Fingerprint Feature Extraction & Matchingshodhganga.inflibnet.ac.in/bitstream/10603/9044/7/07... ·...
Transcript of Fingerprint Feature Extraction & Matchingshodhganga.inflibnet.ac.in/bitstream/10603/9044/7/07... ·...
80
Fingerprint Feature Extraction & Matching
As discussed previously (Section 2.1.1.2) mainly two types,
fingerprint recognition systems namely Minutiae based matching &
Correlation based matching are found in literature. A feature
extractor finds the ridge endings and ridge bifurcations from the
input fingerprint images. If ridges can be perfectly located in an
input fingerprint image, then minutiae extraction is just a trivial
task of extracting singular points in a thinned ridge map. In
practice, however, it is not always possible to obtain a perfect ridge
map.
Table 3.7
Fingerprint Samples Taken from Same User and Corresponding ROI
User Fingerprint 1 Fingerprint 2 Fingerprint 3 Fingerprint 4
1
Fingerprint
ROI
2
Fingerprint
ROI
The performance of currently available minutiae-extraction
algorithms depends heavily on the quality of input fingerprint
images. Due to a number of factors (aberrant formations of
81
(a)
(b)
Fig. 3.25. Enrollment of Fingerprint (a) Fingerprint Preprocessing & ROI
Extraction (b) Training Samples used for Enrolling Fingerprint in
Database
epidermal ridges of fingerprints, postnatal marks, occupational
marks, problems with acquisition devices, etc.); fingerprint images
may not always have well-defined ridge structures. Reliable
minutiae-extraction algorithms should not assume perfect ridge
structures and should degrade gracefully with the quality of
fingerprint images. We are focusing on correlation based
fingerprints which are robust but less accurate; this accuracy can be
increased by combining with another trait in a multimodal biometric
system. We extract spectral features of a fingerprint based on 2D
Image transforms and wavelets.
82
For feature extraction we are using following variants,
1. Kekre’s Wavelets based Feature Extraction.
2. Partitioned Complex Walsh Plane in Transform Domain. We
use Cal, Sal function plot in complex Walsh plane
3. Partitioned Complex Planes of Kekre’s Transform, Hartley
Transform, Discrete Cosine Transform, Kekre’s Wavelets. The
above mentioned approach is extended to generate Even and
Odd function plot of listed transforms.
We preprocess the fingerprint as discussed above and detect the
Registration Point (Core point or High Curvature point). This point is
used to select the Consistent Region of interest. User can place
fingerprint in varying position hence we take multiple samples to
select the ROI using method discussed above. This is shown in
Table 3.7; the fingerprints are captured from two different users.
3.1.2 Fingerprint Recognition using Kekre’s Wavelets
In the previous sections we have seen the fingerprint acquisition,
preprocessing steps in the design of automatic fingerprint
recognition systems. We have a fingerprint Region of Interest (ROI)
of dimensions 144*144 pixels after above mentioned operations.
This dimension is chosen because this was the maximum size of the
square fitting inside the fingerprints without violating the boundary
of the fingerprint images. This ROI is then used for feature
extraction. A typical fingerprint with above mentioned preprocessing
steps is shown in Fig. 3.25. The core region is also shown. This
region of interest is further used for enrolling fingerprints and
feature extraction. From each user we are selecting five samples
which will be used for training & testing of the algorithms discussed
here. In this section we will discuss texture feature extraction from
fingerprint. Fingerprint contains ridges and furrows, and their
organization is unique for each finger. This information is captured
by texture based features. We are using a new family of orthogonal
wavelets called as Kekre’s wavelets. For generation of Kekre’s
wavelets we need basis function as in case of other families, this
basis functions are generated from Kekre’s Transform matrix [214].
The generation of Kekre’s wavelet along with its properties is
discussed in detail in the next section. Advantage of Kekre’s wavelet
is that they can be generated by any number dimension (e.g.
144x144) using basis as Kekre’s transform. As we have ROI of 144*
83
144 pixels we can generate Kekre’s wavelet of this order without
scaling it to either 128x128 or 256x256 pixels.
3.1.2.1 Kekre’s Wavelets
Kekre’s wavelets are orthogonal family of wavelets. For
generation of Kekre’s wavelets we need basis function as in case of
other families, this basis functions are generated from Kekre’s
Transform matrix.
A. Kekre’s Transform [214]
Let us generate the Kekre’s Matrix [K] for size mxm where m can
be any integer not necessarily the power of 2 as required for many
other conventional transforms. This matrix has all 1’s on the main
diagonal and upper triangle of the matrix. The sub-diagonal just
below the main diagonal has the value (-m+i) where ‘m’ is the
order of matrix and ‘i’ is the column number. Rests of the elements
of lower triangle below the sub diagonal are all zeros. The general
form of Kekre’s matrix [K] can be written as
1 1 1 ….
1 1
-N+1 1 1 …
.
1 1
0 -N+2 1 …
.
1 1
.
.
.
.
.
.
.
.
.
…
.
.
.
.
.
.
.
0 0 0 ….
1 1
0 0 0 ….
-N+(N-1) 1
The formula for generating the element Kxy of Kekre’s transform
matrix is,
Kxy= (3.31)
The properties of Kekre’s Transform are as follows:
1) The Kekre’s transform is real and orthogonal transform.
[K]T [K]=[µ] (3.32)
K NxN =
1 ; x ≤ y
-N + (x-1) ; x= y + 1
0 ; x > y + 1
84
Where [k]T is transpose of [K] and [µ] is a diagonal matrix
and its elements are given by
µ11 = m
µii = (m-i+1)(m-i+2) (3.33)
2) It has a fast algorithm as it contains m(m+1)/2 number of
ones and (m-1)(m-2)/2 number of zeros leaving only (m-1)
integer multiplications and only (m-1)(m/2) additions for
transforming a column vector of dimension mx1. For a normal
matrix transformation we require m2 multiplications and m
(m-1) additions.
3) The transform of a vector f is given by
F = [K] f (3.34)
And inverse is given by
f = [K]T [µ]-1 F (3.35)
K11 K12 K13 … K1 (N-1) K1N
K21 K22
K23 … K2 (N-1) K2N
K31 K32
K33 … K3 (N-1) K3N
.
.
.
.
.
.
… .
.
.
.
KN1 KN2
KN3 … KN (N-1) KNN
Fig. 3.26. Kekre’s Transform (KT) matrix of size NxN
4) All entries in Kekre’s Transform Matrix are integers and hence no
floating point operations are required.
B. Kekre’s Wavelets [215]
Kekre’s Wavelet transform is derived from Kekre’s transform.
From NxN Kekre’s transform matrix, we can generate Kekre’s
Wavelet transform matrices of size (2N)x(2N), (3N)x(3N),……, up to
maximum (N2)x(N2). For example, from 5x5 Kekre’s transform
matrix, we can generate Kekre’s Wavelet transform matrices of size
10x10, 15x15, 20x20 and 25x25. In general MxM Kekre’s Wavelet
transform matrix can be generated from NxN Kekre’s transform
matrix, such that M = N * P where P is any integer between 2 and N
that is, 2 ≤ P ≤ N. Consider the Kekre’s transform matrix of size
NxN shown in Fig. 3.26. MxM Kekre’s Wavelet transform matrix
generated from NxN Kekre’s transform matrix is shown in Fig 3.26.
