Name: MorphoFace

Text: U

Executive summary
Morpho Face Investigate Pilot (MFIP) is a user friendly face recognition
system. It implements the fundamental functional capabilities o f 2D
frontal face image recognition.
MFIP has been developed mainly t o evaluate operationally the benefit
of face recognition solutions.

Its main features are
Load and manage a database of up to 350,000 portraits, with an option for
an extension to 2,000,000 portraits
Search one or more images against the portrait database using the Morpho
face recognition technology
Acquire face images from files, or using a camera or a scanner
Extract face images from video files and search them against the database
Extract faces from images showing multiple people
Check the result of face recognition searches and report on the decision
made
Manage subsets of the portrait database - called watch lists - to enable
search scope restriction
This document is intended for potential users of MFIP; it introduces the
concepts of face recognition and describes how MFIP can be deployed and
used with minimal effort even for users with no or very limited knowledge
of face recognition.

Face recognition concepts
Face recognition is a recent although reliable technology
to compare frontal face images, when the face images are
sharp and not too small (at least 40 pixels between the eyes,
preferably more). Morpho face recognition can search any kind
of face images; even poor quality images, however searches are
more reliable with good quality data.
The main steps of face recognition are listed below for one-tomany comparisons:
F-1-WImage acquisition can be done from live or file
sources (e.g. from a camera or a .JPG file), still
or video images.
During face acquls~t~on,
a unique identifier of
the acqu~redimage must be allocated .

This step aims at finding where the face (or faces)
is located in the image, and pinpointing the eye
centres.
Template Extraction

A template is a representation of the image that
is suitable for image comparison. A template
may represent either visible features of a portrait
(e.g. nose or eyebrows location), or purely
mathematical data such as the results of applying
one or more filters to all or part of the image.

Quality Control
As the image is analysed and the template
com~uted,it is possible to determine whether the
template is reliable or not, and whether the image
is of high or low quality.
A low quality score may require that an operator
manually confirm the location of some feature on
the image: this enhances the overall accuracy of
the system.

Furthermore, in case a face image is not properly
oriented (not fully frontal), a tool is provided to
generate a frontal view of the person from the
existing image data.
Matching
Matching a face template against a collection
of face templates. Each comparison produces a
score: the higher the score, the more likely the two
compared faces are similar.

-

Match results review

mhemk ~

~

3

The matching system returns a (possibly empty)
candidate list, sorted in decreasing scores. If the
%
candidate list is not empty, and if an operator has
specified the request for a visual comparison of
candidates, the face images matching those of the
candidates are retrieved from the database.
Following review, other actions can take place such
as, for Instance, inserting the portrait that was just
processed along with its template In the database
or printing a matching report.

MFlP architecture and technical
characteristics
MFlP relies on client-server architecture
The server includes database and matching services
The clients are running miscellaneous applications to enable the end users
to perform face recognition tasks. MFlP can support up to eight clients, with
deployment restrictions.

YFlP client

I

Figure 1: MFlP log~calarchitecture.

The applications have the following purpose
The Database Creation application reads face images from a directory,
extracts the corresponding templates, and stores both images an templates
within the database
The Quality Control application presents the face images requiring quality
control to operators, so that they can review the eyes location as necessary
The verification application is the main application for operational face
recognition. It includes face image acquisition, one-to-one or one-to-many
search requests against the database with review of matching results, and
video analysis
The Face DB Administration application includes database content display,
import and export capability, and other administration capabilities
The Watchlist Administration application takes care of watchlist creation,
updates and deletion
The Multiportrait application crops face sub images from a picture showing
multiple people. It does not directly integrate with the server: it only
generates sub images.

