cameras to automatically
the settings accordingly to improve the clarity of the video images.
The camera detects the moving objects, movement speed, faces, and light
ually hard to see due to subject movement
the settings in real
This function automatically selects several images suitable for facial
from the multiple face images captured when a person passes in
front of the camera, and it sends only those selected images to the server.
quality images suitable for facial recognition to be sent
without putting a load on the network.
speed matching (matching in approx. 1 second
with a maximum of 30,000 faces registered) and high-speed searching (searching
) can be realized.
Surgical Mask
Facial Recognition Server
Improve face match processing speed
The best case when the time range is narrowed down and the number of search result is 100 faces. Depending on the
function installed on Panasonic
maximize facial recognition engine
This function enables cameras to automatically recognize
the settings accordingly to improve the clarity of the video images.
The camera detects the moving objects, movement speed, faces, and light
ually hard to see due to subject movement
the settings in real-time to capture more
This function automatically selects several images suitable for facial
from the multiple face images captured when a person passes in
front of the camera, and it sends only those selected images to the server.
quality images suitable for facial recognition to be sent
speed matching (matching in approx. 1 second
speed searching (searching
.
Changed by A
Facial Recognition Server
Improve face match processing speed
The best case when the time range is narrowed down and the number of search result is 100 faces. Depending on the
installed on Panasonic
facial recognition engine
the scene and
the settings accordingly to improve the clarity of the video images.
The camera detects the moving objects, movement speed, faces, and light
ually hard to see due to subject movement
time to capture more
This function automatically selects several images suitable for facial
from the multiple face images captured when a person passes in
front of the camera, and it sends only those selected images to the server.
quality images suitable for facial recognition to be sent
speed matching (matching in approx. 1 second
speed searching (searching
Changed by A
Facial Recognition Server
Improve face match processing speed
The best case when the time range is narrowed down and the number of search result is 100 faces. Depending on the
installed on Panasonic
facial recognition engine
the scene and
the settings accordingly to improve the clarity of the video images.
The camera detects the moving objects, movement speed, faces, and light
ually hard to see due to subject movement
time to capture more
from the multiple face images captured when a person passes in
front of the camera, and it sends only those selected images to the server.
quality images suitable for facial recognition to be sent
speed matching (matching in approx. 1 second*3
even
speed searching (searching
Changed by Aging
Facial Recognition Server
Improve face match processing speedImprove face match processing speed
3.2 Cost Reduction
With conventional facial recognition systems, all captured images are sent to the
server and the server performs face detection and recognition resulting in the
processing load being concentrated on the server. Systems also tend to be
large-scale because of the large bandwidth required to send all the images and the
large volume of hard drive space required to save those images.
To provide a solution to data storage, Panasonic’s facial recognition system uses
the Best Shot function to only send Best Shot images to the server, thereby making
it possible to eliminate the need for wide bandwidth and reduce the network load.
In addition, performing facial recognition on the server using the Best Shot images
reduces server load and hard drive volume usage so up to 20*5
cameras can be
connected to a single server.
3.3 System expandability
・Simple batch registration of a maximum of 10,000 faces is possible with the
standard Facial Recognition Server Software (WV-ASF950). The optional Face
Registration Expansion Kit (WV-ASFE951W) enables registration of up to 30,000
faces at large-scale facilities.
・Face detection, face matching, and tracking with recorded video can be performed
in the same GUI by integrating management with the WV-ASM300 or
WV-ASE231W client software for Panasonic i-PRO monitoring systems. This
eliminates the need for the dedicated operation and management required for
conventional facial recognition systems and enables work to be centralized.
*5 The number of cameras that can be connected depends on the number of people passing by the camera and the retention
period for face images.
・Additional
The face recognition accuracy improved by the deep learning techn
be used for unregistered face detec
In areas where outsiders and suspicious persons are not allowed to enter such as
the backyard of a store and company's development
persons other than registered faces with high accuracy and to notify an alert to field
surveillance
As a system configuration, i
maximum 10 cameras can be connected in a face matching server.
*6
Detect unregistered faces in approx. 3 second even with a maximum of 30,000 faces registered depending of setting
Unregistered
3.4 Data protection
Whilst working to increase the accuracy of facial recognition, Panasonic is also
providing
security of important data
as a whole.
・Can be selected to store facial
・Face thumbnail in DB will be encrypted.
・SSL communication(Camera
・Registered faces
・Face matching
・Retention period and number of face data
・Log data for all the operation can be saved.
Additional Unregistered
The face recognition accuracy improved by the deep learning techn
be used for unregistered face detec
In areas where outsiders and suspicious persons are not allowed to enter such as
the backyard of a store and company's development
persons other than registered faces with high accuracy and to notify an alert to field
surveillance can be reinforce
As a system configuration, i
maximum 10 cameras can be connected in a face matching server.
