3D Palmprint Verification Project

1
3-D Palmprint Verification Project

• Chhaya Methani

2
Aim of the project

• To develop a technique for access control systems
  using palmprints as signature for civilian based
  applications such that it -
• Handles images taken in natural settings and a
  variable background if possible which also lets
  us use an image capturing model that is simple to
  design ,easy to use and is probably unaffected by
  changes in the surrounding environment
• Allows user some flexibility with respect to the
  desired positioning of the palm by doing away
  with the traditional setup of fixing the hand in
  a setup having pegs etc
• Has an efficiency comparable to other techniques
  in use currently
• Is intuitive to use

3

• Approaches tried so far and problems with them

4
Camera calibration using Dynamic Silhouettes

• Idea To calibrate a network of cameras using
  dynamic silhouettes as defined in Camera Network
  Calibration from Dynamic Silhouettes by Sudipta
  Sinha
• Modeling our problem In our setup, the camera is
  stationary and the hand moves. Using principles
  of relative motion, if we consider that the
  camera has different orientations, we should be
  able to find out the orientation of the hand by
  calibrating the camera.
• Problems
• It needs multiple frames of the object since it
  uses a RANSAC based approach to refine the error,
  while we have only a limited number of images of
  the user
• It is difficult to repeat the process for every
  new user in an online recognition system

5
Recognition of planar shapes in multiple views by
considering shape primitives

• Idea A generalized approach in multiview pattern
  recognition for planar shapes having projective
  deformations based on shape properties
  multiview relations as described in Fourier
  Domain Representation Of Planar Curves for
  recognition in multiple views by Sujith k, Dr.
  C.V. Jawahar Dr. PJN
• Modeling our problem A palm can be approximated
  as planar surface for all practical applications.
  And hence different orientations can be probably
  modeled using the above mentioned techniques
• Problems
• When viewed from different angles, the surface of
  the palm can still be approximated as a plane,
  but the boundary of the shape doesnt essentially
  remain planar because of the thickness of the
  hand.

6
Visual Servoing Based Techniques

• Idea A technique to use visual features to
  control a robot. Enables alignment of an object
  with a manipulator.
• Modeling our problem We are given two different
  orientations of the hand. We need to change the
  camera position from one orientation to another
  so as to reach the orientation from which the
  other image was taken.
• Problems
• Visual Servoing techniques use rigidity of an
  object as an inherent assumption, which does not
  hold true in the case of a palm.

7
Applying homography using a set of corresponding
points

• Idea Two planar surfaces can be related using a
  homography matrix which can be computed using 4
  or more correspondences
• Modeling our problem A hand can be approximated
  to a planar surface. And corresponding points on
  the fingers etc can be easily found out
• Results of the algorithm On applying the
  approach to a set of real data, inconsistent
  results were observed. Some images showed an
  improvement in matching score, while on the other
  hand, some images showed degradation.

8
Results of applying Homography

• Example1
• Matching Scores
• Before applying
• Homography0.14
• After applying
• Homography0.11

9
Contd

• Example2
• Matching Scores
• Before applying
• Homography0.15
• After applying
• Homography0.16

10
Possible problems with using Homography and their
probable solutions

• The corresponding points that we obtain might end
  up being collinear since they lie on the
  periphery of the palm as shown in this figure.
  This can be solved by trying to capture more
  corresponding points.
• The matching technique might not be suitable for
  the approach. Other matching techniques need to
  be looked into.
• Homography is very sensitive to the noise in the
  data. If corresponding points have matching
  errors, it can result in poor performance. This
  still remains a problem.

11
Common principle in all the above approaches

• Till now, we have been looking into ways to
  determine the change in orientation between two
  images or to transform one image to closely match
  the other images orientation.
• This essentially tries to undo the differences in
  rotation, translation and makes scale changes so
  as both the extracted palms have the same size.
• Major problem with this kind of an approach is
  the change in scale.
• What if the palms are actually very different in
  size? For eg. One palm belongs to a child, and
  another to an adult with large hands.

