Face Recognition

1
Face Recognition

• Joshua I. Cohen

2
Overview

• Last Time 50-60 success rate
• Margin of error very small
• - 450 or less is a match
• - greater than 450 not a match

3
Industry Update

• Industry Update
• - Visionics Corporation (VSNX)

4
Industry Update Contd

• - Viisage Technology Inc. (VISG)

5
Data Acquisition

• Digital Camera
• - 150 images Amir, Brian, Josh, Kevin, Ming,
  Raghavan, Tom, Vinesh
• - 200 images unused
• - Final Analysis Amir, Josh, Kevin, Ming,
  Vinesh

6
Data Acquisition Contd

• Adobe PhotoShop Pre-Processing
• - color dodge where necessary
• - crop

7
Data Acquisition Contd

• Video Camera w/ IR Filtering
• - unsuccessful

8
Problems with Previous Algorithm

• Light Colored Shirts

9
Problems with Previous Algorithm

• Different Length Neckline - large cost
• - disturbs alignment

10
Problems with Previous Algorithm

• Hair Cropping

• Edit Cost
• Curve Matching Not Aligned

11
Finalized Algorithm

• Pre-Processing
• a) Convert To Gray Scale
• b) Binary Threshold
• c) Centroid
• d) Filter Noise Using Connected Regions
• e) Fill Region Up Until Contour
• f) Crop Image Area
• g) Normalize
• h) Fill Border To Black
• I) Save to .bmp File

12
Finalized Algorithm Contd

1. Convert .bmp to .pgm
2. Isolate Largest Contour
3. Profile Curve
4. Curve Matching
5. Compute Euclidean Costs
6. Parse Data to Extract Costs

13
Finalized Algorithm Contd

• Good Alignment - DPMatchVis

14
Finalized Algorithm Contd

• Computing Euclidean Cost
• 1) readMapFile
• 2) readRaphC
• 3) registerCorrespondence
• 4) rotateTranslateContour
• 5) euclidean distance

• plotRegisteredContours
• Mean, Median, Standard Deviation
• Higher Dimensional Analysis

15
Results Unedited
16
Results Unedited Contd

• Correct Match 22/37 or 59.46
• Without Tom 12/27 or 44.44

17
Results Unedited Contd

• Threshold 10 34/37 Correct, 64/361 False
  Positives

18
Results Nose
19
Results Nose Contd

• Correct Match 6/17 or 35.29

20
Results Nose Contd

• Threshold 12 - 13/17 Correct, 40/119 False
  Positives

21
Results Planar vs Planar
22
Results Planar vs Planar

• Correct Match 30/75 or 40
• Without Kevin 29/60 or 48.33

23
Results Planar vs Planar

• Threshold 9 71/75 Correct, 155/420 False
  Positives

24
Results H.D. Planar vs Planar 1

• Correct Match Across Rows 8/10 or 80
• Correct Match Down Columns 9/10 or 90

25
Results H.D. Planar vs Planar 1

• Threshold 8 9/10 Correct, 6/50 False Positives

26
Results H.D. Planar vs Planar 2

• Correct Match Across Rows 7/10 or 70
• Correct Match Down Columns 9/10 or 90

27
Results H.D. Planar vs Planar 2

• Threshold 6 9/10 Correct, 1/50 False Positives

28
Results H.D. Planar vs Non-Planar
29
Results H.D. Planar vs Non-Planar

• Correct Match Across Rows 1/10 or 10
• Correct Match Down Columns 4/10 or 40

30
Results H.D. Planar vs Non-Planar

• Threshold 8 6/10 Correct, 9/50 False Positives

31
Results H.D. Planar vs Planar Aberrations
32
Results H.D. Planar vs Planar Aberrations

• Correct Match Across Rows 4/16 or 25
• Correct Match Down Columns 6/16 or 38

33
Results H.D. Planar vs Planar Aberrations

• Threshold 9 9/16 Correct, 13/80 False Positives

34
Results Analysis

• Summary
• Planar vs Planar
• - Individual vs General Threshold
• When does the system fail?

35
Future Work

• Fiducial Point Extraction
• - Scale-Space Approach
• - Andrew Williss f2D utility from Sharp package

36
Future Work Contd

• Invariance occlusions, articulations,rotations,
  
• aberrations in facial
  expressions

• Tracking vs Recognition

37
References

• Ross Cutler, Face Recognition Using Infrared
  Images and Eigenfaces, April 1996.
• Thomas B. Sebastian, Philip N. Klein, Benjamin B.
  Kimia, Alignment-based Recognition of Shape
  Outlines, 2001.
• Zdravko Liposcak and Sven Loncaric, Face
  Recognition from Profiles Using Morphological
  Operators, 1998.
• Zdravko Liposcak and Sven Loncaric, A
  Scale-Space Approach to Face Recognition from
  Profiles, 1999.