Journal of Pattern Recognition Research 7 (2012) 155-174

Received Jul 18, 2011. Revised Jul 26, 2012. Accepted Sep 28, 2012.

A Multi-Level Model for Fingerprint Image Enhancement

Gabriel Babatunde Iwasokun [email protected]

Department of Computer Science
Federal University of Technology, PMB 704, Akure, Ondo State, Nigeria
Oluwole Charles Akinyokun [email protected]
Department of Computer Science
Federal University of Technology, PMB 704, Akure, Ondo State, Nigeria
Cleopas Officer Angaye [email protected]
National Information Technology, Development Agency (NITDA)
PMB 564, Area 11, Garki, Abuja, Nigeria
Olatubosun Olabode olabode [email protected]
Department of Computer Science
Federal University of Technology, PMB 704, Akure, Ondo State, Nigeria

Abstract
Fingerprint has continued to play very important roles in human identification and ver-
ification. A number of Automatic Fingerprint Identification Systems (AFIS) have been
developed for human security and control. Several yardsticks including the degree or level
www.jprr.org

of verification and identification are being considered for the determination of the sound-
ness and contribution of these systems. The well being of the fingerprint image in term
of quality as well as the adequacy level of the implemented algorithm are some of the
basic parameters which determine the performance level of AFIS in their assigned tasks.
In this paper, an existing mathematical algorithm for fingerprint image enhancement is
firstly modified before implementation. The motive behind the modification is to present
a faster and more accurate algorithm. The modified algorithm consists of sub-models
for the different stages of fingerprint image enhancement. During implementation, two
categories of images; namely synthetic images and real fingerprints images obtained from
the FVC2002 fingerprint database were used to determine the performance level of the
different sub-models and the whole algorithm. Obtained results reveal superiority in the
performance of the modified models over the original versions in terms of metric values,
completion time and quality. The results also show the significant importance of each
enhancement stage.
Keywords: AFIS, Pattern recognition, pattern matching, fingerprint, minutiae, image
enhancement

1. Introduction
Fingerprint is presently one of the essential variables used for the enforcement of security
and maintenance of reliable identification of individuals. Because of its uniqueness and
reliability, fingerprint is used for control of human access to highly secured places like
offices, equipment rooms, control centers and so on. The use of fingerprints for human
traffic control spans controlling of access to highly secured places like offices, equipment
rooms, control centers and so on. It is also used during voting, examination, operation of
bank accounts among others.
The result of the survey conducted by the International Biometric Group (IBG) in 2004
on comparison between the use of fingerprint and other biometrics is presented in Fig. 1.

c 2012 JPRR. All rights reserved. Permissions to make digital or hard copies of all or part of
this work for personal or classroom use may be granted by JPRR provided that copies are not
made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. To copy otherwise, or to republish, requires a fee and/or
special permission from JPRR.
 Iwasokun, Akinyokun, Angaye, Olabode

Fig. 1: Comparative survey of fingerprint with other

The result shows that a substantial margin exists between the uses of fingerprint for
identification over other biometrics such as face, hand, iris, voice, signature and middleware
[1]. The reasons attributed to this substantial margin include [1, 2, 3, 4]:

a. Wide variations in fingerprints from individual to individual.

b. Every fingerprint exhibits high degree of consistency due to the fact that changes in
scale does not necessarily translate to changes in appearance unlike in other biomet-
rics.

c. Every contact between the fingertip and the surface produces a fingerprint.

Similarly, the much larger market of personal authentication via fingerprints is attributed
to the following reasons:

a. There is a reasonable number of small and inexpensive fingerprint capture devices in

existence.

b. There are fast and low cost computing devices that support fingerprint based systems.

c. Presence of high speed recognition devices for meeting the needs of many applications.

d. The explosive growth of network and Internet transactions.

e. The increasing awareness of ease-of-use component for reliable security.

