Gender Classification Based on Feature Selection Using Genetic Algorithm

1
Gender Classification Based on Feature Selection
Using Genetic Algorithm

• Zhiming Liu
• Instructors Dr. George Bebis
• Dr. Sushil J. Louis

2
Outline

• Introduction
• Method
• Experimental results
• Conclusion and discussion

3
Introduction

• Problem
• Gender classification use the facial image to
  classify gender.
• Application A successful gender
  classification method can be applied in Human
  Computer Interaction (HCI) systems, passive
  surveillance and control in "smart buildings"
  (e.g., restricting access to certain areas based
  on gender) and collecting valuable demographics
  (e.g., the number of women entering a retail
  store on a given day).

4
Introduction

• Motivation
• Some researches have shown that the different
  eigenvectors (eigenfaces) encode the different
  information of object.

Fig.1. eigenfaces (from left to right) No. 1, 2,
11, 17, 23 and 62.
The first 2 eigenvectors encode light variation
information. The eigenvectors 11 and 17 look more
like female than other eigenvectors while
eigenvector 23 look like male because it
obviously has beard. The eigenvector 62 has the
glasses information.
5
Methodology

• Encoding
• Use the binary string to encode the gender
  eigenfaces. The entire eigenspace is divided into
  four sub-spaces
• Female eigenspace Ef 01
• Male eigenspace EM 10
• Common gender eigenspace Ec 11
• Other eigenfaces 00
• The entire eigenspace has 250 eigenfaces,
  the length of chromosome is 500.

6
Methodology

• Fitness function
• Compute the projection coefficients of three
  eigenspaces
• wf, wm and wc
• One SVM is trained to classify these
  coefficients.
• If the sample image is female, the input
  vector is formed in order
• wc wf wm
• If the sample image is male, the input
  vector is formed in order
• wc wm wf

7
Methodology

• Fitness function (cont.)
• Compute the residual errors between original
  images and the reconstructed images in gender
  eigenspaces.
• For the female sample Gf, compute the
  reconstructed image Ff using wf and wc and the
  reconstructed image Fm using wm and wc.
• And then compute the residual errors ef Gf
  Ff and em Gf Fm.
• For the male sample Gm, we use the similar
  method to compute the residual errors.

8
Methodology

• Fitness function (cont.)
• One SVM is trained to classify the residual
  errors.
• For the female sample, the input vector is
  ef em.
• For the male sample, the input vector is em
  ef.
• Up to now, we get two SVM classification
  results. Let Accuracycoeffi denote the accuracy
  of SVM using the projection coefficients as the
  input vectors, and Accuracyerr denote the
  accuracy of SVM using the residual errors as the
  input vectors.
• fitness 104 ( Accuracycoeffi
  Accuracyerr) 2Zeros

9
Experimental results

• Dataset
• The dataset contains 400 frontal facial images
  from 400 distinct people.
• Normalization size and intensity

Use a three-fold cross-validation procedure to
compute the average accuracy, by keeping 300
images to train, 50 images for validation and 50
images for testing.
10
Experimental results

• Parameters of GA
• crossover uniform, probability 0.9
• mutation multi-point, probability 0.1
• selection ranking
• population size 100
• generation 100

11
Experimental results

• Test results
• As mentioned in the definition of fitness
  function, there are two SVMs to evaluate the
  validation dataset. Therefore, a combination of
  two SVMs is used to decide the classification in
  test dataset, by integrating the outputs of two
  SVMs classifiers.
• When training SVMs, output -1 labels female,
  output 1 labels male. Using sum rule, we define
  the following integration strategy
• Ocomb Osvm_coeffi
  Osvm_err
• where Osvm_coeffi and Osvm_err are the
  outputs of two SVMs for each test image.

12
Experimental results

• Test results (cont.)

Osvm_coeffi -1, 1
Coefficients
Ef EM Ec
Projection

SVM
Test Image
S
Ocomb
Osvm_err -1, 1
Errors
Ef Ec
SVM
Reconstruction
Em Ec
If Ocomb is negative, the classification result
is female, otherwise is male.
13
Experimental results

• Test results (cont.)
• traditional method manually select
  eigenfaces
• TABLE 1
• Correct Rate Using Manually Selected Eigenfaces.

14
Experimental results

• Test results (cont.)
• TABLE 2
• Correct Rate Using The Eigenfaces Selected By GA.

15
Experimental results

• Test results (cont.)
• TABLE 3
• Distribution
  Number of The Eigenfaces Selected By GA.

16
Experimental results

• Test results (cont.)

Fig. 2. Reconstructed images using the selected
eigenfaces. First row original images
Second row using the female eigenfaces
Third row using the male eigenfaces.
17
Conclusion and discussion

• Conclusion
• This project proposes a new method to classify
  gender using facial images. Comparing with the
  traditional method, the accuracy of proposed
  method is better.
• Present problem the reconstruction results
  show that some eigenfaces encode not only female
  information but also male information. The used
  encoding scheme does not seem to be the best to
  encode these information.
• Future work
• Use real encoding might overcome the present
  problem.

18
Thanks!