Dynamic Face Recognition Committee Machine

1
Dynamic Face Recognition Committee Machine

• Presented by
• Sunny Tang

2
Outline

• Introduction
• Previous Work
• Dynamic Architecture
• Face Verification System
• Conclusion Future Work

3
Committee Machine

• Train a committee of estimators and combine the
  individual predictions
• Motivation
• Achieve better performance
• Reduce computational complexity
• Type of Committee Machine
• Static structure
• Dynamic structure

4
Static Structure

• Ignore input signals
• Fixed weights
• Examples
• Majority Voting
• Ensemble averaging
• Bagging

5
Dynamic Structure

• Employ input signal to improve the classifiers
• Variable weights
• Examples
• Gating networks
• Hierarchical mixture of experts

6
Previous Work

• Static Face Recognition Committee Machine consist
  of 5 experts

7
Drawback

• Expert weighting depends on overall performance
  of a particular face database
• Weight is fixed once the system is trained
• Only frontal faces are used for identification /
  verification

8
Dynamic Architecture

• Keep performance of experts on different face
  databases
• Gating Network consisting of a neural network to
  determine which performance to use as weight

Database Image
ORL 400
Yale 165
CVL 800
Umist 560
HRL 1370
Feret 1200
9
Dynamic Architecture

• Gating Network
• Input image x
• Output
• Px Performance for xs database

r
?
x
x
10
Face Verification System

• Biometric Security Application
• Personal authentication
• Target
• Low false acceptance
• Low false rejection
• Two face images are used
• Frontal
• Profile
• Hierarchical Structure

11
System Snapshot
12
Hierarchical Dynamic Architecture
13
Face Verification System

• Accept when
• Both frontal and profile results match the
  claimed identity
• Each committee machine has overall confidence
  over a selected threshold
• Feedback Mechanism
• Adjust individual experts weight
• Update corresponding performance

14
Conclusion Future Work

• Conclusion
• We propose a framework for a dynamic committee
  machine
• We design a face verification system for security
  purpose
• Future Work
• Work on the system and get experimental result