Technically, MFlP requires the following environment
Platform:

- For 350,000 faces or less: client, server or standalone configurations are
based on 32 bits PC running Windows XP professional SP2

- For more than 350,000 and up to 2,000,000

faces, it is required to have
a server based on 64 bit PC running Windows Server 2003 R2 SP2 standard
64 bits; clients are 32 bit PC running Windows XP professional SP2

Database management: SQL server for workgroup. For very small
configurations of 5000 faces or less, it is possible to use SQL server express
Compatible image format: jpeg, jpeg 2000, Tiff and bmp
Compatible video file format: .avi files with mpeg4 or mjpeg codec
Scanner: Fujistu FI-60 or Epson V700
Camera: Canon Reflex EOS 450D or Canon Reflex EOS 1000D or Canon
Powershot 5x1 10 IS.

Loading the database
Because most users of MFlP already have an existing collection of
faces images, it is required before MFlP can be used to load all these
images into the system.
Loading means:
Inserting the images into MFIP database
Extracting a template from each image
Storing the template into MFIP database and MFIP rnatcher

The Database Creation application achieves all these tasks.
It works as follows:

1 A user opens the Database Creation application from the desktop

2 He or

she selects one or more directories where the face images to be
inserted into MFlP are stored

3 He or she starts the loading process
4 The loading process reads the selected image files one by one, extracts the
associated templates, and creates into the database and into the matcher
the corresponding entries. Each portrait in the database is identified with
the original name of the image file (without its extension)

S A report is generated

A

During the operation, a progress bar shows that the application is active
and how many images have already been processed.

-*-

0
1
.
1
1

Figure 2:
the Database Creation
application interface,
showing progress during
loading.

gure 3: Quality Control application,
jt of all images with a quality score

elow 200.

It may happen that some of the images loaded into the database could be
of low quality and that the face could not be easily located in the image. To
overcome this potential issue, it is necessary to have an operator review the
face location on low quality images. The "Quality Control" application is
available to perform this task.
The Quality Control application reads the database to find the low quality
images, using a quality score computed during template extraction (for
instance within the Database Creation application). It displays the list of all
images with a quality score below a certain level.
From this list, the operator can select and review any face image, and change
as necessary the points showing the eyes location. If some image does not
include any face or if it is not suitable for face recognition (e.g. profile view)
then the operator can remove it from the database.

Searching face images
The main puqmw af MFlP is to acquire face images and search them
against the already existing collection of faces. Once searching b
done, it is possible to review the pmpased candidate imags, me55
whether they really come from the same person a the search image,
and insert them hto the database.
All these tasks are performed using the Verification application that can be
started from the user desktop.

Figure4 Verifmtion
application, start xreen.

>

There are mainly 2 kinds of 2D image searching: 1:l searches (also known as authentication)
and 1:n searches (also known as ~dentification).The verification application can handle both:
1:I searches are handled from the rrghtmost buttons of the top tool bar, and l:n searches are
handled from the leftmost buttons of the top tool bar.
To acquire face images, three means are proposed by MHP:
Acquisition from a scannec depending on a configuration parameter, MFlP can scan
a picture and look for a face in the scanned image
Acquisition from a camera: in case of live acquisition, dedicated face acquisrtion software
is avarlable. It helps the users getting hrgh quality portraits, cropped as recommended in the
ICAO standard, and it displays compliancy status of the acquired and cropped image for all
/CA0 cr~teria
Acquisition from files. This acquisition mode is the only one that is compatible with
batch operations. the user can select multiple face images (up to 100) to be compared to
the database

1
.

1
l

;

Whatever acquisition means IS used, the
acqurred images are analysed automatically,
a template is extracted, and a quality mark
is given. If quality is not high enough,
the acquired images must be reviewed to
perform a qual~tycheck.

Figure 5:Acquisition from files: defining the quality check
and matching parameters.

1

1I

Quality checking includes:
The ability to change eyes location (called
2DQCl
The ability to "turn the face" using a 3 D
model, in order to reconstruct a frontal
view of the searchedperson (called 3D QC)
Once the images are corrected, the
searches are done automatically.

Fi ure 6: Qualitycontrol within the verification application.
--

In case on an authentication, the end users
can check the result using the following
screen:

Figure 7: Authentication verification screen within the
Verification application.

Figure 8: Identification verification screen within the
Verification application.

Figure 9: Multi Portrait application.

