Partial

1
Partial Holistic Face Recognition on FRGC-II
data using Support Vector Machine Kernel
Correlation Analysis a brief summary

• M. Savvides, R. Abiantun, J. Heo, S. Park, C.
  Xie, and B.V.K Vijayakumar

2
Presentation Outline

• Background Material
• Correlation Filters ? MACE ? Kernel Correlation
  Filters
• Support Vector Machines
• Partial vs. Holistic Face Recognition
• Results

3
Correlation Pattern Recognition

• Matched Spatial Filter (MSF)
• Uses frequency domain
• Optimal in noise
• Spatial freq. domain proven in (ATR)
• No need for image segmentation
• Closed form expression
• - Need new classifiers for multiple orientations

4
Synthetic Discriminant FunctionHester and
Casasent (1980)

• Uses multi-class objects (same object from
  different views) as reference functions
• System of linear equations determines an Average
  Filter
• Constraint shift invariance
• Relative peak values used to detect object
• - Sidelobes can lead to misclassifications (due
  to constraints it only controls values near the
  origin)

5
(No Transcript)
6
Peek Sidelobe Ratio
7
Min. Avg. Corr. Energy (MACE)Mahalanobis, Kumar,
Casasent (1987)

• Minimizes sidelobe energy (fewer
  misclassifications)
• Kallman (optimum) is too computation intensive
• With only a few training images, outperforms most
  others (in less time)
• Can be shown it is Preprocessing Invariant

8
(No Transcript)
9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
Correlation Pattern Recognition - Cambridge Press
2005
13
(No Transcript)
14
(No Transcript)
15
(No Transcript)
16
Class Dependant Feature Analysis (CFA)

• One Filter designed for each of the 222 people in
  the training set
• One against all 12,776 images
• Closed form solution to produce a correlation of
  1 for authentic class 0 for everyone else

17
(No Transcript)
18
In general linear subspace methods e.g. LDA, PCA,
are extended into higher dimensions to compensate
or non-linear distortions.Here they use the
Kernel Trick
19
Kernel Trick

• Uses inner products of functions that map to
  higher dimensions
• Do not actually have to compute the higher
  dimensional mappings

20
Kernel Correlation Filters

• Now use the 222 Kernel versions of the CFA
  filters to get KCFA filters
• Nearly three fold increase over other linear
  methods

21
Support Vector Machines

• Instead of training the SVM on the images
  (directly) they are trained on the KCFA
  coeffiecients

22
(No Transcript)
23
Final Holistic Results
24
Partial vs. Holistic Face Recognition

• Experiments using the above techniques were done
  on three regions of the face
• Mouth region
• Nose region
• Eyes and eye-brow region

25
(No Transcript)
26
ROC Curves by Region
27
(No Transcript)
28
(No Transcript)
29
(No Transcript)
30
Conclusions

• MACE filters are less dependant on alignment
• KCFA allows working in lower dimensional space
• SVMs offer a general improvement over other
  similarity metrics and can be trained on lower
  dimensional data
• Fusion based techniques show improved results in
  Correlation Filters as well