The most important part of fingerprints mostly used for different forms of identifications
and verifications are the features they possess. Fingerprint features exhibit uniqueness
from fingerprint to fingerprint. This uniqueness is defined by feature type, position and
orientation. Two types of feature exists; namely global and local features [5, 6, 7]. Global
features are those characteristics of the fingerprint that could be seen with the naked eye.
They are characteristics that capture the global spatial relationships of a fingerprint. Mostly
known global features are ridge pattern, type, orientation, spatial frequency, curvature,
position and count. Others are type lines, core and delta areas. The Local Features are
the minutiae consisting of the tiny, unique characteristics of fingerprint ridges mainly used

156
 A Multi-Level Model for Fingerprint Image Enhancement

for positive identification. They possess the information that is in a local area and are
invariant with respect to global transformation. Two or more impressions of a finger may
have identical global features with different local features (minutia points). An example is
shown in Fig. 2 with ridge patterns (a) and (b) representing different impressions of the
same finger (person). A local feature is read as bifurcation in (a) while it appears as a ridge
ending in (b).

(a) Bifurcation (b) Ridge ending

Fig. 2: Different minutiae for different impressions of the same finger

Section 2 of this paper presents the different stages of the modified fingerprint image
enhancement algorithm while Sections 3 and 4 present the experimental results and the
conclusion respectively.

2. Fingerprint Image Enhancement

Reliable and sound verification of fingerprints in any AFIS is always preceded with a proper
minutiae detection and extraction. A fingerprint image is firstly enhanced before the de-
tection and extraction of minutiae. A well enhanced image will provide a clear separation
between the valid and spurious minutiae. Spurious minutiae are those minutiae points that
are created due to noise or artifacts and they are originally not part of the fingerprint. This
paper adopts with major modifications, the algorithm implemented in [8, 9] for fingerprint
image enhancement. The overview of the algorithm is shown in Fig. 3. Its main steps
include image segmentation, local normalization, filtering and binarization/thinning.

Image Filtering

Image Image Local

Segmentation Normalization

Image Local Image Local

Normalization Normalization

Image Local Image

Normalization Binarization/Thinning

Fig. 3: The conceptual diagram of the fingerprint enhancement algorithm

157
 Iwasokun, Akinyokun, Angaye, Olabode

2.1 Image Segmentation

There are two regions that describe any fingerprint image; namely the foreground region and
the background region. The foreground regions are the regions containing the ridges and
valleys. As shown in Fig. 4, the ridges are the raised and dark regions of a fingerprint image
while the valleys are the low and white regions between the ridges. The foreground regions
often referred to as the Region of Interest (RoI) is shown for the image presented in Fig.
5. The background regions are mostly the outside regions where the noises introduced into
the image during enrolment are mostly found. The essence of segmentation is to reduce the
burden associated with image enhancement by ensuring that focus is only on the foreground
regions.

Background
region
Ridges

Valleys Foreground
region

Fig. 4: Ridges and valleys on a fingerprint image. Fig. 5: A fingerprint image and its foreground and
background regions.

The background regions possess very low grey-level variance values while the foreground
regions possess very high grey-level variance values. A block processing approach used in
[8, 9] is adopted in this research for obtaining the grey-level variance values. The approach
firstly divides the image into blocks of size W ×W and then the variance V (k) for each of
the pixels in block k is obtained from:

W W
1 XX
V (k) = (I(i, j) − M (k))2 (1)
W2
i=1 j=1
W W
1 XX
M (k) = J(a, b) (2)
W2
a=1 b=1

where I(i, j) and J(a, b) are the grey-level value for pixel (i, j) and (a, b) respectively in
block k.
2.2 Image Local Normalization
Normalization is performed on the segmented fingerprint image ridge structure so as to
standardize the degree of variations in the image grey-level values. By normalization, the
grey-level values are made to fall within certain range that is good enough for improved
image contrast and brightness. The first of the tasks of image normalization implemented
in [8, 9] and adopted for this research is the division of the segmented image into blocks of
size S×S. The grey-level value for each pixel is then compared with the average grey-level
value for the host block. For a pixel I(i, j) belonging to a block of average grey-level value