K NxN =
85
First ‘N’ number of rows of Kekre’s Wavelet transform matrix is
generated by repeating every column of Kekre’s transform matrix P
times. To generate remaining (M-N) rows, extract last (P-1) rows
and last P columns from Kekre’s transform matrix and store
extracted elements in to temporary matrix say T of size (P-1) x P .
Fig. 3.27 shows extracted elements of Kekre’s transform matrix
stored in T.
K(N-P+2) (N-P+1)
K(N-P+2) (N-P+2) … K(N-P+2) N
K(N-P+3) (N-P+1) K(N-P+3) (N-P+2)
….. K(N-P+3)N
.
.
.
.
….
….
.
.
KN(N-P+1)
KN(N-P+2) ….. KNN
Fig. 3.27.Temporary Matrix T of size (P-1) x P
Values of matrix T can be computed as,
T(x, y) = K( N-P+(x+1), N-P+ y) ; 1≤ x≤ (P-1) , 1≤ y≤ P (3.36)
First row of T is used to generate (N+1) to 2N rows of Kekre’s
Wavelet transform matrix. Second row of T is used to generate
(2N+1) to 3N rows of Kekre’s Wavelet transform matrix, Like wise
last row of T is used to generate ((P-1) N + 1) to PN rows [236].
We have used Kekre’s Wavelet Transform Matrices of Size 128,
64, 32 Generated from Kekre’s Transform Matrix of Size 64, 32, 16
respectively. We calculate Wavelet energy feature for the fingerprint
image using these wavelet matrices.
Properties of Kekre’s Wavelet Transform:
1. Orthogonal- The transform matrix K is said to be orthogonal
if the following condition is satisfied.
[K][K]T = [D], Where D is a diagonal matrix.
Kekre’s Wavelet Transform matrix satisfies this property
and hence it is orthogonal. The diagonal matrix value of
Kekre’s transform matrix of size NxN can be computed as
(3.37)
T =
86
2. Asymmetric- As the Kekre’s transform is upper triangular
matrix, it is asymmetric.
3. Non Involutional - An involutionary function is a function
that is its own inverse. So involution transform is a
transform which is inverse transform of itself. Kekre’s
transform is non-involution transform.
4. Transform on Vector -The Kekre’s Wavelet transform on a
column vector f is given by
F = [KW] f (3.38)
And inverse is given by
f = [KW]T [D]-1 F (3.39)
5. Transform on 2D Matrix- Kekre’s Wavelet transform on 2D
matrix f is given by
[F] = [KW] [f] [KW]T (3.40)
Obtaining Inverse:
Calculate Diagonal matrix D as,
[D] = [KW][KW]T (3.41)
D1 0 0 0 0 0
0 D2 0 0 0 0
0 0 D3 0 0 0
0 0 0 … 0 0
0 0 0 0 … 0
0 0 0 0 0 DN
Inverse is calculated as
[f] = [KW]T [ Fij / Dij ] [KW] (3.42)
Where Dij = Di * Dj ; 1≤ i ≤ N and 1≤ j ≤ N
Next an example of Kekre’s wavelet transform is given,
Fig. 3.34 shows the Kekre’s Wavelet transform matrix of size 15 x
15 generated from the Kekre’s transform matrix of size 5 x 5. Here
We have M =15, N =5 and P =M/N=3
1 1 1 1 1
-4 1 1 1 1
0 -3 1 1 1
0 0 -2 1 1
0 0 0 -1 1
D =
K5X5 =
87
As shown in Fig. 3.28, all the columns of Kekere’s transform matrix
are repeated P=3 times to generate first N=5 number of rows of
Kekre’s Wavelet transform matrix. To generate remaining (M-N) =
10 rows, extract last (P-1) = 2 rows and last P=3 columns from
Kekre’s transform matrix and store these elements into temporary
matrix T. The temporary matrix T is as follows .
-2 1 1
0 -1 1
The first row of T [-2 1 1 ] is used to generate next 5-10 rows of
KW transform matrix as shown above. Second row of T [0 -1 1] is
used to generate last 11-15 rows of KW transform matrix.
Fig. 3.28. 15x15 Kekre’s Wavelet Transform Matrix Generated from 5x5
Kekre’s Transform Matrix
3.1.2.2 Feature Vector Extraction
With the development of wavelet theory, Wavelet Analysis has
been valued highly in various domains of research. It is a powerful
tool of multi-resolution analysis. Here we construct wavelet energy
feature (WEF) by the high frequency to describe the fingerprint
images texture and use it to describe the ridges & principle lines.
We take the Kekre’s Wavelet (KW) Transform of the selected
fingerprint ROI. The wavelets will capture localized spectral
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
-4 -4 -4 1 1 1 1 1 1 1 1 1 1 1 1
0 0 0 -3 -3 -3 1 1 1 1 1 1 1 1 1
0 0 0 0 0 0 -2 -2 -2 1 1 1 1 1 1
0 0 0 0 0 0 0 0 0 -1 -1 -1 1 1 1
-2 1 1 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 -2 1 1 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 -2 1 1 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 -2 1 1 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 -2 1 1
0 -1 1 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 -1 1 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 -1 1 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 -1 1 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 -1 1
5th column of
KT repeated P
=3 times
1st
column of
KT repeated P
=3 times
2ndt column of
KT repeated P
=3 times
3ndt column of
KT repeated P
=3 times
4th column of
KT repeated P
=3 times
KW15X15 =
88
information from the ROI. We have the ROI of Size 144 X 144
Pixels. At each level Mth order KW matrix (of MXM Size) is generated
by M/2 order Kekre’s Transform Matrix (N=M/2, P=2). The wavelet
energy in horizontal, vertical and diagonal directions at the i-level
can be, respectively, defined as:
2
1 1
( ( , ))M N
h
i i
x y
E H x y (3.43)
2
1 1
( ( , ))M N
v
i i
x y
E V x y (3.44)
2
1 1
( ( , ))M N
d
i i
x y
E D x y (3.45)
(256x256)
(128x128)
(128x128)
LL LH
(128x128)
(128x128)
HL HH
(a) (b)
Fig. 3.29. Wavelet Decomposition of Selected ROI of Fingerprint Image
(a) Selected ROI (b) Kekre’s Wavelet First level Components
These energies reflect the strength of the images’ details in different
direction at the i-level decomposition. Hence the feature vector
1,2,3,...,( , , )h v d
i i i i kE E E where K is the total number of wavelet
decomposition level, can describe the global details feature of
fingerprint texture. One such decomposition is shown in Fig. 3.29.
