Face Recognition Committee Machine
description
Transcript of Face Recognition Committee Machine
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 chosenCount 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
CaptureEnsemble
ServerRecognition
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!