Face Detection, Recognition and Reconstruction using Eigenfaces

1
Face Detection, Recognition and Reconstruction
using Eigenfaces

• By
• Muhammad Faisal Azeem

2
Objective

• Creating a Face Space using Eigen Faces as
  its Basis
• Projecting Faces on this Face Space to obtain
  Eigenface components
• Detection and identification of human faces using
  these components
• Face Reconstruction using Eigen Faces
• Differentiating Between Face images and other
  images

3
Eigenfaces

• Principal components of the distribution of
  faces, Eigenvectors of covariance matrix of a set
  of images
• Use to encode characteristics of faces which
  contain maximum information (variation)
• Faces correspond to points in Face space
• Treat face as a 2-D pattern and do pattern
  recognition
• Face images approximated as weighed sum of
  Eigenfaces
• Using these weights as a comparison criteria for
  face recognition

4
Eigenfaces..

• 120128 image point in 15360 dim space
  Vector of length 15360
• Images of faces of 16 people, 18 images per
  person
• Training sets of 8 or 16 images
• C A x AT, A F1 F2 . . . FM, Fi Gi - ?,
  (Avg(Gi))
• Uk, eigenvectors of C, each corresponding to some
  eigenvalue

5
Face Recognition

• 8 Eigenfaces obtained from a training set of 8
  images.
• Test set consists of 16 images (8 of them are the
  same persons as in training set)
• Finding weights of Eigenfaces and using Euclidean
  distance to classify images
• Images far away from training sets treated as
  unrecognized images

6
Issues

• Trying to recognize faces under variation of
  lighting, scaling and head tilt
• For Different Head tilts ended up having separate
  training sets and database
• Scaling proved to be a harder problem, pretty bad
  results for scaling
• Lighting results turned out to be the best
• Initially some problems with equations

7
Example Eigenfaces
First Five Eigenfaces and Average Image
8
Example Eigenfaces ...
First Four Eigenfaces and Average Image for
tilted head training set
9
Experimental Results

• For upright heeds, 100 correct face
  recognition, 87.5 face recognition
• For right head tilt head, 86 correct face
  recognition, 87 face recognition
• For right head tilt head, 83 correct face
  recognition, 75 face recognition
• Better face recognition if willing to trade off
  on accuracy

10
Face Reconstruction

• 15 Eigenfaces obtained from a database of 16
  images.
• Removing some image information and projecting
  the images on the face space.
• Recognizing images as one of images in the
  database
• Using recognized image to reconstruct images

11
Experimental Results

• Accurate when data lose is small

12
Face Detection

• Differentiating Between face images and other
  images
• To check whether image is a face, distance
  between image and its projection is used
• If distance less than threshold ?d image is a
  face otherwise not a face
• Also part of face recognition algorithm

13
Face Detection . . .

• Examples of face images and other images

Recognized as Face. Was used in training
Recognized as Face. Not used in training
Recognized as not a Face Of course not used in
training
14
Future Work

• Make a better system for recognition of
  variations in scaling
• A better classification of head tilt so we dont
  have search through all databases
• A bigger training set to capture more variations

15
Credits

• Images taken from the Vision and Modeling Group
  at the MIT Media Lab
• Research Papers by M. Turk and A. Pentland used
  as reference