Using above mentioned vector, the features extracted from the
whole ROI don’t preserve the information concerning the spatial
89
location of different details of ridges and principle lines, so its ability
to describe fingerprint uniqueness is weak. In order to deal with this
problem, we divide the detail images into S×S non-overlap blocks
equally, and then compute the energy of each block. Then, the
energies of all blocks are used to construct a vector. This is shown
in Fig. 3.30. Finally the vector is normalized by the total energy. We
are also normalizing this vector each level and Component wise
also. These normalized vectors are named as wavelet energy
feature.
WEF has a strong ability to distinguish fingerprints. According to
these figures, WEFs of the same fingerprints are very similar while
those of different fingerprint are quite dissimilar. This is helpful for
fingerprint recognition.
Fig. 3.30. Dividing Wavelet Components into 4x4 Non-Overlapping Blocks
Only Horizontal, Vertical and Diagonal Components are divided into 4x4
blocks. Each component gives 16 values and per level we get 48 values of
wavelet energy
Each component is divided into 4x4 (SxS) non-overlapping blocks;
hence for a single component (LH, HL or HH) we have 16 wavelet
energy values. In a single decomposition we have 3 components
(LH, HL & HH) hence we get total 48 components. We normalize the
feature vector by dividing energy of components at each level by
the sum of all the components energy at that level. We have such 5
levels (J) of decompositions hence we have total 3xSxSxJ i.e. 240
(3*4*4*5) values in the wavelet energy feature vector. We call this
feature vector as Kekre’s Wavelet Energy Feature Vector (KWEFV).
KWEFV = {WE0, WE1, ….., WEn) n=3xSxSxJ. (3.46)
90
(a)
(b)
Fig. 3.31. Kekre’s Wavelet Energy Feature Vector Plot (a) Normalized by
Total Energy (b) Normalized by Level-wise Energy
(a)
(b)
Fig. 3.32. Kekre’s Wavelet Energy Distribution for Each Component (a)
Energy of Each Component (b) Energy of Each Level
91
For the fingerprint shown in Fig. 3.29 the plot of Wavelet Energy
Coefficients (KWEFV) are shown in Fig. 3.31. The coefficients are
normalized in two different ways. In one method we use total
energy to divide all coefficients, the plot is shown in Fig. 3.31 (a)
and in other method the sum of coefficient of each level
(LH+HL+HH) is used for normalization, the plot is shown in Fig 3.31
(b). We have total 15 Components (3*J), We also find total energy
of each component for analysis. The component energy is
normalized by total energy at the level. The distribution is a coarser
estimate of spectral content. This is shown in Fig. 3.32.
3.1.2.3 Relative Wavelet Entropy
We are using two metrics for analyzing similarity between two
energy distributions (Kekre’s Wavelet Energy Feature Vector). We
are using Wavelet Energy Entropy [216] as one of the similarity
measure & other is the Euclidian distance between two sequences.
We have normalized the wavelet energy feature vector hence this
can be treated as a probability distribution.
Let us now suppose that we have two different probability
distributions {pi} and {qj}, with ∑j pj = ∑j qj =1. Here consider
them as wavelet energy distributions. We define Relative Wavelet
Entropy as,
0
( | ) lnn
j
wt j
j j
pS p q p
q (3.47)
Which give degree of similarity between two probability distributions
with respect to each other. The RWE is a positive real number and it
vanishes when j jp q . Fig. 3.33 presents three different relative
wavelet energy (Probability) distribution levels (j=1,2,..,5). It is
clear from the figure that distributions A and B are quite similar,
and present broad band spectra. In contrast, distribution C sows a
clear dominance of the resolution level j=-2. According to this
description, for the total WE the following relation can be expected:
SWT(A) ≈ SWT(B) > SWT(C). We can see that SWT(B|A) ≈0 and
SWT(C|A)>>0. When Corresponding numerical values for the
distribution are used, a very good match with previous relation is
obtained. We are using this metric for matching the Wavelet Energy
Distribution per level as well as for full sequence of wavelet energy
coefficients.
92
(a) (b)
(c)
Fig. 3.33. Relative Wavelet Energy (Probability) Distribution for Wavelet
Resolution Levels (j=1,…..,5). Distribution A,{pj}={0.06, 0.10, 0.30,
0.34, 0.20}; B,{pj}={0.03, 0.10, 0.12, 0.33, 0.37}; C,{pj}={0.03, 0.11,
0.11, 0.70, 0.05}. The WE values for these distributions are
SWT(A)=1.440, SWT(B)=1.37, SWT(C)=0.994. Taking Distribution A as
reference, SWT(B|A)=0.015 and SWT(C|A)=0.220
3.1.2.4 Results
We are using scanned fingerprint from 57 different persons
captured using FS88 fingerprint scanner. For each person 5 samples
were taken for training & testing. Imposter fingerprints were
selected from other person’s samples. We have extracted Wavelet
Energy Feature Vectors using Kekre’s Wavelets and Haar Wavelets
(Modified). As discussed earlier the feature vector is normalized per
level. Fig. 3.25 & Fig. 3.26 shows some of these feature vectors. We
are representing the fingerprint by its feature vector (KWEFV). We
analyze similarity in three modes.
The feature vector is normalized level wise i.e. for all 48
components (16 LH, 16 HL, 16 HH) and then we can evaluate
the Relative Wavelet Energy Entropy for this normalized
0
0.1
0.2
0.3
0.4
1 2 3 4 5
A
Wavelet Resolution Level
Re
lati
ve E
ner
gy
0
0.1
0.2
0.3
0.4
1 2 3 4 5
B
Re
lati
ve E
ner
gy
Wavelet Resolution Level
0
0.2
0.4
0.6
0.8
1 2 3 4 5
C
Wavelet Resolution Level
Re
lati
ve E
ner
gy
93
feature vector, final matching is given by summation of all
levels relative entropy.
Euclidian Distance (ED) between two KWEFV sequences (Seq.
X & Seq. Y).
2
, ,
0
( )n
x i y i
i
ED KWEFV KWEFV (3.48)
RWEE (Relative Wavelet Energy Entropy) for two normalized
KWEFV directly. We take full wavelet energy coefficient
sequence and normalize it by total energy. This distribution is
then used for finding full sequence relative entropy.
We have performed total 672 tests in parallel for intra class
matching i.e. genuine fingerprint testing and 5330 tests for inter
class matching i.e. cross matching and imposter testing. Euclidian
distance and relative energy entropy for each test is calculated. We
have divided the test results in two classes as Genuine Tests
Distance and Forgery Tests Distance (Imposter distance). The range
of distance values against the participation in specific class is shown
in Fig. 3.34. We can see that two peaks for two test classes. For
genuine fingerprints the relative distance lies in the range of 10 to
65 and that for imposter fingerprint lies in the range of 45 to 110.