In case of identification, the user can check
the candidates using the following screen.
In both cases, tools are provided to make
image comparison easier:
Change the magnification rate
Adjust image size so that faces are seen at
the same scale
Convert to grey scale
Create an horizontal or vertical merge of
the search and candidate images
Make a hit no hit decision
Following a search, a matching report can
be printed, and the searched image can be
inserted into the database.

In some cases, a single image includes many
portraits: the Multi Portrait application can
be used to find and extract conveniently
portraits from such images, individually or
by batch.
The Multi Portrait application is started
from the user's desktop. It asks the user
to select one or more images, and extracts
all individual portraits it can find on these - images. Whileextraction Brunning, a progre25~
. .
bar is displayed.
-The face images extracted can be displayed
for review, and are saved in a directory.
Once they have been extracted, they can be
selected from the Verification application
for further processing.

Searching face images extracted from
video files
Because sometimes fare images could be found in video faotage,
MFIP includes a video analyser capability. This capabiJQ is available
inside the Verification application, at the same level as the static file
searching capability. Video analysis includes extracting and tracking
face i m a g barn
~
a video file, and matching the extraded images
against the database.
To trigger the analysis of a video, the user first needs t o select a video file from
a directory. A short wizard enables the operator t o define the face extraction
and matching parameters.
Then the video file is analysed, and displayed at the same speed. The face
images extracted from the video are displayed t o the user on the right size
of the screen.

Once the analysis is completed,
the operator can display any subsequence of the video, forward
and backward, accessed either
sequentially or directly from any
extracted face image (access to
the first frame in which the person
can be seen, or to the last image
on which a person can be seen).

Figure 10: Multi Portrait application.

It n also possible to select any of the
extracted face ~mages,and check
the corresponding search results
for
The search result d~splayn s~m~lar
faces extracted from vldeos and for
statlc face ~mages.Thesamedecls~on
and tools are available

Managing MFlP
Because it is necessary to manage the face image database and look
at its content, MFlP includes the Face DB Administration application.
It offers:
Database management functions
Slmple databasesearchesand display
of result~ngface images
Face record deletion and edition
(associatecomments to face records)
Export selected portrait identifiers
in a file
Export portra~trecords fimage and
template)

Semier management
Reload the matcher database from
the SQL server database to ensure
synchronicity

Figure 11: Face DB administration application.

Reports generatton
Quality control statistics
Database face records list

I

Statisticson quality and size of face
images

Backup and restore operation are performed using the SQL server interface
directly.
The Watch list management application manages sets of portraits that can be
used to restrict search scope. It includes the following capabilities:
Create a watch list
Delete a watch list
Insert into a watch list all faces specified by identifier in a file

-

MFlP also includes miscellaneous capabilities such as automatic import from a
directory, and a limited set of web services, mainly des~gnedfor demonstratron
purpose.

Sample project description
Because MFlP has been specifically designed for technology trials, its
implementation is extremely simple.

The first step is to make a simple project description:
Purpose of the trial
Size of the portrait database
Number of workstations required
Morpho will help the potential MFlP user to define the project as necessary.
When the scope of the project is defined, Morpho defines a suitable hardware
configuration and it is agreed between Morpho and the potential customer
who will provide the hardware and SOL server. Of course Morpho always
supplies the face recognition system itself and its associated services.
Upon contract signature, a delivery date is agreed. As MFlP is not customizable,
it can be delivered very shortly after contract signature, as soon as the
necessary hardware is available and installed. Because MFlP is so user friendly,
training can be completed in 2 or a maximum of 3 days after installation, if
the potential customer wants help with the loading of the initial database.
It is recommended to include in the contract a maintenance option, whereby
telephone support will be provided during working hours, and fix will be
delivered to the customer in the unlikely event of failure.

9 SAFRAN
U

I

Document Path: ["brochure444.pdf"]

e-Highlighter

Click to send permalink to address bar, or right-click to copy permalink.

Un-highlight all Un-highlight selectionu Highlight selectionh