Documents
Resources
Learning Center
Upload
Plans & pricing Sign in
Sign Out

Face Recognition Committee Machine

VIEWS: 10 PAGES: 23

									 Face Recognition
Committee Machine

Term Three Presentation
         by
    Tang Ho Man
                   Outline

 Introduction
 Algorithms Review
 Face Recognition Committee Machine (FCRM)
 Distributed Face Recognition System (DFRS)
 Experimental Results
 Conclusion and Future Work
Q&A
                   Introduction

 Applications in security
     Authentication
     Identification
 Authentication measures
     Password
     Card/key
     Biometric
                   Introduction

 Face Recognition
     Training phase
     Recognition phase
 Objectives
     Comparison of different algorithms
     Face Recognition Committee Machine
     Distributed Face Recognition System
                     Review

 Algorithms in Committee Machine
     Eigenface
     Fisherface
     Elastic Graph Matching (EGM)
     Support Vector Machine (SVM)
               Review – Eigenface

 Application of Principal Component Analysis (PCA)
     Find eigenvectors and eigenvalues of covariance matrix
      C from training images Ti:




 Training & Recognition
     Project the images on face space
     Compare Euclidean distance and choose the closest
      projection
               Review – Fisherface

 Similar to Eigenface
      Application of Fisher’s Linear Discriminant (FLD)
      Minimize inner-class variations and maintain between-
       class discriminability
 Projection finding
      Between class scatter
      Within class scatter
      Projection
                      Review – EGM

 Based on dynamic link architecture
      Extract facial feature by Gabor wavelet transform as a jet
      Face is represented by a graph G consists of N nodes of jets
 Compare graphs by cost function
      Edge similarity
      Vertex similarity
      Cost function
                    Review – SVM

 Look for a separating
  hyperplane H which
  separates the data with the
  largest margin
 Decision function
 Kernel function
      Polynomial kernel
      Radial basis kernel
      Hyperbolic tangent kernel
                FRCM - Overview

 Mixture of five experts
      Eigenface
      Fisherface
      EGM
      SVM
      Neural network
                  FRCM - Overview

 Elements in voting machine
     Result r(i)
         Individual expert’s result for test image
     Confidence c(i)
         How confident the expert on the result
     Weight w(i)
         Average performance of an expert
      FRCM - Result & Confidence

 Eigenface, Fisherface, EGM
     Use K nearest-neighbour classifiers
       Five nearest training set images are chosen
       Count number of votes for each recognized class

     Result
     Confidence
       FRCM - Result & Confidence

 SVM
      One-against-one approach with maximum voting used
      For J different classes, J(J-1)/2 SVM are constructed
      Confidence:

 Neural network
      Binary vector of size J for target representation
      Result:
           Class with output value closest to 1
      Confidence:
           Output value
           FRCM - Voting Machine

 Ensemble results, confidences from experts to
  arrive a final result
 Score function:

 Final result – Highest score class
 Advantages
      High performance
      High confidence
                         DFRS

 Motivation
     Real face recognition application
     Face recognition on mobile device
 Consists of
     Face Detection
     Face Recognition
              DFRS - Limitations

 Memory
     Little memory for mobile devices
     Requirement for recognition




 Processing power
                 DFRS - Overview

 Client-Server approach
     Client
       Capture
       Ensemble

     Server
         Recognition
               DFRS - Testing

 Implementation
     Desktop (1400MHz)
     Notebook (300MHz)
      Experimental Results - Database

 ORL Face Database
     40 people
     10 images/person



 Yale Face Database
     15 people
     11 images/person
     Experimental Results - ORL

 ORL Face database
     Experimental Results - Yale

 Yale Face Database
       Conclusion and Future Work

 Conclusion
     Comparison of different algorithms
     Committee machine improves accuracy
     Feasible on mobile device
 Future Work
     Use of dynamic structure
     Include more expert in the committee machine
     Implementation on PDA/Mobile
Question & Answer Section

         Thanks!

								
To top