Fig. 3.34. Relative Probability for Matching Distance of Genuine and
Forgery Tests
Two clear classes can be seen, with threshold distance as 55. This can be
used for designing classifier
0
5
10
15
20
25
30
35
10
20
30
40
50
60
70
80
90
10
0
11
0
12
0
13
0
14
0
15
0
16
0
17
0
18
0
19
0
20
0
21
0
22
0
23
0
24
0
25
0
Pro
bab
ility
(%
)
Distance Probability for Genuine and Forgery Tests
P-GEN
P-Forge
Distance
94
We are using simple Euclidian Distance based K-nearest
neighborhood classifier (K-NN). Evaluation metrics such as False
Acceptance Rate (FAR), False Rejection Rate (FRR), True
Acceptance Rate (TAR) and True Rejection Rate (TRR) are
evaluated. They are specified as follows,
Total Number Imposter Fingerprints Accepted as Genuine
Total Number of Forgery Tests PerformedFAR
(3.49)
Total Number Genuine Fingerprints Rejected as Imposter
Total Number of Genuine Matching Tests PerformedFRR
(3.50)
Total Number Genuine Fingerprints Accepted
Total Number of Genuine Matching Tests PerformedTAR
(3.51)
Total Number Imposter Fingerprints Rejected
Total Number of Forgery Tests PerformedTRR
(3.52)
A. Relative Energy Entropy of Full Sequence of Wavelet Energy
We evaluate the Relative Energy Entropy for two KWEFV Full
Sequence.
Fig. 3.35. Test Results for Relative Entropy of Full Sequence of Energy
Feature Vector FAR-FRR Plot
(RKEEF : Relative Kekre’s Energy Entropy Full Sequence)
Higher the value lesser the matching and for matching
fingerprints the RWEE value is low. We plot the FAR-FRR curve for
accuracy calculations. We can see that from Fig. 3.35 the Equal
Error Rate (EER) i.e. the rate at which FAR and FRR both are
0
10
20
30
40
50
60
70
80
90
100
0 5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
10
0
10
51
10
FRR
FAR
RKEEF FAR vs FRR Plot
% R
ate
Distance Threshold
EER = 22%
PI= 78%
95
minimum is at 22%, We are evaluating relative Energy Entropy
(REE) between two wavelet energy sequences; one such sequence
is shown in Fig. 3.31 (b). The EER for FAR-FRR analysis should be
low (ideally zero). We define Performance Index (PI) of a biometric
system based on this as follows,
PI=100-EER (3.53)
To indicate the performance of the systems hereafter
Performance Index (PI), Equal Error Rates (EER) of FAR-FRR Plot
and Correct Classification Ratio (CCR) are used. Correct
classification ratio is defined previously. This convention is followed
throughout the thesis.
B. Relative Energy Entropy of Energy Sequence Normalized
Levelwise
Here we have calculated the Euclidian Distance between two
KWEFV sequences. The sequence is normalized for each level. This
mode of matching gives maximum accuracy. Performance Index
(PI) of this system is 86% and 14% EER for FAR-FRR, which is
shown below.
Fig. 3.36. Test Results for Relative Entropy for Level wise Normalized
Sequence FAR-FRR Plot
C. Euclidian Distance of KWEFV Normalized Level Wise
Here we have calculated the Euclidian Distance between two KWEFV
sequences. The analysis is shown in Fig. 3.37.
0
10
20
30
40
50
60
70
80
90
100
0
20
40
60
80
10
0
12
0
14
0
16
0
18
0
20
0
22
0
24
0
26
0FRR
FAR
RKEEF FAR vs FRR Plot
Distance Threshold
% R
ate
PI= 86%
EER = 14%
96
Fig. 3.37. Test Results for Euclidian Distance Between Wavelet Energy
Sequences ( FAR-FRR Plot)
The sequence is normalized for each level, i.e. by total energy of
each level. This mode of matching gives maximum accuracy. It has
PI of 89% and 11% EER for FAR-FRR.
D. Fusion of Relative Entropy & Euclidian Distance based Metrics
Here we have fused the score of Relative Entropy and Euclidian
Distance based distance. We have used linear fusion of score where
the final score Df is given by fusion of Relative Entropy ER & El
(Levelwise Entropy) and Normalized KWEFV distance DWN.
Df = W1*ER + W2*El+ W3*DWN (3.54)
Where W1, W2 & W3 are fusion weights, W1=0.001, W2= 10, W3=10.
These weights are empirically decided so that their contribution
towards final distance is equal; this is also called as score
normalization.
For Kekre’s wavelet based feature vector the observed Range
for ER is (1890- 3940), for El the range is (0.805 – 1.564) and that
of Dwn is (0.2950-0.5245), after weighting this range becomes
(1.890- 3.940), (8.05 – 15.64) & (2.950-5.245) respectively. Fusion
of these scores gives the Performance Index of 90% and EER as
10%.
The similar feature vectors are extracted using Haar wavelets
also and performance of both the wavelets is compared.
Comparison of above mentioned techniques and given in Fig. 3.38.
0
10
20
30
40
50
60
70
80
90
100
0
40
80
12
0
16
0
20
0
24
0
28
0
32
0
36
0
40
0
44
0
48
0
52
0
FRR
FAR
RWEEL FAR vs FRR Plot
Distance Threshold
% R
ate
EER= 11%
PI= 89%
97
Fig. 3.38. Comparison of Kekre’s & Haar Wavelet Based Fingerprint
Matching Techniques
We can see that Euclidian distance based feature vector
classification method has highest Performance Index. The
performance of Level wise entropy is next best. The fusion
performed here is acting like weighted average. We can conclude
that Euclidian distance based classifier works best for the wavelet
energy based fingerprint matching. Besides this Kekre’s Wavelet
based feature vector gives higher PI for fusion.
Table 3.8
Summary of Fingerprint Matching Tests
Sr. Type of Wavelet Accuracy (%) - CCR
1 Kekre’s Wavelets 84.40
2 Haar Wavelets 81.15
Finally we have performed total 520 tests for intra class matching &
5330 tests for inter class matching for both Kekre’s & Haar
Wavelets. The total accuracy (CCR) is for Kekre’s wavelets based
fingerprint recognition system is 84.40% and that of Haar wavelet is
81.15%; we have used simple Euclidian distance based K-NN
classifier with three training samples per person. This is an example
of multi-algorithmic biometric fusion. Performance of Kekre’s
Wavelet is better and they give higher PI & CCR. Next we discuss
use of Walsh, Hartley, Kekre’s Transform, DCT & Kekre’s Wavelets
for fingerprint recognition.
KWEFV RWEEL RKEEF Fusion
PI-Kekre 89 86 78 90
CCR-Kekre 86.22 81.52 72.04 84.4
PI-Haar 89 84 75 88
CCR-Haar 85.87 79.21 70.25 81.15
0102030405060708090
100
% R
ate
Comparison of PI & EER (FAR-FRR) for Wavelet Energy Based Fingerprint Matching Techniques
98
3.1.3. Fingerprint Recognition using Partitioned Complex
Walsh Plane in Transform Domain
Here we discuss a method which deals with fingerprint
identification in the transform domain is considered. The one-step
Walsh transform i.e. either the row or the column transforms of the
fingerprint is subjected to partitioning to generate the feature
vector. This process is based on Cal & Sal Functions of Walsh
Transform, next we discuss the Walsh transform & it’s Cal, Sal
functions.