158
 A Multi-Level Model for Fingerprint Image Enhancement

of M , the result of comparison produced a normalized grey-level value N (i, j) defined by

the formula:

 M0 + V0 (I(i, j) − M )2 ∗ V −1 0.5 ,


if I(i, j) > M,
N (i, j) = (3)
 M − V (I(i, j) − M )2 ∗ V −1
0.5 , otherwise.
0 0

where an assumed value of M0 is set for the desired mean and an assumed value of V0 is
set for the desired variance.
2.3 Image Filtering
Normalized fingerprint image is filtered for enhancement through removal of noise and other
spurious features. Filtering is also used for preserving the true ridge and valley structures.
The fingerprint image filtering structure adopted for this research is in the following phases:
2.3.1 Orientation Estimation
Orientation estimation is the first of the prerequisites for fingerprint image filtering. In
every image, the ridges form patterns that flow in different directions. The orientation of a
ridge at location x, y is the direction of its flow over a range of pixels as shown in Fig. 6.

θ
(x,y)

Fig. 6: The orientation of a ridge pixel in a fingerprint

A modified version of the Least Square Mean (LSM) fingerprint ridge orientation estima-
tion algorithm proposed in [8, 9] was implemented in this research. The modified algorithm
involves the following steps:

a. Firstly, blocks of size S×S were formed on the normalized fingerprint image.

b. For each pixel, (p, q) in each block, the gradients ∂x (p, q) and ∂y (p, q) were computed as
the gradient magnitudes in the x and y directions, respectively. ∂x (p, q) was computed
using the horizontal Sobel operator while ∂y (p, q) was computed using the vertical
Sobel operator.

   
1 0 −1 1 2 1
 2 0 −2   0 0 0 
1 0 −1 −1 −2 −1

Horizontal Sobel Operator Vertical Sobel Operator

c. The local orientation of each pixel in a fingerprint image was computed by using its
S×S neighborhood in [8, 9]. This was modified in this research by dividing the image

159
 Iwasokun, Akinyokun, Angaye, Olabode

into S×S blocks and the local orientation for each block centered at pixel I(i, j) was
then computed from:

i+ S2 j+ S2
X X
Vx (i, j) = 2∂x (p, q)∂y (p, q) (4)
p=i− S2 p=i− S2

i+ S2 j+ S2
X X
Vy (i, j) = 2∂x2 (p, q) − ∂y2 (p, q) (5)
p=i− S2 p=i− S2

1 Vy (i, j)
θ(i, j) = tan−1 (6)
2 Vx (i, j)
where θ(i, j) is the least square estimate of the local orientation of the block centered
at pixel (i, j).

d. The orientation image is then converted into a continuous vector field defined by:

ϕx (i, j) = cos(2θ(i, j)), (7)

ϕy (i, j) = sin(2θ(i, j)), (8)

where ϕx and ϕy are the x and y components of the vector field, respectively.

d. Gaussian smoothing is then performed on the vector field as follows:

Sϕ Sϕ
2
X 2
X
ϕX (i, j) = G(p, q)ϕx (i − ps, j − qs), (9)
S S
p=− 2ϕ q=− 2ϕ

Sϕ Sϕ
2
X 2
X
ϕY (i, j) = G(p, q)ϕy (i − ps, j − qs), (10)
S S
p=− 2ϕ q=− 2ϕ

where G is a Gaussian low-pass filter of size Sϕ ×Sϕ .

f. The orientation field O of the block centered at pixel (i, j) is finally smoothed using
the equation:

1 ϕ′y (i, j)
O(i, j) = tan−1 ′ (11)
2 ϕx (i, j)

