SUBJECT SPECIFIC FACE RECOGNITION

1
SUBJECT SPECIFIC FACE RECOGNITION

• J.W.H. Tangelder (CWI), M. Bicego (UNISS) D.
  González Jiménez (UVIGO)
• J.L. Alba-Castro (UVIGO), E. Grosso(UNISS) M.
  Tistarelli(UNISS), B.A.M. Schouten(CWI)

2
Overview

• Face recognition using most dissimilar patches
  (UNISS)
• Face recognition using most discriminative Gabor
  features (UVIGO)
• Authentication by Bayesian fusion of subject
  specific face descriptors (CWI)
• Future research directions

3
Face recognition using most dissimilar patches

• Key concepts
• Subject specific face recognition
• Face distinctiveness
• Subject specific face recognition
• A template should be tailored to the subject
  peculiarities
• Face distinctiveness
• Use as template the most distinctive parts of
  the subjects face

4
Computing most distinctive parts

• Distinctiveness is related to saliency
• But.
• Saliency is an unary operation
• Tsotsos et al. 95Lindeberg 93Salah et al.
  02Lowe 04
• Distinctiveness is related to a n-ary operation
• Finding differences between the client face and
  all other faces

5
Computing most distinctive parts

• Basic scheme
• Binary operator finding differences between a
  pair of faces
• Face sampling, multiscale patch extraction and
  description
• Projection in the feature spaces and weighting

6
Computing most distinctive parts

• Face sampling, multiscale patch extraction and
  description
• Sampling points are the edges of the face
  random points
• Logpolar patches are extracted at each point

7

• Log polar patches
• Multiresolution information
• they describe the mapping postulated to occur
  between the retina and the visual cortex
• Grosso Tistarelli 2000

8

• Projection in the feature space and weighting

The weight of a patch is proportional to the
distance to the other set
9
Generalization to many to many

• Simple generalization
• Extract and project all client faces patches
• Extract and project all impostor faces patches
• Compute most distinctive client patches
• Patches which are very different from all other
  impostor patches (the rest of the world) are the
  most distinctive
• Obviously it depends on the size of the impostor
  set

10
Basic experimental evaluation an impairment
test

• Idea given a simple authentication scheme
  (Euclidean distance between images)
• Compare
• Method 1 delete from the images the K most
  important patches (of the claimed identity)
• Method 2 delete the same amount of random points
• Look for the worst behaviour

11
Results

• Banca database protocol MC
• Result deleting the most distinctive points is
  really deleterious for autentication
• Next step design a direct face authentication
  methodology

12
Face recognition using most discriminative Gabor
features

• Gabor features selection is splitted into 2
  stages
• First Layer A set of points are selected by
  exploiting facial structure, and Gabor features
  are computed.
• Second Layer A subset of the initial group of
  points are preserved based on their respective
  Gabor features accuracy

13
First Layer Shape-driven location

• Ridges Valleys detector
• Sampling points from lines depicting facial
  structure
• Preliminary set of points for client C

14
Extracting texture

• 40 Gabor filters are used
• A feature vector (Gabor jet) is extracted at each
  shape-driven point
• Set of jets

Jpk
15
Second Layer Accuracy-based selection (I)

• Goal Given a set of
• Available images for client C
• Available impostor images

Find the best subset of nodes for client C
.....
.....
16
Second Layer Accuracy-based selection (II)

• Each jet is evaluated as an individual classifier
• Only the nodes whose jets are good at
  discriminating between client C and impostors are
  preserved.
• Finally, for client C, we get

17
Results

• Banca database protocol P.

ERROR RATES ()
18

Authentication by Bayesian Fusion of Subject
Specific Face Descriptors

• Idea
• Extract features at most discriminative face
  parts for both clients and impostors
• Model the distribution of features by both a
  Gaussian client model and a Gaussian impostor
  model
• Take an authentication decision by thresholding
  the sum of log-likelihood-ratios of the client
  and impostor

19
Client and impostor feature templates
Template of client feature models built on 5
locations using 5 client images
Template of impostor feature modes built on 5
locations using 5 world images
20
Gaussian client and impostor feature models
For each client c, for each location
we estimate the probability of feature
value x by a Gaussian client model of its
atypical client-specific feature
distribution And by a Gaussian non-client model
of its typical client-specific feature
distribution
21
Taking the authentication decision
For each location the log-likelihood ratio is
used to assign confidence that a feature value is
client-specific
Based on the weighted sum of these
log-likelihood-ratios over all reference
points the claim of a client c is accepted when
it exceeds a given threshold
reject
accept
22
Future research directions

• Generalizing methods developed for authentication
  to face recognition
• Fusion using both Gabor jets and log polar patch
  features
• Developing new approaches to compute the matching
  from the features in one image to the features in
  another image
• Applying a cascaded approach to face recognition
  Compare first, the most salient part of a face
  against other faces
• Applying methods to video (tracking landmarks,
  pose correction)