3.1.3.1 Walsh Functions [217]
Walsh functions are a set of orthogonal functions which can be
used to represent any discrete-time signal. The Walsh functions (W0
- W7) as shown in the Fig.3.39 are generated from square wave
functions of different sequency.
Fig. 3.39. First Eight Walsh Functions
The even functions (C0 - C3) are called Cal functions and the
odd functions (S1-S4) are called Sal functions. The basic square
wave functions are S1, S2 and S4. C0 is DC component and the
remaining functions are generated from the basic square waves by
EX-OR operation (equivalent to multiplication). This operation
generates only the difference sequency functions (as opposed to the
case of sinusoidal signals where both difference and sum
frequencies are generated) e.g. C1 = S1 ⊕ S2, here S1 and S2
being odd function, their EX-OR operation results in an even
function (C1). Similarly EX-OR operation of an even and odd
function generates an odd function e.g. S3 = S4 ⊕ C1, which can
further be simplified to S3 = S4 ⊕ S2 ⊕ S1, showing that all
99
functions are generated from the basic square waves S1, S2 and
S4.
Walsh functions can be ordered in a number of ways.
The sequency ‘k’ of a Walsh function is defined as half the number
of zero crossings in one cycle of the time base. Walsh functions with
non-identical sequencies are orthogonal, as are the functions W(n,
2k) and W(n, 2k+1). The product of two Walsh functions is also a
Walsh function. Harmuth in [217] designates the even Walsh
functions Cal(k) and the odd Walsh functions Sal(k)[218],
Cal (n, k) = W(n, 2k) (3.55)
Sal (n, k) = W(n, 2k+1) (3.56)
where ‘k’ is the sequency.
The Walsh transform matrix (W) is then generated by sampling
these Walsh functions at the middle of the smallest time interval.
The matrix, as in Eqn. (3.57) is obtained, which can be directly used
to generate the transform coefficients of a discrete signal both of 1-
D and 2-D as shown in Eqn. 3.55 and Eqn. 3.56 respectively,
(3.57)
F1D=W.f (3.58)
F2D=W.f.WT (3.59)
The interpretation of Walsh transform of a 2-D signal can be
understood by Fig. 3.46, where first the row transform is calculated
and then the column transform. The final output has DC component
in the top left corner and the sequency components increase
leftwards and downwards.
In the current approach, we are first generating the
intermediate transform, i.e. the row transform (or column
transform) of a fingerprint image as shown in Fig. 3.40., which have
DC component as its first row (or column) and higher sequency
components (Sal and Cal) as the following rows (or columns).
100
Fig. 3.40. Transform of a 2D Function
(a) (b)
Fig. 3.41. a) Row Transform and b) Column Transform of a Fingerprint
3.1.3.2 Complex Walsh Plane [219] & Feature Vector
Generation
The Cal and the Sal components of the same sequency are
grouped together and are considered to be in the four quadrants of
2-D complex coordinate plane as listed in Fig.3.41. This complex
plane is now partitioned into different numbers of blocks.
Fig. 3.42. Complex Walsh Plane
101
The complex plane consisting of same-sequency (Sal, Cal)
components is now partitioned 256 square blocks as shown in
Fig.3.43. For each block a feature vector is generated which is the
mean value of all the transform coefficients in that block, as well as
the number of points i.e. the density is also considered. Such a
complex Walsh plane & its partitioning is shown in Fig. 3.43.
(a) (b)
Fig. 3.43. Complex Walsh Plane (a) Partitioned Cal+jSal Function Plot of
Row Transform (b) Partitioned Cal+jSal Function Plot for Column
Transform
This value is unique for each fingerprint as the sequency
distribution of each fingerprint is unique in different blocks. As
compared to all or those transform coefficients which contain major
part of signal energy feature vectors generated using partitioning
are much less in number and hence the reduction in processing time
and complexity. The blocks generated are square shaped and the
mean values of the transform coefficients in each block are
calculated as in Eqn. 3.60, where Mk is the mean and N is the
number of coefficients in a block, which form the features. The DC
component, separate means of the Sal and Cal component and the
last sequency component together form the feature vector, and
hence the number of features is 2S+2, where S is the number of
blocks.
1
1 n
k i
i
M WN
(3.60)
The features obtained from the test image are compared with
those obtained from the stored fingerprint in the database and the
102
results matched. The Euclidian distances between the feature
vectors of the test image and the database images are calculated.
The minimum distance gives the best match.
3.1.3.3 Fingerprint Feature Vector Extraction
As discussed above the intermediate Walsh transform is used to
generate the complex Walsh plane, this plane is then partitioned
into 256 blocks (16x16). In each block the mean is calculated as
well as DC component and the last sequency component is together
treated as feature vector.
(a) (b)
(c) (d)
Fig. 3.44. Partitioned Complex Walsh Plane of Fingerprint
(a) Row Transform Function Plot for Full Fingerprint
(b) Column Transform Complex Function Plot for Full Fingerprint
(c) Row Transform Complex Function Plot for Core Point ROI
(d) Column Transform Complex Function Plot for Core Point ROI
103
In the previous sections methods for fingerprint segmentation as
well as core point detection are discussed. The current method is
tested on the segmented fingerprint as the detected core area. The
fingerprint matching is performed in two modes
1. Full Segmented fingerprint is considered for feature vector
generation
2. First the core point (or registration point) on the fingerprint is
located. A region of interest of m*m pixels is segmented and
considered for feature extraction. Both the plots are shown in
Fig. 3.44.
The Optical fingerprint scanner Futronics FS88 is used for live
fingerprint capture, this device gives image of size 320*480 pixels
at 500dpi. This image is preprocessed and segmented to remove
the background as well as noise. Then the core point is detected
from the fingerprint. The Region of Interest (ROI) of size 144*144
pixels is the segmented from the preprocessed fingerprint. Finally,
full segmented fingerprint as well as the core point ROI is available
for feature extraction. This procedure is shown below,
The segmented fingerprint as shown in Fig. 3.11 and Core
Point ROI as shown in Table 3.7 are then subjected to the
intermediate transform generation. The complex Cal + jSal function
plots are generated by intermediate Row & Column Walsh
transform. As discussed previously the complex Walsh plane is
partitioned. The complex Walsh plane plots for full finger as well as
core point ROI is shown in Fig. 3.44. As discussed earlier each plot
gives 2S+2 coefficients, we have 256 blocks in each plot, hence one
plot gives 514 (256*2 +2) coefficients. For each type of input i.e.
segmented fingerprint and core point ROI we have two plots, one
for row and one for column transform hence we have 1028 (514 *2)
coefficients for each type of fingerprint input. Finally we have 1028
coefficients in the feature vector of segmented fingerprint and
another 1028 for core point ROI. Similar Feature vector is
generated for Density of the points in complex Walsh Plane for each
fingerprint input. This feature vectors are used for enrollment and
matching of the fingerprints.
