A presentation about Palmprint Recognition

beeuppityΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

126 εμφανίσεις

Mustafa Berkay Yılmaz


Palmprint (what is)


Features


Acquisiton


Enhancement


Matching Methods


Classification


Databases


Comparison with Fingerprint


P
alms of human hands


C
ontain
s

unique pattern of ridges and

valleys


L
arger

th
a
n finger


E
xpected

to be even more reliable


More expensive scanners


Bigger area


H
ighly

accurate biometric system could be
combined

using hands


U
sing a high
-
resolution

scanner that would collect all
the features


H
and geometry, ridge and valley features, principal
lines, wrinkles
, even fingerprints, hand veins...


Chinese merchants

(
History
)


R
ecord
ed

palm prints and footprints on paper with ink

to
distinguish people



C
orrelate
d
with
medical



disorders

e.g. Genetic

disorders


and Downs

syndromes



Palmistry (chiromancy)


Palm reading
in order to reveal


the past, present and

future


High
Distinctiveness


High
Permanence (duration)


High
Performance


Medium
Collectabillity


Medium
Acceptability


Medium
Universality


Medium
Circumvention (fooling)


Principal lines and wrinkles:


Geometric features


width, length and area of palm


coarse measurement and relatively easily

duplicated


n
ot

sufficiently distinct


Line features


length, position,

depth and size of various lines


principal lines
:
not sufficiently distinctive


w
rinkles
:
distinctive
,
not easily duplicated


Point features or minutiae


similar to fingerprint minutiae


ridges, ridge endings, bifurcation and dots


d
atum

point: endpoints of principal lines


Reader Types:


Capacitive


Optical (generally used)


Ultrasound


Thermal


Low resolution readers (
<
100 dpi)


principal

lines and wrinkles


High resolution readers (
>
400 dpi)


also
point features and minutiae


Off
-
line


inked onto paper and later scanned


On
-
line


palm prints are directly scanned


Real
-
time


palm prints are scanned and processed in real
-
time.


C
hange

in scale


caused by increasing

or varying the distance between reader
and palm



C
apturing

a clear image of the hollow of the palm which
may not fully

contact the reader


provi
de

curved readers that

fully contact all parts of the palm



S
hifting

position, closing fingers or placing the hand


on different parts of the reader


design hollows for palm and fingers to occupy


provid
e

pins to separate and

locate the hand on the scanner



N
eed

to touch the hand reader


h
ygiene

and latent prints


Conventional enhancement techniques


Histogram equalization


Low
-
pass filtering


Phase Congruency [1]






[1]
Palmprint

Image Enhancement Using Phase Congruency
,
Yunyong

Punsawad

and
Yodchanan

Wongsawat
.
IEEE

International Conference on Robotics and
Biomimetics
,
2009.


Topological Inconsistency:


ridge/valley pattern has local not global consistency


Image Quality Discrepancy:


always poorer

q
uality

compared to
fingerprint
and
online palmprint


Image Size Difference:


always bigger compared to fingerprint


needs efficient algorithm


Method in [2]


Step 1:
Estimate and modify the orientation
field


Similar to fingerprint

[2] RESEARCH ON OFFLINE PALMPRINT IMAGE ENHANCEMENT. Yan Zheng, Guang
Shun Shi, Lin Zhang, Qing Ren Wang, Ya Jing Zhao. ICIP 2007.


Step 2:
Remove noises in a grey
-
scale image


filter image according to its orientation field


Ex. horizontal line


Filters of other
orientations are the rotations of
this filter



Step 3: B
inarization


based on local threshold


used to

convert grey
-
scale images into binary images


l
ocal

threshold is determined by Otsu threshold selection
algorithm


Step 4:
Noises removal in a binary image


morphological operators


M
inutiae
-
based matching


most widely used technique


location, direction, and
orientation of each point


higher recognition accuracy


performs poorly with low quality
images


does not take advantage of
textural or visual features


time consuming because of
minutiae extraction


C
orrelation
-
based matching


lin
e

up the palm images and subtract the
m


determine if the ridges in the two palm images correspond


less tolerant to elastic, rotational, and translational variances
and noise within the image


R
idge
-
based matching


ridge pattern landmark features such as sweat pores, spatial
attributes, and geometric characteristics of ridges, and/or local
texture analysis


f
aster

than minutiae


overcomes difficulties associated with extracting minutiae from
poor quality images


lower distinctiveness than the minutiae



Method in [3] (texture based)


Step 1: binarize the image using a threshold





Step 2:
Obtain the boundaries of the gaps


using a

boundary tracking algorithm



[3] On
-
Line Palmprint Identification. David Zhang, Wai
-
Kin
Kong, Jane You, Michael Wong. IEEE TRANSACTIONS ON
PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL.
25, NO. 9, SEPTEMBER 2003.


