A Face processing system Based on Committee Machine: The Approach and Experimental Results

1
A Face processing system Based on Committee
Machine The Approach and Experimental Results

• Presented by Harvest Jang
• 29 Jan 2003

2
Outline

• Introduction
• Background
• Face processing system
• System Architecture
• Face Detection Committee Machine
• Face Recognition Committee Machine
• Experimental result
• Conclusion and Future work

3
Introduction

• Information retrieval from biometric technology
  become popular
• Human face is one of the input source that can
  get easily for further processing
• A wide range of usage for face processing system,
  for example,
• Person identification system
• Video content-based information retrieval
• Security entrance system

4
Background

• Homogenous committee machine
• Train experts by different training data sets to
  arrive a union decision
• For example
• Ensemble of networks
• Gating network
• Mixture of experts (neural networks or RBF)
• We propose a heterogeneous committee machine for
  face processing
• Train different classifiers from different
  approaches to make the final decision
• Capture more features in the same training data
• Overcome the inadequacy of each single approach

5
Face processing system

• Three main components
• Pre-processing
• Face Detection Committee Machine (FDCM)
• Face Recognition Committee Machine (FRCM)

Fig 1 System architecture
6
Pre-processing

• Transform to YCrCb color space
• Use ellipse color model to locate the flesh color
• Perform morphological operation to reduce noise
• Skin segmentation to find face candidates

Fig 2 2D projection in the CrCb subspace (gray
dots represent skin color samples and black dots
represent non-skin tone color)
Fig 3 Pre-processing step (a) original images,
(b) binary skin mask, (c) binary skin mask after
morphological operation and (d) Face candidates
7
Pre-processing

• To detect different size of faces, the region is
  resized to various scales
• A 19x19 search window is searching around the
  re-sized regions
• Histogram equalization is performed to the search
  window

Fig 4 Face detection step
8
Face Detection Committee Machine

• Compose of three approaches
• Neural network
• Sparse Network of Winnow (SNoW)
• Support vector machine (SVM)

Fig 5 System architecture for FDCM
9
FDCM Problem modeling (1)

• Based-on the confidence value of each expert
  i

Fig 6 The distribution of confident value of the
training data from three different approaches
10
FDCM Problem modeling (2)

• The confidence value of each expert are
• Not uniform function
• Not fixed output range (e.g. 1 to 1 or 0 to 1)
• Normalization is required using statistics
  information getting from the training data
• where is the mean value of training face
  pattern from expert i and is the standard
  derivation of training data from expert i

11
FDCM Problem modeling (3)

• The information of the confidence value from
  experts can be preserved
• The output value of the committee machine can be
  calculated
• where is the criteria factor for expert i and
  is the weight of the expert i

12
Face Recognition Committee Machine

• Mixture of five experts

Fig 7 System architecture for FRCM
13
FRCM

• Result r(i)
• Individual experts result for test image
• Confidence c(i)
• How confident the expert on the result
• Weight w(i)
• Average performance of an expert

14
FRCM Problem modeling (1)

• Eigenface, Fisherface, EGM
• K nearest-neighbor classifiers
• SVM
• One-against-one approach used
• For J different classes, J(J-1)/2 SVM are
  constructed
• Result value
• Confidence value
• where c(i) is the confidence value for expert i,
  r(i) is the result of the expert i and v() is
  the highest votes in class j

15
FRCM Problem modeling (2)

• Neural network
• Result value
• Class with output value closest to 1
• Confidence value
• Output value
• Score function
• where c(i) is the confidence value for expert i
  and w(i) is the weight of the expert i

16
Experimental result - FDCM

• CBCL face database from MIT
• Training set (2429 face pattern, 4548 non-face
  pattern with 19x19 pixel)
• Testing set (472 face pattern, 23573 non-face
  pattern with 19x19 pixel)

Table 1 experimental results on images from the
testing set of CBCL database
17
Experimental result - FDCM

• To better represent the detectability of each
  model, ROC curve instead of single point of
  criterion response

Fig 8 The ROC curves of committee machine and
three different approaches
18
Experimental result - FRCM

• ORL Face Database
• 40 people
• 10 images/person
• Yale Face Database
• 15 people
• 11 images/person

19
Experimental result - FRCM

• ORL Face database

20
Experimental result - FRCM

• Yale Face Database

21
Conclusion and Future work

• We propose a heterogeneous committee machine
  approaches for face processing
• Face Detection Committee Machine (FDCM)
• Face Recognition Committee Machine (FRCM)
• Combine the state-of-the-art approaches
• Improve in accuracy and experimental results are
  satisfactory
• We have implemented a real-time face processing
  system
• Can detect and tracking the face automatically
• Work well for upright frontal face in varies
  lighting conditions
• We may use other biometric module such as
  fingerprint and hand geometry to improve the
  accuracy of the system

22

• Thank you!