Face Image Based Gender Classification using Minmax Modular Classifier

1
???????????????????????? Face Image Based
Gender Classification using Min-max Modular
Classifier
????? ? ?
? ???? ?? ? ? ?
????????? ? ? ? ???????
?????? ?????????
2
Outline

• Introduction to gender classification problem
• Face image preprocessing and feature extraction
  for gender information
• Min-max modular classifier with different task
  decomposition strategies for gender
  classification
• Experiments and discussion
• Conclusion and future works

3
Introduction

• History of Gender classification problem
• Gender classification based on methods from face
  recognition problem
• Feature Subspace, Gabor, EGM, Texture, Genetic
  features
• Classifier Neural Network, SVM, M3-SVM
• Specialty of Gender classification
• Binary class problem
• High generalization ability
• Large-scale data required

4
Previous Work
5
Motivation of thesis

• Objectives
• Analyze generalization ability systematically on
  more complex face images
• Improve classification performance on large-scale
  database
• Solutions
• Enrich features extracted from face images to
  enhance representation ability
• Assemble high-performance modular classifiers to
  solve large-scale classification

6
Outline

• Introduction to gender classification problem
• Face image preprocessing and feature extraction
  for gender information
• Min-max modular classifier with different task
  decomposition strategies for gender
  classification
• Experiments and discussion
• Conclusion and future works

7
Face Image Preprocessing

• Convert color images into gray-scale
• Easier and more efficient to process
• Geometric normalization
• Scale conversion, translation, rotation
• Histogram equalization
• Decrease the illumination effects
• Warping and Masking
• Erase background clutter

8
Face Image Preprocessing (cont.)
9
Feature Extraction

• Special requirements of gender information
• Extract features from single image
• Synthesize features from a database
• Feature categories for gender information
• Holistic features Gary-scale, Gabor wavelet
  filter (Wiskott, 1997)
• Local features Local Binary Pattern (LBP)
  (Ojala,1996)
• All are general models without selection
• A new Embedded SIFT features
• Scale invariant feature transform (Lowe,1999)
• From Person-specific to an uniform model

10
Our new feature extraction method- embedded SIFT
feature
Feature Set Synthesis
Key-points Detection
Image Set
Male features
Feature Description
Female features
11
SIFT key-points detection
Detect maxima and minima of Difference of
Gaussian in scale space
12
Feature set synthesis
13
Key-point description

• Key-point descriptor
• Thresholded image gradients are sampled over
  neighboring array of locations
• Create array of orientation histograms

14
Outline

• Introduction to gender classification problem
• Face image preprocessing and feature extraction
  for gender information
• Min-max modular classifier with different task
  decomposition strategies for gender
  classification
• Experiments and discussion
• Conclusion and future works

15
(Lu and Ito, 1997, 1999)
Independent Subproblems
Task Decomposition
Solution
MIN
MAX
MassivelyParallel Learning
MIN
Module Combination
16
Task decomposition method of Min-max modular SVM

• Random partition method (Lu, 1999)
• Divide the data set randomly
• Hyper-plane method (Wang, 2005)
• Divide the data set using a hyper-plane
• Prior knowledge method (Lian, 2005)
• Divide the data set with prior knowledge
• Equal clustering method (Wen, 2005)
• Divide the data set with clustering method

17
Merits and Demerits of Existing Task
Decomposition Strategies

• Random partition is simple and direct but not
  stable.
• Hyper-plane partition is useful for sparse data
  mostly.
• Prior knowledge is the most effective but not
  available in many cases.
• Equal Clustering concentrates too much on
  balancing sub-problems.

18
New task decomposition strategy based on spectral
clustering

• Spectral clustering
• The data set in an arbitrary feature space is
  represented as a weighted undirected graph
• Conduct spectral graph partitioning under certain
  constrains
• Advantage
• Capture the intrinsic data distribution as much
  as possible

19
Spectral Clustering Algorithm

• Step1 Construct affinity matrix of data points
  with each element denoting the similarity between
  two samples.
• Step2 Transform the affinity matrix into a
  Laplacian matrix and compute the eigenvalues and
  corresponding eigenvectors.
• Step3 Normalized the eigenvectors and then
  conduct a normal clustering method (K-means, etc)
  on the space constructed by eigenvectors.
• Step4 Assign the cluster labels to data points
  according to the clustering results from Step3.

20
Decomposition examples on toy data set
21
Decomposition examples on toy data set (cont.)
Spectral K-means
Equal Clustering
Spectral Equal Clustering
Random decomposition
22
Outline

• Introduction to gender classification problem
• Face image preprocessing and feature extraction
  for gender information
• Min-max modular classifier with different task
  decomposition strategies for gender
  classification
• Experiments and discussion
• Conclusion and future works

23
Experiment Setup

• Face Database
• CAS-PEAL
• 595 males and 445 females
• Different variations including Pose, Expression,
  Illumination, Accessory and etc.

24
Experiment-1

• Objective
• Evaluate the proposed Embedded SIFT feature
• Compare with LBP and Gabor filter features
• Setup
• Training 400 male and female frontal images
• Testing SVM with Linear, Poly3 and RBF kernel

Note all probe images are randomly chosen
25
Experiment-1 (cont.)
26
Experiment-2

• Objectives
• Evaluate the Spectral clustering method on task
  decomposition for Min-max modular SVM
• Compare with other strategies
• Setup
• Training 3600 images
• Testing 8746 images of 9 kinds of poses ranged
  from -30o to 30o
• Using Gray-scale vectors as features

27
(No Transcript)
28
Experiment-2 (cont.)
Experiment results of M3-SVM using different
stategies
29
Experiment-3

• Objective
• Assemble SIFT features with M3-SVM
• Setup
• Data similar as in Experiment-1
• SVM is using RBF kernel with C128, s0.125
• Training set is divided into 3,6,9 parts
  respectively for M3-SVM under different task
  decomposition strategies

30
Experiment-3 (cont.)
31
Discussion

• Experiment-1
• The proposed embedded SIFT features outperform
  other traditional effective features
• The new features are invariant to pose,
  expression, individual and accessory variations
  but not so effective on illumination variations
• Experiment-2
• The proposed Spectral clustering strategy can
  improve the performance of M3-SVM
• The strategy also shorten the training time
• However, the extra matrix construction increases
  the preprocessing time complexity
• Experiment-3
• The combination of the new feature and
  classification indeed outperforms the traditional
  methods for gender classification

32
Outline

• Introduction to gender classification problem
• Face image preprocessing and feature extraction
  for gender information
• Min-max modular classifier with different task
  decomposition strategies for gender
  classification
• Experiments and discussion
• Conclusion and future works

33
Conclusion and future work

• The contributions of the thesis
• Propose a new feature for representing gender
  information of face images based on SIFT features
• Explore the spectral clustering algorithm on task
  decomposition for M3-SVM
• What to do next
• Increase the stability of new features on
  illumination effects
• Improve the computation cost of task
  decomposition of M3-SVM

34
Publications

• Jun Luo and Bao-Liang Lu, Gender Recognition
  Using Min-Max Modular Support Vector Machine with
  Equal Clustering, Proceedings. of International
  Symposium on Neural Network (ISNN 2006), Lecture
  Notes in Computer Science, Springer, vol. 3972,
  pp.210-215, Chengdu, 2006
• Jun Luo, Yong Ma, Erina Takikawa, Shihong Lao,
  Masato Kawade, and Bao-Liang Lu, Person-Specific
  SIFT Features for Face Recognition, Accepted by
  International Conference of Acoustics, Speech and
  Signal Processing (ICASSP 2007), Hawaii, USA, 2007

35
Thank You!