3.1.3.4 Results
To test the matching algorithm, 285 samples collected from 57
persons (5 samples per person) have been used. The experiment
was conducted on two different modes; one is on Full Segmented
104
Fingerprint and another is on Fingerprint Core point ROI. Total 2421
different tests are performed. Equal Error Rate (EER) is evaluated
and PI (Performance Index) is calculated. While testing the DC &
Sequency components and its effect on matching is also evaluated.
For each type of fingerprint input the feature vector is generated in
following variations
1. Row transform mean feature vector (Row TRF).
2. Column transform mean feature vector (Col TRF).
3. Row density feature vector (Row Density).
4. Column density feature vector (Col Density).
5. Fusion of above mention feature vectors with DC & Sequency
components (Fusion).
The fusion is performed by score normalization. The
normalization is performed by weighting the distance by specific
coefficient decided empirically to give proper weightage to each
feature vector. These feature vectors are extracted for full
fingerprint as well core point area of fingerprint. From each user five
samples are taken out of that two samples are used for testing and
remaining samples are used for training.
(a) (b)
Fig. 3.45. FAR-FRR Analysis for Walsh Cal-Sal based fused Feature
Vectors (a) Plot for Fingerprint Core point ROI Feature Vector (b) Plot for
Full Segmented Fingerprint Feature Vector
For classification Euclidian distance based K-NN classifier is used.
This algorithm is tested on a machine running Windows XP SP3,
with AMD Athlon 64FX Processor running at 1880 MHz and 1.5 GB
of RAM. The FAR-FRR plot for Fused Feature vector matching for full
0102030405060708090
100
40 80 120 160 200 240 280 320 360
% R
ate
Fingerprint Core Point Feature Vector Fusion FAR-FRR
FRR
FAR
Threshold
EER= 05% PI = 95%
0102030405060708090
100
60 100 140 180 220 260 300
% R
ate
Full Finger Feature Vector Fusion FAR-FRR
FRR
FAR
Threshold
EER= 19% PI=91%
105
finger and core point area are shown in Fig. 3.45. The plot shows
that EER for core point ROI based feature vector is high (EER= 5%)
for FAR-FRR Analysis, this shows high accuracy as compared to
feature vector of full segmented finger (EER =19%). This is mainly
because the core point ROI is more consistent and has less
variations due to change in finger placement, pressure applied,
dryness of finger etc. The results are summarized in Fig. 3.46 for
all the feature vectors variations.
Fig. 3.46. Performance Comparison for Feature Vector Variants of
Partitioned Walsh Cal-Sal Function Fingerprint Matching.
Score Fusion based Matching Gives Higher Performance Index; this is
Indicated by Bars in Red Colour. (TRF: Transform, FV: Feature Vector)
It is clearly seen that the core point based feature vectors give
higher accuracy. As compared to the density based feature vector
the coefficient mean based feature vector give higher accuracy. The
Fusion of Row & Column Transform mean & Density with DC &
Sequency coefficient gives 95% PI for core point based feature
vector and that of full segmented fingerprint’s feature vector gives
81% PI. The individual Row & column transform mean based
feature vectors have 85% & 88% PI for core point ROI based
feature vector, this shows that due to fusion of feature vector with
DC & Sequency component the performance has improved. Similar
improvement can be seen with full fingerprint’s feature vectors.
In the next section we continue this feature extraction
mechanism for Even and Odd functions (similar to Cal & Sal function
of Intermediate Walsh transform) of intermediate transforms of
Hartley Transform, Kekre’s Transform, DCT & Kekre’s Wavelets.
RowTRF-CorePoint
ColTRF-CorePoint
Row-Density-CorePoint
Col-Density-CorePoint
Fusion-CorePoint
FV
RowTRF-Full
Finger
ColTRF-Full
Finger
Row-Density
-FullFinger
Col-Density
-FullFinger
Fusion- Full
FingerFV
PI 85 88 80 82 95 80 70 66 69 81
0102030405060708090
100
% P
I
Partitioned Walsh Cal Sal Function Based Fingerprint Matching Performance Comparison of PI
106
3.1.4. Fingerprint Matching using Partitioned Complex Plane
in Transform Domain of Hartley Transform, Discrete Cosine
Transform, Kekre’s Transform and Kekre’s Wavelets
The feature vector extraction mechanism discussed earlier is
extended here for other transforms. The complex plane is generated
by Cal & Sal function of Intermediate Walsh transform as shown in
Fig. 3.43; in case of other transforms Cal and Sal functions
analogous part is Even & Odd functions. The even and odd function
are used to generate complex transform plane by plotting complex
pair (Even+jOdd), then the above discussed feature vectors are
extracted for fingerprint matching. We discuss the results one by
one for the listed transforms.
3.1.4.1 Fingerprint Matching using Partitioned Hartley Plane
in Transform Domain
A discrete Hartley transform (DHT) is a Fourier-related transform
of discrete, periodic data similar to the discrete Fourier transform
(DFT), with analogous applications in signal processing and related
fields. Its main distinction from the DFT is that it transforms real
inputs to real outputs, with no intrinsic involvement of complex
numbers. Just as the DFT is the discrete analogue of the continuous
Fourier transform, the DHT is the discrete analogue of the
continuous Hartley transform, introduced by R. V. L. Hartley in 1942
[220].
The real valued Hartley transform [221] is an alternative to the
more familiar complex valued Fourier transform [222], [223].
Because any two dimensional intensity distributions has only real
values, the Fast Hartley transform technique is an alternative to the
Fast Fourier transform for transforming and inverse transforming
images.
The Hartley transform has the advantage of having the same
formula for forward and inverse transform (it is involutionary) and
tends to require less memory space since complex numbers are not
used, and a separate inversion program is not required. Regardless
of speed, the Hartley transform approach is worthy of consideration
for image processing applications due to its theoretical elegance,
symmetry of the inversion formula, and the economy of memory
utilization. In the case of two-dimensional real image data, the
discrete Hartley transform H (u, v) is defined as
107
1 1
0 0
1( , ) ( , ) [2 ( )]
( )
M N
x y
ux vyH u v f x y cas
MN M N (3.61)
The kernel [2 ( )]ux vy
casM N
is not separable into products of
factors, unlike exp[ 2 ( )]ux vy
iM N
.
(a) (b)
(c) (d)
Fig. 3.47. Partitioned Complex Hartley Plane of Fingerprint
(a) Row Transform Function Plot for Full Fingerprint
(b) Column Transform Complex Function Plot for Full Fingerprint
(c) Row Transform Complex Function Plot for Core Point ROI
(d) Column Transform Complex Function Plot for Core Point ROI
Since the One dimensional FHT has space advantage over the
one dimensional FFT for real numbered data, a method by Bracewell
[224] which is directly analogous to the two-dimensional FFT can be
108
used for the two dimensional fast Hartley transform. It takes the
one dimensional Hartley transform of the rows one by one, and then
transform the columns using the expression. Same approach is
implemented here. The intermediate Hartley transform is generated
by taking transform of rows first and then the even and odd rows of
this transforms are used for generating the coefficients for
(Even+jOdd) complex plane plot. These functions are similar to Cal
& Sal functions as used to generate the (Cal+jSal) complex plane.