160
 A Multi-Level Model for Fingerprint Image Enhancement

2.3.2 Ridge Frequency Estimation

The second prerequisite for fingerprint image filtering is the ridge frequency estimation. In
any fingerprint image, there is a local frequency of the ridges that collectively form the
ridge frequency image. The ridge frequency is obtained from the extraction of the ridge
map from the image. The extraction of the ridge map involves the following steps:
a. Compute the consistency level of the orientation field obtained from the first prereq-
uisite in the local neighborhood of a pixel (p, q) with the following formula:

1
s X
Co (p, q) = |θ(i, j) − θ(p, q)|2 (12)
n2
(i,j)∈W

d, if d < 180,
|θ(i, j) − θ(p, q)| = (13)
d − 180, otherwise.

d = (θ(i, j) − θ(p, q) + 360) mod 360 (14)

where W represents the local neighborhood around (p, q), which is an n×n local win-
dow, θ(i, j) and θ(p, q) are local ridge orientations at pixels (i, j) and (p, q) respectively.
b. If the consistency level is below a certain threshold Fc , then the local orientations in
this region are re-estimated at a lower image resolution level until the consistency is
above Fc . After the orientation field is obtained, the following two adaptive filters are
applied to the image:

−2


 −(2πδ)−0.5 (e−δ ), if i = l(j) − d, j ∈ ρ,



ht (p, q, i, j) = −0.5 −δ −2 (15)
 (2πδ) (e ), if i = l(j), j ∈ ρ,



0, otherwise.


−2


 −(2πδ)−0.5 (e−δ ), if i = l(j) + d, j ∈ ρ,



−2
hb (p, q, i, j) = (2πδ)−0.5 (e−δ ), if i = l(j), j ∈ ρ, (16)




0, otherwise.


l(j) = j tan(θ(p, q)), (17)

−1
d = Y ∗ 2 cos θ(p, q) , (18)

 
ρ = Y | − 0.5 sin(θ(p, q)|, |0.5 sin(θ(p, q)| (19)
The two filters are capable of stressing under different condition the local maximum grey
level values along the normal direction of the local ridge orientation. The normalized image
is first convolved with these two masks, ht (p, q, i, j) and hb (p, q, i, j). If both the grey level
values at pixel (p, q) of the convolved images are larger than a certain threshold Fridge , then
pixel (p, q) is labeled as a ridge.

161
 Iwasokun, Akinyokun, Angaye, Olabode

2.3.3 Gabor Filtering

Having obtained the prerequisites, Gabor filtering is then used to improve or enhance the
fingerprint image to obtain a finer structure. Gabor filtering involves the removal of noise
and artifacts and its general form is:
n  o
G(x, y : f, θ) = exp 0.5 a2 ∗ δx−2 + b2 ∗ δy−2 cos(2πf a) (20)

where f is the frequency of the cosine wave along the direction θ from the x-axis, and δx
and δy are the space constants along x and y axes respectively. a = x sin θ + y cos θ and
b = x cos θ + y sin θ.
The values of the space constants δx and δy for the Gabor filters were empirically de-
termined as each is set to about half the average inter-ridge distance in their respective
direction. δx and δy are obtained from δx = kx F and δy = ky F respectively. F is the
ridge frequency estimate of the original image, and kx and ky are constant variables. The
value of δx determines the degree of contrast enhancement between ridges and valleys while
the value of δy determines the amount of smoothing applied to the ridges along the local
orientation.
2.4 Image Binarization/Thinning
The image obtained from the Gabor filtering stage is binarized and thinned to make it
more suitable for feature extraction. The method of image binarization proposed in [10]
is employed. The Method sets the threshold (T ) for making each cluster in the image as
tight as possible, thereby minimizing their overlap. To determine the actual value of T , the
following operations are performed on set of presumed threshold values:

a. The pixels are separated into two clusters according to the threshold.

b. The mean of each cluster are determined.

c. The difference between the means is squared.

d. The product of the number of pixels in one cluster and the number in the other is
determined.