Step 3:
Compute the tangent of the two gaps







Step 4:
Line up (x1, y1) and (x2, y2) to get the
Y
-
axis of the
palmprint

coordinate system


Step 5:
Extract a sub
-
image of fixed size
based on the coordinate system






Use Gabor filter for feature extraction


Use hamming distance
for
palmprint

matching


Other alignment methods exist


Given ROI, c
alculat
e

moments
to
estimate

orientation

[4]






C
alculated

based on the whole

hand region rather than only
contour points

[4] Palmprint Recognition Under Unconstrained Scenes. Yufei Han, Zhenan Sun, Fei
Wang, and Tieniu Tan. Asian Conference on Computer Vision, 2007.


Alignment using homography [5]


Step 1:
Compute interest points in each image by
richly sampling points on the edge map


Step 2:
Use correlation

to compute the putative
correspondences


If

the number of correspondences is less than some
threshold, label it as imposter

[5] Pose Invariant Palmprint Recognition.
Chhaya

Methani

and
Anoop

M.
Namboodiri
.
International Conference on Biometrics (
ICB 2009
)


Step 3:
RANSAC based robust
Homography

estimation







Step 4:
Output the image obtained by

transforming
second i
mage with H


Alignment based on heart line [6]


H
eart

line is the most stable of the three principle
lines


always starts in the same region of the
palmprint


Use
2D
Morlet

wavelet transform

to capture the
high frequency responses where the intensity
values change abruptly

[6] On Latent Palmprint Matching.
Anil K. Jain and
Meltem

Demirkus
. MSU Technical
Report, 2008.


P
ixel
s
on the heart line provide high magnitudes
of at least one of the
Morlet

coefficients


Get
candidate heart line pixels
set
S


Apply RANSAC to fit a line to set S


Obtain
heart line start point and the angle


Make alignment using these information


Many other texture feature extraction methods, for example
DCT features are used [7]

[7]
Texture Based
Palmprint

Identification Using DCT Features
. Manisha P. Dale,
Madhuri A. Joshi, Neena Gilda.
Seventh International Conference on Advances in
Pattern Recognition
,
2009
.


Method in [8]


Step 1: Line feature extraction

[8]
Palmprint

Verification: An Implementation of Biometric Technology
. Wei Shu,
David Zhang. ICPR 1998.


Step 2: Line feature matching


Look at the
Euclidean distances between the


endpoints of
two
line segments


If distance between both ends are small


Two lines are considered as the same


Prior to recognition [9]


can greatly reduce
palmprint

matching

time for a
large database


C
lassify

palmprints

into ten

categories
,
examples:

[9] Palmprint Classification. Li Fang, Maylor K.H. Leung, Tejas Shikhare, Victor Chan, Kean Fatt
Choon. IEEE International Conference on Systems, Man, and Cybernetics. 2006.


Australia houses the largest repository of
palm prints in the world


The Australian

National Automated Fingerprint
Identification System (NAFIS) stores over 4.8
million palm

prints


PolyU
-
Palmprint
-
Database (The Hong Kong


Polytechnic University)


contains 7752 grayscale images corresponding to
386 different individuals


Larger surface


More features


Line features of palm are stable through life


Palm is less likely damaged


Low resolution images are enough


Faster preprocessing & feature extraction


Fingerprint

recognition state of art:


Verification EER: 0.1 %


http://www.eyenetwatch.com/access_control/bioscrypt
-
vflex
-
fingerprint
-
scanner.htm


Winner of FVC 2002


Palm
print

recognition state of art:


Verification EER: 0.07 %


Embedded Palmprint Recognition System on Mobile
Devices. Han, Y.F.[Yu
-
Fei], Tan, T.N.[Tie
-
Niu], Sun,
Z.N.[Zhe
-
Nan], Hao, Y.[Ying]. ICB (International
Conference on Biometrics) 2007.


[10]
BIOMETRIC TECHNOLOGIES
-

PALM
AND HAND
. Chris Roberts, May 2006.


[11] Palm Print Recognition, National Science
and Technology Council, 2006.


Thanks for Listening