Pose Invariant Palmprint

Recognition
Chhaya Methani and Anoop Namboodiri
Centre for Visual Information Technology
IIIT, Hyderabad, INDIA
IIIT Hyderabad
 Palmprint Aquisition
• Controlled pose, scale, and
illumination
• High accuracy

• Fixed Scanner/Camera
• Restricted Palm position
• Palmprint-Specific
• Can we use a generic
camera as the
acquisition device?
IIIT Hyderabad
 Unrestricted Palmprint Imaging
• Minimal Constraints
• Intuitive, user friendly
• New applications
• Multibiometric sensor
IIIT Hyderabad
 Challenges
• Background
• Illumination
• Contrast
• Noise
• Pose
• Scale
IIIT Hyderabad
 Previous Work
• Background
– Skin Color
– Hand Shape
• Illumination
– Normalize
Stenger et al. “Model-Based Hand Tracking Using a Hierarchical
• Noise Bayesian Filter”, TPAMI 28(9), Sept. 2006
Shadow,Wrinkles,
Pixel Noise.
– Good features
• Scale
IIIT Hyderabad

JDoublet, et al. “Contactless hand Recognition Using Shape and

Texture Features”, ICSP 2006
 Variations in Pose
• Induce perspective line distortions
• Associated with scale changes
• Performance degradation EER: ~22%
• Dataset: 100 palms, 5 images per palm.

• Solution Directions:
1. Compute Pose-Invariant Features
2. Correct Pose variations
• Non-rigid transformations are difficult
to model
• Assumption of planarity
IIIT Hyderabad
 Invariance to Perspective Projection
• Cross Ratio, defined by 5
coplanar points
• Assume a stretched out
palm to be planar
• Sensitive to point position
• Need reliable point
detection
• Point matches found
using SIFT
IIIT Hyderabad

• Zheng, Wang and Boult : “ Application of Projective Invariants in Hand Geometry Biometrics”,
IEEE Transactions on Information Forensics and Security, 2007.
 Finding Pose Transformation
Parameters
• Palm considered a planar surface.
• Homography defines transformation parameters
between 2 planes given 4 point correspondences
are known.
–

– Where x'/c and y'/c is the resulting point.

• 4 distinctive point correspondences needed.
IIIT Hyderabad
 Solution using Interest Points
• We use a combination of stable
points and a set of interest
points as candidate matches.
• Stable/Valley points are the
consistent points.
• 4 valley points available.
• Only 2 can be used. Valley Points
• Rest of the points must be
selected from the palm lines.
• Thus, we choose a bag of
candidate interest points.
• These points are refined later to
get reliable interest points.
IIIT Hyderabad
 Proposed Solution

Image
Image Preprocessing Image Feature
& Alignment Matching
Acquisition Extraction
Palm Extraction
IIIT Hyderabad
 Image Acquisition
• Fixed Camera and Background
• Flexible Palm pose
and position
• Natural Illumination
variations
• Sample Image

Image
Image Preprocessing Image Feature
Matching
Acquisition & Alignment Extraction
IIIT Hyderabad

Palm Extraction
 Image Preprocessing & Palm Extraction
• Finger valley points are used to extract ROI and correct
in-plane rotations

Image
Image Preprocessing Image Feature
Matching
Acquisition & Alignment Extraction
IIIT Hyderabad

Palm Extraction
 Proposed Solution – Image Alignment
• Assumption of planarity of the palm surface
• Homography can be used to estimate pose
• 4 distinct point correspondences needed.
• Use a combination of stable
points and interest points
Valley Points

• Other interest points?

Back to the same problem!

Image
Image Preprocessing Image Feature
IIIT Hyderabad

Matching
Acquisition & Alignment Extraction
Palm Extraction
 Proposed Solution – Image Alignment
• Assuming
Descriptorsequal
Correspondences
Similar areprobability
process made
is found of occurrence
followed
using
usingfor
11x11 for all palm
correlation
the windows
second points
on the line,
around eacha richly
of thesampled pointinterest
candidate set is chosen
pointson the
palm line

Image
Image Preprocessing Image Feature
Matching
Acquisition & Alignment Extraction
IIIT Hyderabad

Palm Extraction
 Proposed Solution – Image Alignment

Input
The
• The
Final best
to of
final
set RANSAC
set of parameters
transformed
inliers based
in both theHomography:
image. found by
template andRANSAC
thethe
set 2image
are.
used for
valley points
the final
and transformation.
iterative selection of the other
two from the interest points.

Image
Image Preprocessing Image Feature
Matching
Acquisition & Alignment Extraction
IIIT Hyderabad

Palm Extraction
 Proposed Solution: Computing Features
and Matching
• Thresholded Gabor responses
dist(final) = min(dist(fixed), dist(corrected))

• D. Zhang, A. W. K. Kong, You, J., Wong M., “Online Palmprint Identification” , PAMI 2003.
Image
IIIT Hyderabad

Image Preprocessing Image Feature

Matching
Acquisition & Alignment Extraction
Palm Extraction
 Datasets

• 100 palms,
5 images per palm
• Pose variations up
to 45 degrees

• 50 palms, from
PolyU dataset
• 10 synthetic poses
per palm:
0 - 45 degrees
IIIT Hyderabad
 Results
• Comparison of EER values
Method Synthetic Data Real Data
0◦- 20◦ 20◦-30◦ 30◦-35◦ 35◦-40◦ 40◦-45◦
Fixed Pose 0.01% 3.24% 3.71% 16.93% 30.92% 22.07%
Approach
Blind Pose 16.48% 12.40% 11.14% 14.98% 11.92% 16.51%
Approach
Proposed 0.47% 4.19% 11.14% 14.98% 11.92% 8.71%
Approach
IIIT Hyderabad
 Results: Synthetic Data
IIIT Hyderabad
 Results: Real Data
IIIT Hyderabad
 Results
•• (q)
(r)
(p) :The
: There
GAR
Semilog drop
low
iscurve
aofeven
sharp
GAR with
drop
toinobserve
case
high
in the
of
FAR.
proposed
GAR. approach is earlier.
the highlighted data.
• Indicates imposter
genuine pairspairswith
withlow
increased
highsimilarity.
similarity.
similarity.
• Reasons: Inherent
Blur, wrinkles,
Pixel saturation,
in the algorithm.
etc.
specular reflections of skin etc.
IIIT Hyderabad
 Video Based Palmprint Recognition
Base Image Base Image + 2 Base Image + 6 Base image + 10

12.46% 10.92% 9.83% 7.87%

• Successive addition of Gabor responses.

• Images shown after adding 2, 6 and 10 images respectively.
IIIT Hyderabad
 Conclusion/Observation
• Proposed view invariant recognition system for Palmprint.
• Very difficult to find point correspondences for palm.
• Solution using point correspondence of stable and interest
points.
• RANSAC based Homography used to choose from approximate
point correspondences.
• Major role played by illumination variations and noise.
• Video based palmprint recognition is a possible solution.
• Future Work: To study the effects of video based palmprint
recognition in further in more detail.
IIIT Hyderabad
IIIT Hyderabad

Thank You