The complex plane for Hartley transform is shown in Fig. 3.47. Plots
for (Even+jOdd) function points of full finger as well as core point
ROIs are shown. The feature vector as discussed earlier are
extracted from them and used for matching in exactly same way as
discussed for Walsh transform. The performance comparison of the
feature vectors is shown below.
Fig. 3.48. Performance Comparison for Feature Vector Variants of
Partitioned Hartley Even-Odd Function Fingerprint Matching.
Score Fusion based Matching Gives Higher Performance this is Indicated
by Bars in Red Colour. (TRF: Transform, FV: Feature Vector)
The core point based feature vectors give higher accuracy as
they have higher PI (Performance Index), we can see that first five
columns (belonging to fingerprint core point ROI based feature
vectors) have high EER as compared to last five columns for TAR-
TRR analysis. The fusion of row & column transform mean & Density
with DC & Sequency coefficient gives 94% PI for core point based
feature vector and that of full segmented fingerprint’s feature vector
gives 84% PI. The individual Row & column transform mean based
feature vectors have 88% & 89% PI for core point ROI based
RowTRF-CorePoint
ColTRF-CorePoint
Row-Density-CorePoint
Col-Density-CorePoint
Fusion-CorePoint
FV
RowTRF-Full
Finger
ColTRF-Full
Finger
Row-Density
-FullFinger
Col-Density
-FullFinger
Fusion- Full
FingerFV
PI 88 89 82 80 94 83 77 70 71 84
0102030405060708090
100
% P
I
Partitioned Hartley Even Odd Function Based Fingerprint Matching Performance Comparison of PI
109
feature vector, this shows that due to fusion of feature vector with
DC & Sequency component the performance has improved. Similar
improvement is seen with full fingerprint’s feature vectors as shown
in Fig. 3.48. Next we discuss Discrete Cosine Transform (DCT)
based feature vectors.
3.1.4.2 Fingerprint Matching using Partitioned DCT Plane in
Transform Domain
(a) (b)
(c) (d)
Fig. 3.49. Partitioned Complex DCT Plane of Fingerprint
(a) Row Transform Function Plot for Full Fingerprint
(b) Column Transform Complex Function Plot for Full Fingerprint
(c) Row Transform Complex Function Plot for Core Point ROI
(d) Column Transform Complex Function Plot for Core Point ROI
110
Fig. 3.50. Performance Comparison for Feature Vector Variants of
Partitioned DCT Even-Odd Function Fingerprint Matching.
Score Fusion based Matching Gives Higher Performance this is Indicated
by Bars in Red Colour. (TRF: Transform, FV: Feature Vector)
The discrete cosine transform (DCT) represents an image as a
sum of sinusoids of varying magnitudes and frequencies. The DCT
function for 2-dimensional image is given by Eqn. 3.62 & 3.63.
1 1 (2 1) (2 1)
B2 20 0
mnC CpqM N m p n p
A os osM Nm n
p q (3.62)
1 If 0
2 If 1 1
p pM
p MM
1 If 0
2 If 1 1
q qN
q NN
(3.63)
Where Bpq are called the DCT coefficients of A which can be an
image data A(m, n). The DCT decomposes a signal into its
elementary frequency components. When applied to an MXN
image/matrix, the 2D-DCT compresses all the energy information of
the image and concentrates it in a few coefficients located in the
upper-left corner of the resulting real-valued MXN DCT frequency
matrix [225]. First the DCT kernel of desired size is generated and
then it is used for generating intermediate transform matrix of the
input fingerprint image. The intermediate transforms are generated
as discussed for Walsh & Hartley transform to generate the complex
plane in transform domain using Even and Odd functions of the
RowTRF-CorePoint
ColTRF-CorePoint
Row-Density-CorePoint
Col-Density-CorePoint
Fusion-CorePoint
FV
RowTRF-Full
Finger
ColTRF-Full
Finger
Row-Density
-FullFinger
Col-Density
-FullFinger
Fusion -Full
FingerFV
PI 80 75 71 75 79 75 67 66 63 70
0
10
20
30
40
50
60
70
80
90%
PI
Partitioned DCT Even Odd Function Based Fingerprint Matching Performance Comparison of PI
111
intermediate transform. The even and odd functions are used to
plot (Even + jOdd) points in complex plane and then the feature
vector variants are extracted as discussed earlier. The complex
plane for DCT is shown in Fig 3.49. This plane is partitioned and
mean and density based feature vectors are extracted. The
matching results for these feature vectors are summarized in Fig.
3.50. The fusion of even and odd functions and their mean as well
as density with DC & Sequency component gives highest PI of 79%
for fingerprint core point ROI. The feature vector generated for full
fingerprint however have low PI of 70%.
(a) (b)
(c) (d)
Fig. 3.51. Partitioned Complex Kekre’s Transform Plane of Fingerprint
(a) Row Transform Function Plot for Full Fingerprint
(b) Column Transform Complex Function Plot for Full Fingerprint
(c) Row Transform Complex Function Plot for Core Point ROI
(d) Column Transform Complex Function Plot for Core Point ROI
112
3.1.4.3 Fingerprint Matching using Partitioned Kekre’s Plane
in Transform Domain
The Kekre’s transform is discussed in detail in section 3.1.2.1.
The Kekre’s transform matrix is generated as discussed and then
the intermediate row and column transform of the fingerprint input
image is taken to extract the even and odd function points and
these points are plotted in complex plane. The partitioned complex
plane of Kekre’s transform even and odd function point are shown
in Fig. 3.51.
The Performance comparison of feature vectors extracted from
this is given in Fig.3.52. Here the Column Transform mean based
feature vector of fingerprint core point ROI gives best performance
of (PI=84% for Core TRF-Core Point Feature Vector). Fusion of
feature vectors has lower performance (PI=83% for Fusion-Core
point FV). For full fingerprint fusion of feature vectors gives PI of
81%, much higher than individual feature vectors EER. In the next
section we discuss the Kekre’s wavelets based feature vectors
performance.
Fig. 3.52. Performance Comparison for Feature Vector Variants of
Partitioned Kekre’s Transform Even-Odd Function Fingerprint Matching.
