Dynamic Scalable Distributed Face Recognition System Security Framework

1
Dynamic Scalable Distributed Face Recognition
System Security Framework

• by
• Konrad Rzeszutek
• B.S. University of New Orleans, 1999

2
Overview

• Purpose
• History of face recognition
• Problems
• Solution
• Apollo
• Components of Apollo
• Face recognition technology used
• Motion detection
• Future work

3
Applications of face recognition

• Surveillance Systems
• Biomedical Systems (eye-replacement)
• Military (anti-terrorist groups)
• Security (logon authorization)
• Autonomous vehicle navigation
• many more

4
History

• Sir Francis Galton (1888) Automatic method of
  classification of French prisoners. Called it
  mechanical selector.
• Late 1960s started.
• In 1980 research picked up dramatically.
• Two branches of face recognition
• Geometric-features
• Template matching

5
Geometric - profile

• Profile features.
• 8-100 control points
• Six control points using B-spline
• U.S. INS uses this one extensively.

6
Geometric - frontal

• Frontal features
• 8-16 features
• Various distance from right and left eye to nose,
  nose to chin, eye to eye, etc.
• Nose width, chin radii, eyebrow thickness, etc

7
Template matching

• Face images are represented as vectors in an
  array (each image is identified as ?k)
• Computations are carried on the model arrays
  resulting in hash values.
• The matching image hash value is compared against
  the template images hash values.

8
Template matching, part 2

• The distance from the training images hash value
  determines the match.
• Euclidian distance mostly used.

9
Principal Component Analysis

• Turk and Pentland Eigenface.
• Most simplest uses the whole image face as a
  template.
• Variations of this use infrared images.

10
Template matching .. more

• Isodensity line maps (brightness of image viewed
  as height of the mountain isodensity lines
  corresponds to contour lines of equal altitude).
• Neural network eye and mouth regions feed into
  multi-layer perception engine that carries of the
  classification
• .. Other are mostly various combinations of these
  two branches of face-recognition technologies.

11
Problems

• Work done on a very selective set of face images,
  mostly
• In upright position
• Lighting and background controlled
• Either in frontal or profile view
• Have no occlusions, facial hair
• Most test cases are white males

12
Solution

• A distributed system capable of handling large
  load of images, analyze them in near-real time,
  provide support for future enhancements and be
  scalable to the load.
• Separates the functionality of a security system
  in three modules recognition, notification, and
  replay.

13
Apollo
14
Components

• Ares - the thin-client providing camera feed.
• Hermes the police officer directing traffic
• Demeter the storage for later replay of
  camera-feed
• Nemesis the face recognition engine
• Mors the notification event server

15
Ares

• Passes the real-time camera feed through a
  motion-detection engine.
• Transmits the feed to Nemesis for
  face-recognition and Demeter for storage.
• Uses Jini/RMI to localize required components.

16
Hermes

• Collects information about load of each
  components.
• Is queried for its knowledge whenever a system in
  a pool requires another component.
• Scalable many of these system (Hermes) can
  coexist and provide the load information.

17
Demeter

• Stores camera-feed for later replay and for
  storing the camera-feed on a archive media (WORM).

18
Nemesis

• Face recognition module. Uses Eigenfaces
  technique to match images in near-real time.
• Can be extended to use more algorithms and check
  image using many techniques.
• If match found, an event is sent to Mors.

19
Mors

• Receives events which notify about a possible
  face match.
• Centralized pool where humans can visually check
  results and carry out proper procedures.

20
Face recognition - Nemesis

• Eigenfaces algorithm finds the PCA of faces, or
  eigenvectors of the covariance matrix.
• Each eigenvalue can be thought as an amount
  which, when subtracted from each diagonal matrix,
  makes the matrix singular. Eigenvectors are
  characteristics vectors of the matrix (from
  Digital Image Processing by Castleman)

21
Eigenvalues and Eigenvectors

• We are looking for ? (eigenvectors) and ?
  (eigenvalues) defined as
• C? ? ?
• Where C is our covariance matrix of the
  normalized face-vector ??1 ?2 ?M

22
Weights

• After the computation of eigenvalues and
  eigenvetors, we use M most significant
  eigenfaces (where each eigenface is the linear
  combination of eigenvalues and the face-image) to
  form a face subspace.
• From the face subspace we calculate the weights
  (where ?T?1 ?2 .. ?M)
• ?k ?kT ? k 1,2, , M

23
Matching

• We use the calculated weights to determine if the
  image is recognized.
• Usually we use Euclidian distance.

24
Motion Detection

• Motion detection is used on the client side
  Ares.
• It saves bandwidth and saves only frames that
  have content.
• Algorithm uses two threshold functions
• The first is used to accommodate for possible
  artifacts introduced by the camera.
• Second determines the if there is motion
  depending on the count of clusters of pixels
  that changed.

25
Motion Detection, .. more

• Red is the cluster count.

26
Future work

• Use more face recognition technologies so each
  can complement each other.
• Expand framework to include other recognition
  technologies iris, speech, etc.
• Improve motion detection engine.
• Face operations automatically removing
  background.
• Generate from one face a multitude of other faces
  with different alternations bear (or lack of
  it), long hair, etc to expand possibly match.