Enhancing the Performance of Face Recognition Systems

1
Enhancing the Performanceof Face Recognition
Systems

• Presenter Dr. Christine Podilchuk
• Professors Richard Mammone, Joe Wilder
• Students Anand Doshi, Aparna Krishnamoorthy,
  Robert Utama
• WISE Lab, CAIP Center
• http//www.caip.rutgers.edu/wiselab

2
Project Description

• Funded by Dept of Defense, Technical Support
  Working Group (TSWG)
• Scope of Work Preprocessing technology to
  improve existing state-of-the-art face
  recognition systems
• - commercial system provided by Viisage
  (technology from MIT, Media Lab)
• - Rutgers technology
• Problems addressed blur and illumination
  correction

3
Problem
Preprocessing for Face Recognition

• Current state-of-the-art face recognition
  systems degrade significantly in performance due
  to variations in illumination and blurring

Solution
PREPROCESSING RESTORATION/ ENHANCEMENT
FACE RECOGNITION SYSTEM
IMAGE CAPTURE
DEBLURRING (due to mismatch in camera resolution,
image scale, and motion blur) ILLUMINATION
CORRECTION (due to mismatch in lighting
conditions in both indoor and outdoor
environments)
4
Preprocessing for Face Recognition
Solution

• Projection onto Convex Sets (POCS) framework
• A priori knowledge of the blur, illumination
  and/or face can be incorporated into the POCS
  framework
• Deblurring and illumination correction processes
  are duals of each other
• - the deblurring process operates in the Fourier
  domain
• - the illumination correction operates in the
  spatial domain

5
Resolution Enhancement
Problem recognition performance drops when
image resolution of training and testing images
vary.
Training image
Testing image Same resolution EER 8
Testing image Lower resolution EER 23
6
Resolution Enhancement
7
Illumination Correction
Enrollment Failure (no preprocessing) 44
Training image A
Testing image B
Enrollment Failure (with preprocessing) 10
Preprocessed Image B
8
Future Work

• Improve algorithms for deblurring and
  illumination correction
• Test algorithms on additional databases (varying
  cameras, resolutions, viewing angles, lighting
  conditions)
• Devise models of convex sets for faces, blur
  models and illumination models
• Generate ROC curves for performance before and
  after preprocessing
• Test our preprocessing algorithms on commercially
  available systems
• For current updates, visit http//caip.rutgers.edu
  /wiselab