Score Fusion based Matching Gives Higher Performance this is Indicated
by Bars in Red Colour. (TRF: Transform, FV: Feature Vector)
RowTRF-CorePoint
ColTRF-CorePoint
Row-Density-CorePoint
Col-Density-CorePoint
Fusion-CorePoint
FV
RowTRF-Full
Finger
ColTRF-Full
Finger
Row-Density
-FullFinger
Col-Density
-FullFinger
Fusion -Full
FingerFV
PI 76 84 80 81 83 64 72 68 69 81
0
10
20
30
40
50
60
70
80
90
% P
I
Partitioned Kekre's Transform Even Odd Function Based Fingerprint Matching Performance Comparison of PI
113
3.1.4.4 Fingerprint Matching using Partitioned Kekre’s
Wavelet’s Plane in Transform Domain
(a) (b)
(c) (d)
Fig. 3.53. Partitioned Complex Kekre’s Wavelet Plane of Fingerprint
(a) Row Transform Function Plot for Full Fingerprint
(b) Column Transform Complex Function Plot for Full Fingerprint
(c) Row Transform Complex Function Plot for Core Point ROI
(d) Column Transform Complex Function Plot for Core Point ROI
The Kekre’s Wavelets are new family of wavelet, discussed in
section 3.1.2.1. Here the wavelets are used in a novel way to
extract the intermediate row & column wavelet transform. The
wavelet matrix is generated form Kekre’s transform matrix with
proper selection of M,N & P Parameters (p=M/N, takes as P=2 for
current feature vector generation using Kekre’s wavelets). The
wavelet matrix is used for taking intermediate column & row
transform. The even and odd functions of the intermediate
transform are used to plot (Even + jOdd) points in complex plane.
114
Here the wavelets are used for extracting the complex plane based
features in transform domain rather than the texture features of the
fingerprints. The complex plane plots are shown in Fig. 3.53. The
feature vectors are generated as discussed previously and the
fingerprint matching analysis is shown in Fig. 3.54.
Fig. 3.54. Performance Comparison for Feature Vector Variants of
Partitioned Kekre’s Wavelets Even-Odd Function Fingerprint Matching.
Score Fusion based Matching Gives Higher Performance this is Indicated
by Bar in Red Colours. (TRF: Transform, FV: Feature Vector)
Maximum PI is given by fusion of mean & density feature
vectors with DC & Sequency components , it is equal to 88% for
fingerprint core point ROI. The same for full fingerprint feature
vector gives 75% PI. It is seen that the performance of fingerprint
core point ROI based feature vector is higher as compared to full
fingerprint.
Next we compare the performance of all the techniques based on
partitioning of complex plane. Table 3.9 (a) & (b) and Fig. 3.55 give
comparison of Correct Classification Rates (CCR) of the fingerprint
Matching. CCR for fusion based feature vector is given.
RowTRF-CorePoint
ColTRF-CorePoint
Row-Density-CorePoint
Col-Density-CorePoint
Fusion-CorePoint
FV
RowTRF-Full
Finger
ColTRF-Full
Finger
Row-Density
-FullFinger
Col-Density
-FullFinger
Fusion- Full
FingerFV
PI 86 84 80 79 88 70 65 73 72 75
0102030405060708090
100
% P
I
Partitioned Kekre's Wavelets Even Odd Function Based Fingerprint Matching Performance Comparison of PI
115
Fig. 3.55. Performance Comparison for Accuracy (Correct Classification
Rate) of All the Transforms Discussed Above for Generation of
Partitioned Complex Plane in Transform Domain
Table 3.9
(a) Summary of Fingerprint Matching Tests for Partitioned Complex
Plane for Finger Core Point ROI Testing
Sr. Transform Successful Failure Total %Accuracy (CCR)
1 Walsh 457 73 530 86.20
2 Hartley 455 75 530 85.84
3 DCT 420 110 530 79.20
4 Kekre’s Transform 433 97 530 81.70
5 Kekre’s Wavelet 447 83 530 84.40
(b) Summary of Fingerprint Matching Tests for Partitioned Complex
Plane for Full Segmented Finger Testing
Sr. Transform Successful Failure Total %Accuracy (CCR)
1 Walsh 425 105 530 80.19
2 Hartley 410 120 530 77.36
3 DCT 403 127 530 76.04
4 Kekre’s Transform 411 119 530 77.55
5 Kekre’s Wavelet 406 124 530 76.60
86.2 85.84 79.2 81.7 84.4
80.19 77.36 76.04 77.55 76.6
0
10
20
30
40
50
60
70
80
90
100
Walsh Hartley DCT Kekre’s Transform Kekre’s Wavelet
%Accuracy (CCR)-Finger Core %Accuracy (CCR)-Full Finger
Partitioned Complex Plane Fingerprint Matching Testing Summary - % Accuracy (CCR) Comparison
116
Walsh transform based feature vectors give highest CCR followed by
Hartley Transform and Kekre’s Wavelets. DCT based feature vectors
has lowest accuracy for fingerprint matching. Specifically fingerprint
core point ROI based feature vectors have higher CCR as compared
to full segmented fingerprint. This is shown in Table 3.9 (a) & (b)
respectively. This shows the localizing core point or high curvature
point gives consistent texture pattern for feature vector extraction
and has higher matching rate.
3.1.5 Summary
In this chapter fingerprint matching using Kekre’s wavelets and
partitioned complex plane in transform domain of Walsh, Hartley,
Kekre’s Transform, DCT & Kekre’s wavelets is discussed. The
Kekre’s wavelets in the first approach are used for texture feature
extraction of the full fingerprint the Kekre’s wavelet based feature
vector gives 93% PI and 74% of correct classification ratio (CCR).
Fig. 3.56. Performance Comparison for Feature Vector Variants of
Partitioned Walsh, Hartley, DCT, Kekre’s Transform & Kekre’s Wavelet.
(Performance Index is Compared)
In another approach partitioned complex Walsh plane is used for
feature vector generation. Cal, Sal functions from intermediate
Walsh transform on rows and columns of fingerprint input image is
RowTRF-CorePoint
ColTRF-CorePoint
Row-Density -CorePoint
Col-Density -CorePoint
Fusion- CorePoint
FV
RowTRF-Full
Finger
ColTRF-Full
Finger
Row-Density -FullFinger
Col-Density -FullFinger
Fusion- Full
FingerFV
Walsh 85 88 80 82 95 80 70 66 69 81
Hartley 88 89 82 80 94 83 77 70 71 84
DCT 80 75 71 75 79 75 67 66 63 70
Kekre 76 84 80 81 83 64 72 68 69 81
KWT 86 84 80 79 88 70 65 73 72 75
0102030405060708090
100
Pe
rfo
rman
ce In
de
x
Performance Comparison of Walsh, Hartley, DCT, Kekre's Transform & Kekre's Wavelet Partitioned
Complex Plane in Transform Domain
117
used to find mean & density (Cal + jSal) points in complex Walsh
plane. These feature vectors are fused with DC & Sequency
information to achieve higher EER. First row in the table shown in
Fig. 3.56 shows EER for Walsh transform based feature vectors.
The complex Walsh Plane based on Cal & Sal functions of walsh
transform is applied to Hartley, DCT, Keker’s Transform & Kekre’s
Wavelets. The Cal & Sal functions here are referred as Even & Odd
Functions. The intermediate transform are used to generate
complex plane plots of (Even + jOdd) points. The table in Fig. 3.56
shows the PI comparison, Walsh & Hartley transforms have best PI
as well as Correct Classification Rate (CCR). This section has
discussed various correlation based fingerprint matching methods.
In the next sections different palmprint & Finger-knuckle print
matching algorithms will be discussed.