12
A different angle

• Use the original set of images instead of
  transforming them.
• Extract different sized palm images based on the
  size of the hand and then apply appropriate
  matching technique.
• Same underlying assumption that the palm has been
  rotated only to such an extent that atleast the
  palm lines are clearly visible, should suffice.
• The complete algorithm needs to be modified to
  accommodate this change.
• Step-by-step analysis of the matching mechanism
  to follow.

13
Stages in the Palmprint matching mechanism

• Dataset Collection (to be discussed in the last)
• Preprocessing steps
• Extraction of the region of Interest from the
  preprocessed image (palm)
• Choice of appropriate sub-windows for feature
  extraction
• Choice of the feature to be extracted and the
  corresponding Matching Technique.

14
Preprocessing Steps

• Normalizing the image
• Appropriate thresholding techniques (depends on
  dataset capture)
• Reference Point Extraction
• Approach being followed so far
• Using PCA to align hand
• Computing the reference points after computing
  the boundaries of fingers using geometrical
  properties of the hand as shown previously
• Possible Alternative Using hands curvature
  based properties to determine the required points
  as shown.
• Not sure how well it responds to occlusions.

15
Extraction of the Region of Interest

• So far, we extract a square palm of fixed size
  from all images.
• Now, we extract a square palm image of variable
  side length depending on the distance between the
  two extreme fingers.

16
Choice of appropriate sub-window for Feature
Extraction

• The extracted region is usually divided into
  fixed sized squares with some overlapping part.
• In the changed approach, with different sized
  palm windows, elliptical haar like features need
  to be considered.
• The sub-window is 1/3rd of a fixed size
  concentric chain of half ellipse, and the
  features can be marked as shown below.

17
Choice of feature to be extracted and the
corresponding Matching Mechanism

• There can be following types of features
• Feature vector consists of the entire length of
  Image which is filtered with directional filters
  and then matched pixel by pixel using logical
  operations.
• A feature vector consisting of the mean and
  variance of the gray level of all the pixels in
  every sub-window. The length of the feature
  vector in such a case is the no. of sub-windows.
• Extraction of line features and creases etc.
• Statistical signatures
• Feature vector having moments of various orders
• DCT (Discrete Cosine Transform)
• Wavelet based transformation to be applied and
  then the following signatures can be used for
  each sub-window
• The average gravity center signature
• The density signature
• The spatial dispersivity signature

18
How to compare feature vectors with variable
length

• Score(i,j) ? feature Vi(n) feature Vj(n)
• min(Ni,Nj)
• where-
• Vi feature vector i
• Vj feature vector j
• Ni length of feature vector i
• Nj lebgth of feature vector j
• n goes from 1 to min(Ni,Nj)

19
Database Collection

• Dataset can be collected in the follwoing two
  situations
• Moderately controlled conditions
• Completely Natural or Uncontrolled conditions
• Both have their own merits and demerits.

20
Dataset collection with moderately controlled
conditions

• Such conditions can be useful as they help give
  some control on the algorithm and on the other
  hand, can be difficult to implement for practical
  systems
• Such a dataset has been collected under the
  following moderated conditions
• Background was a black sheet
• Partially controlled illumination settings, as in
  effect of day and night create a difference in
  the illumination effects
• Resolution of the image 1728 x 1152
• Total images 760 (76 users, each giving 5 left
  hand and 5 right hand images)
• Problems with thresholding because of the noise
  effect created by the background lighting
  conditions
• Possible solution Applying an advanced level
  binarization algorithm like
• 2D Otsu Thresholding of gray scale images
• Hue histogram based thresholding of color images
  etc

21
Dataset collected under natural conditions

• Problem acquiring such a dataset
• Needs training a cascaded classifier for which a
  database of positive examples containing
  synthetic images generated in the required
  parameter space is needed
• OpenCV implementation of cascaded classifiers
  return an approximate bounding box, and not the
  boundary as has been extracted in the paper and
  which is more useful as such
• Increases computational complexity

22
Finally

• What dataset to be used?
• Should both the approaches be implemented and
  compared?

23
Future Work

• Implementing and comparing results
• Integrate Elastic Matching

24

• Thank You