Patchbased Gabor Fisher Classifier for Face Recognition

1
Patch-based Gabor Fisher Classifier for Face
Recognition

• Yu Su, Shiguang Shan, Xilin Chen, Wen Gao
• ICT-ISVISION Joint RD Lab for Face Recognition,
  Institute of Computing Technology, CAS, China

2
Outline

• Previous Works
• Motivation Basic Idea
• Proposed Method
• Experiments
• Summary

3
Previous Works

• Gabor based methods has become dominative methods
  in face recognition
• Gabor provides multi-scale, multi-orientation
  local features
• Typical Gabor based methods for face recognition
• Elastic Bunch Graph Matching (EBGM)
• Pre-assigned landmarks. People probably recognize
  face differently from machine
• It highly depends on the accuracy of landmark
  locations

L.Wiskott, J.M.Fellous, N.Kruger,
C.v.d.Malsburg, Face Recogniton by Elastic Bunch
Graph Matching, T-PAMI, 1997
4
Previous Works

• Gabor based methods has become dominative methods
  in face recognition
• Gabor provides multi-scale, multi-orientation
  local features
• Typical Gabor based methods for face recognition
• Elastic Bunch Graph Matching (EBGM)
• Gabor Fisher Classifier (GFC)
• Pre-assigned uniformed mesh
• Higher complexity

C. Liu and H. Wechsler, Gabor Feature Based
Classification Using the Enhanced Fisher Linear
Discriminant Model for Face Recognition, T-IP,
2002
5
Previous Works

• Gabor based methods has become dominative methods
  in face recognition
• Gabor provides multi-scale, multi-orientation
  local features
• Typical Gabor based methods for face recognition
• Elastic Bunch Graph Matching (EBGM)
• Gabor Fisher Classifier (GFC)
• AdaBoosted Gabor Fisher Classifier (AGFC)
• Using AdaBoost to select Gabor features
• Only those most discriminative Gabor features are
  kept for further classification

S.Shan, P.Yang, X.Chen, W.Gao, AdaBoost Gabor
Fisher Classifier for Face Recognition, Proc. of
AMFG 2005
6
Problems

• Holistic representation have three disadvantages
• Lost of the spatial information of features
• Suppose of the identical discriminative capacity
  of the features
• Too large image variation to be modeled by linear
  subspace

7
Outline

• Previous Works
• Motivation Basic Idea
• Proposed Method
• Experiments
• Summary

8
Motivation Our Solution

• Motivation
• Discriminative patches are more efficient than
  the whole face image
• Some trivial parts will decrease the
  discriminative capability of the feature
• Even more, some extrinsic variances provide wrong
  information for further classification
• Our solution
• Partition the face image, and use the
  discriminative patches only

9
Proposed Method

• Patch-based Gabor Fisher Classifier
• Partition face image into several discriminative
  patches by certain learning process.
• Get multiple Gabor feature segments by
  concatenating all the features within the
  corresponding patch.
• Design a FDA (Fisher Discriminant Analysis)
  classifier based on each feature segment.
• Combine all these component FDA classifiers.

10
Learning discriminating patches

• Use AdaBoost to select image patches
• AdaBoost is a powerful feature selection method.
• Convert multi-class problem to two-class problem.
• Intra-personal difference
• Extra-Personal difference
• Candidate patches set is formed by exhaustively
  enumerating the spatial position and patch size.
• Patches are considered as features which can be
  selected one by one from the candidate set.

11
Train and combine component GFCs

• Train multiple GFCs by applying FDA on each
  selected patch.
• Combine all the component GFCs by sum rule.

12
Outline

• Previous Works
• Motivation Basic Idea
• Proposed Method
• Experiments
• Summary

13
Experiments

• Proposed method (PGFC) vs. GFC AGFC
• Database
• FERET
• CAS-PEAL-R1
• Some parameters
• Image size 64 X 64
• Gabor feature 5 scales, 8 orientations
• Feature dimension
• GFC 10240 Gabor features
• AGFC about 3000 Gabor features
• PGFC up to 100 Gabor patches
• Note that the patch size is restricted within 32
  by 32 to avoid the problems of holistic
  representation.

14
Experiments on FERET

• Training set
• Standard FERET training set (1002 images from 429
  subjects) is used to learn the most discriminant
  patches and train FDA classifiers.
• Gallery
• 1196 subjects with only 1 image / subject
• Probe sets
• fafb (1195 subjects, 1 image / subject,
  expression)
• fafc (194 subjects, 1 image / subject, light)
• Dup1 (243 subjects, 722 images, aging, less than
  3 years )
• Dup2 (75 subjects, 234 images, aging, more than 1
  year later)

15
Performance comparisons
16
Performance comparisons (cont.)
17
Experiments on CAS-PEAL-R1

• Training set
• 1200 images from 300 subjects
• Gallery
• 1040 subjects, 1 image / subject
• Probe sets
• Accessory, 2646 images from 438 subjects
• Background, 650 images from 297 subjects
• Expression, 1884 images from 377 subjects
• Lighting, 2450 images from 233 subjects

18
Performance comparisons
19
Outline

• Previous Works
• Motivation Basic Idea
• Proposed Method
• Experiments
• Summary

20
Summary

• Proposed an effective patch-based method for face
  recognition
• The method overcomes some disadvantages in other
  Gabor-based methods
• Considering the spatial information of Gabor
  features
• Considering the different discriminative capacity
  of Gabor features
• reducing the variation of face which is difficult
  to be model by linear subspace methods such as
  FDA
• Experiments on two large databases FERET and
  CAS-PEAL-R1 show that the proposed algorithm is
  more effective than GFC and AGFC.

21
Thanks!