The success of these operations depends on the difference between the means of the
clusters. The optimal threshold is the one that maximizes the between-class variance or,
conversely, the one that minimizes the within-class variance. The within-class variance of
each of the cluster is then calculated as the weighted sum of the variances from:

2 2 2
σwithin (T ) = nB (T )σB (T ) + nO (T )σO (T ) (21)

T
X −1
nB (T ) = p(i) (22)
i=0

N
X −1
nO (T ) = p(i) (23)
i=T

where:
2 (T ) = the variance of the pixels in the background (below) threshold,
σB

162
 A Multi-Level Model for Fingerprint Image Enhancement

σO2 (T ) = the variance of the pixels in the foreground (above) threshold,

p(i) is the pixel value at location i, N is the intensity level and [0, N − 1] is the range of
intensity levels.
The between-class variance, which is the difference between the within-class variance and
the total variance of the combined distribution, is then obtained from:

2
σbetween (T ) = σ 2 − σwithin
2
(T ) (24)

2
σbetween (T ) = nB (T )[A] + nO (T )[B] (25)

A = (µB (T ) − µ)2 (26)

B = (µO (T ) − µ)2 (27)

where σ 2 is the combined variance, µB (T ) is the combine mean for cluster T in the back-
ground threshold, µO (T ) is the combine mean for cluster T in the foreground threshold and
µ is the combined mean for the two thresholds. The between-class variance is simply the
weighted variance of the cluster means themselves around the overall mean. Substituting
µ = nB (T )µB (T ) + nO (T )µO (T ) into Equation 25, the result is:

2
 2
σbetween (T ) = nB (T )nO (T ) µB (T ) − µO (T ) (28)
Using the following simple recurrence relations, the between-class variance is successfully
updated by manipulating each threshold T using a constant value p as follows:

nB (T + 1) = nB (T ) + p (29)
nO (T + 1) = nO (T ) − p (30)
−1
µB (T + 1) = (µB (T )nB (T ) + pT ) ∗ (nB (T + 1)) (31)
µO (T + 1) = (µO (T )nO (T ) − pT ) ∗ (nO (T + 1))−1 (32)

3. Experimental Results
A modified version of the fingerprint enhancement algorithm proposed in [8, 9] was imple-
mented in this research by using MATLAB Version 7.2 on the Windows Vista Home Basic
operating system. The experiments were performed on a Pentium 4 - 1.87 GHz processor
with 1024MB of RAM. The purpose of the fingerprint enhancement experiments is to ana-
lyze the performance of the modified algorithm under different conditions of images as well
as generate the metrics that could serve the basis for the comparison of the results from
the research with results from related works. Two sets of experiment were conducted for
the performance analysis. The first set of experiments was on synthetic images. The orien-
tation estimation, ridge frequency estimation and Gabor filtering experiments all employed
the circsine function [11] to generate the synthetic images. Where necessary, the MATLAB
imnoise function was also used to generate noise and artifacts on synthetic images. The
second set of experiments was on the FVC2002 fingerprint database. Fig. 7(a), 7(b) and
7(c) are synthetic images of size 200×200 and wavelength 8. They were obtained with
the imnoise function using the salt and pepper noise level of 0, 0.2 and 0.4 respectively.
The results of the ridge orientation estimation experiments on each of the three images are

163
 Iwasokun, Akinyokun, Angaye, Olabode

shown in Fig. 7(d), 7(e) and 7(f) respectively. These results show that for noise levels of
0 and 0.20, the modified ridge orientation algorithm effectively generated ridge orientation
estimates that are very close to the actual orientations. However, for image with noise level
of 0.4, the result shows a substantial number of ridge orientation estimates that signifi-
cantly differ from the actual orientations. These results show that the performance of the
algorithm depends on the image noise level. When the noise level on the image is within
reasonable range, the algorithm does well while it fails when the noise level rises beyond
the threshold which was found to be 0.29.