Detect unregistered faces in approx. 3 second even with a maximum of 30,000 faces registered depending of setting
Unregistered Face Detection and Registered Face
3.4 Data protection
working to increase the accuracy of facial recognition, Panasonic is also
capabilities such as
security of important data
as a whole.
Can be selected to store facial
Face thumbnail in DB will be encrypted.
SSL communication(Camera
*Face cannot
egistered faces
ace matching data can be deleted
etention period and number of face data
Log data for all the operation can be saved.
Unregistered Face
The face recognition accuracy improved by the deep learning techn
be used for unregistered face detec
In areas where outsiders and suspicious persons are not allowed to enter such as
the backyard of a store and company's development
persons other than registered faces with high accuracy and to notify an alert to field
can be reinforced.、
As a system configuration, in the case of unregistered face detection
maximum 10 cameras can be connected in a face matching server.
Detect unregistered faces in approx. 3 second even with a maximum of 30,000 faces registered depending of setting
Face Detection and Registered Face
3.4 Data protection
working to increase the accuracy of facial recognition, Panasonic is also
capabilities such as the ones listed
security of important data, and therefore
Can be selected to store facial
Face thumbnail in DB will be encrypted.
SSL communication(Camera
Face cannot be reproduce
egistered faces can be deleted
data can be deleted
etention period and number of face data
Log data for all the operation can be saved.
ace Detection
The face recognition accuracy improved by the deep learning techn
be used for unregistered face detection as well as registered face re
In areas where outsiders and suspicious persons are not allowed to enter such as
the backyard of a store and company's development
persons other than registered faces with high accuracy and to notify an alert to field
、
n the case of unregistered face detection
maximum 10 cameras can be connected in a face matching server.
Detect unregistered faces in approx. 3 second even with a maximum of 30,000 faces registered depending of setting
Face Detection and Registered Face Detection simultaneously.
working to increase the accuracy of facial recognition, Panasonic is also
the ones listed
therefore build safety management into the system
Can be selected to store facial data.
Face thumbnail in DB will be encrypted.
SSL communication(Camera-Server-Client)
reproduced from Metadata.
can be deleted after valid
data can be deleted after
etention period and number of face data
Log data for all the operation can be saved.
etection setting
The face recognition accuracy improved by the deep learning techn
tion as well as registered face re
Detect unregistered person
In areas where outsiders and suspicious persons are not allowed to enter such as
the backyard of a store and company's development department, it is detect
persons other than registered faces with high accuracy and to notify an alert to field
n the case of unregistered face detection
maximum 10 cameras can be connected in a face matching server.
Detect unregistered faces in approx. 3 second even with a maximum of 30,000 faces registered depending of setting
Detection simultaneously.
working to increase the accuracy of facial recognition, Panasonic is also
the ones listed below to strengthen the information
build safety management into the system
Face thumbnail in DB will be encrypted.
Client)
from Metadata.
after valid time frame
after valid time frame
etention period and number of face data collected can be set
Log data for all the operation can be saved.
The face recognition accuracy improved by the deep learning techn
tion as well as registered face re
Detect unregistered person
In areas where outsiders and suspicious persons are not allowed to enter such as
department, it is detect
persons other than registered faces with high accuracy and to notify an alert to field
n the case of unregistered face detection
maximum 10 cameras can be connected in a face matching server.
Detect unregistered faces in approx. 3 second even with a maximum of 30,000 faces registered depending of setting
working to increase the accuracy of facial recognition, Panasonic is also
below to strengthen the information
build safety management into the system
from Metadata.
time frame.
valid time frame.
collected can be set
The face recognition accuracy improved by the deep learning technology can
tion as well as registered face recognition.
Detect unregistered person
In areas where outsiders and suspicious persons are not allowed to enter such as
department, it is detect
persons other than registered faces with high accuracy and to notify an alert to field
n the case of unregistered face detection setting,
maximum 10 cameras can be connected in a face matching server.
Detect unregistered faces in approx. 3 second even with a maximum of 30,000 faces registered depending of setting
working to increase the accuracy of facial recognition, Panasonic is also
below to strengthen the information
build safety management into the system
collected can be set.
ology can
cognition.
Detect unregistered persons
In areas where outsiders and suspicious persons are not allowed to enter such as
department, it is detect*6
persons other than registered faces with high accuracy and to notify an alert to field
setting,
Detect unregistered faces in approx. 3 second even with a maximum of 30,000 faces registered depending of setting
working to increase the accuracy of facial recognition, Panasonic is also
below to strengthen the information
build safety management into the system
4. Conclusion
By utilizing deep learning technology, Panasonics facial recognition system makes it
possible to build a high-precision facial recognition and video monitoring system
that can perform more complex and sophisticated monitoring centered on people.
Furthermore, by cutting data volumes, this technology will also contribute to
reductions in transmission, network-construction, and operation costs.