(a) Synthetic Image with 0 noise level (b) Synthetic Image with 0.2 noise (c) Synthetic Image with 0.4 noise
level level

(d) Orientation estimate for (a) (e) Orientation estimate for (b) (f) Orientation estimate for (c)

(g) Ridge frequency estimate for (a) (h) Ridge frequency estimate for (b) (i) Ridge frequency estimate for (c)

Fig. 7: Orientation and ridge frequency estimates for synthetic images of different noise levels

Fig. 7(g), 7(h) and 7(i) present the results of the ridge frequency estimation experiments
on the synthetic images of different noise levels shown in Fig. 7(a), 7(b) and 7(c) respec-

164
 A Multi-Level Model for Fingerprint Image Enhancement

tively. In the ridge frequency estimates shown in Fig. 7(g) and Fig. 7(h), there is uniformity
between most of the estimated frequency values for each 32×32 block as against Fig. 7(i)
which shows irregular patterns. Fig. 7(g), 7(h) and 7(i) produced MSE value of 0.0006,
0.0017 and 0.0077 respectively. The ridge frequency estimates and the increasing MSE val-
ues reveal that the performance of the ridge frequency estimation algorithm decreases with
increase in the image noise level. Generally, under noise level of 0.30, it was discovered
that the algorithm produced relatively uniform ridge frequency estimates as shown in Fig.
7(g) and 7(h). When the noise level equals or exceeds 0.3, the orientation estimates shows
non-uniformity as shown in Fig. 7(i). Since the performance of the ridge frequency estima-
tion algorithm depends significantly on the performance of the ridge orientation estimation
algorithm, the failure of the orientation estimation algorithm explains the failure of the
ridge frequency estimation algorithm for noise level above 0.29.
The efficiency of the ridge orientation and frequency estimation algorithms were quantita-
tively measured by estimating the Mean Square Error (MSE) that represents the difference
between the estimated and actual values in radians. Mean square error of the ridge orien-
tation estimates for the synthetic image shown in Fig. 7(a) under different conditions of
noise for both the pixel processing approach in [8] and the block processing approach are
shown in Table 1.
Table 1: Comparison of Mean Square Error for ridge orientation estimates.

Mean Square Error

Noise level
Raymond [8] Current study
0.00 0.0003 0.0000
0.05 0.0009 0.0006
0.10 0.0032 0.0008
0.15 0.0102 0.0019
0.20 0.0246 0.0015
0.25 0.0691 0.0026
0.30 0.1722 0.0405
0.35 0.2330 0.0521
0.40 0.3041 0.0661
0.45 0.4124 0.0803
0.50 0.4262 0.1085

The bar chart of these values is presented in Figure 8. Table and Figure 8 indicate
increasing mean square errors for increasing noise level in both cases. This indicates that
the accuracy of the two approaches decreases with increase in the noise level. It is also
revealed that the orientation estimate is closer to the actual value for the block processing
approach at any noise level. With lower standard deviation of 0.0393 for its MSE values, it
is established that the block processing approach performs better than the pixel processing
approach with MSE values with standard deviation of 0.1686.
This higher performance is attributed to the fact that in the block processing approach,
the degree of variation in the orientation estimates for pixels in a block is reduced to zero
as each pixel assumed the orientation estimate for the center pixel of its host block. This
significantly increases the ability of the algorithm to estimate the ridge orientation close to
the actual value. The MSE results of the implementation of the ridge frequency estimation
algorithm on the synthetic image shown in Fig. 7(a) under different conditions of noise in
[8] and the current study are shown in Table 2. The bar chart of these values is shown in

165
 Iwasokun, Akinyokun, Angaye, Olabode

0.45

0.4

0.35

0.3

MSE value
0.25

0.2 Raymond
0.15 Current study

0.1

0.05

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Noise level

Fig. 8: Bar chart of MSE ridge orientation values of Table 1

Figure 9. It is revealed from Table 2 and Figure 9 that there is increase in the MSE values
as the noise level increases for the two cases. This translates to diminishing performance
due to increasing margin between the actual and the estimated ridge frequency values. It
is also shown that the MSE value is significantly lower in all cases for the current study
than in [8]. With lower standard deviation of 0.0032 for its MSE values, it is established
that the implementation of the ridge frequency estimation algorithm in the current study
yields better results than its implementation in [8] which yielded a standard deviation of
2.8514 for its MSE values. This improvement is also attributed to the superior performance
of the block processing approach to ridge orientation estimation over the pixel processing
approach.
Table 2: Comparison of Mean Square Error for ridge frequency estimates.

Mean Square Error

Noise level
Raymond [8] Current study
0.00 0.0100 0.0006
0.05 0.0211 0.0009
0.10 0.0465 0.0011
0.15 0.0820 0.0012
0.20 0.1702 0.0017
0.25 1.1149 0.0033
0.30 2.0229 0.0041
0.35 4.1149 0.0060
0.40 5.8543 0.0077
0.45 6.1098 0.0080
0.50 7.2616 0.0090

The performance of the Gabor filtering algorithm on a zero, medium and high quality
synthetic image of size 410×410 and wavelength 15 is presented in Figs. 10(d), 10(e) and
10(f) respectively. Parameter values of kx = 0.45 and ky = 0.45 were used to obtain these
results. The results presented in Fig. 10(d) and 10(e) reveal that with zero or medium
level noise density, the filter suitably removed the noise from the image and enhanced it

166
 A Multi-Level Model for Fingerprint Image Enhancement

MSE value
4
Raymond
3
Current study
2

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Noise level

Fig. 9: Diagram of MSE ridge orientation values of Table 2

to a level that is close to the original image. This is partly due to the previous accurate
estimation of the ridge orientation and the ridge frequency for zero or medium noise level
images. However, the experimental result presented in Fig. 10(f) reveals that when the
filter is applied to images with higher noise level, it could not remove the noise effectively as
it produced a significant amount of spurious elements. This failure is due to the inaccurate
estimation of the ridge orientation and the ridge frequency at higher noise levels.
When experiments were performed on real fingerprint images, just like cases in [8], the
best results were obtained for image segmentation with variance threshold of 100. This
threshold value provided the best segmentation results in terms of differentiating between
the foreground and the background regions as shown in Fig. 11(b). The results of segmenta-
tion using threshold value of 95 and 105 are presented in Fig. 11 (c) and 11(d) respectively.
These results show inappropriate segmentation due to inaccurate variance thresholds. With
lower threshold of 95, some of the background regions have been segmented to the foreground
while some foreground regions are also segmented to the background under higher thresh-
old of 105. The result of the normalization experiment on the fingerprint images shown in
Fig. 11(a) is presented in Fig. 12(a). The desire mean of zero and variance of one used
in [8] were adopted and used to normalize the ridges in the images. During normalization,
all positions are evenly shifted along the horizontal axis, which makes the structure of the
ridges and valleys to become well and suitably positioned.
The histogram plots of the original and the normalized images are shown in Fig. 12(b)
and 12(c) respectively. The histogram plot of the original image shows that all the intensity
values of the ridges show irregular frequency values and also fall within the right hand
side of the 0-255 scale, with no pixels in the left hand side. This leads to an image with
a very low contrast. The histogram plot of the normalized image shows that the range
of intensity values for the ridges has been adjusted between 0-0.5 scale such that there is
a more evenly and balanced distribution between the dark and light pixels and that the
ridge frequencies fall within close values. The normalized image histogram plots also show
that the normalization process does distribute evenly the shape of the original image. The
positions of the values are evenly shifted along the x-axis, which means the structure of

167
 Iwasokun, Akinyokun, Angaye, Olabode

(a) 0 noise level image (b) 0.2 noise level image (c) 0.4 noise level image

(d) Filtered image of (a) (e) Filtered image of (b) (f) Filtered image of (c)

Fig. 10: Results of applying a Gabor filter on synthetic images of different noise levels.

(a) Original Fingerprint (b) Segmented Finger- (c) Segmented Finger- (d) Segmented Finger-
image print image with vari- print image with vari- print image with vari-
ance threshold of 100 ance threshold of 95 ance threshold of 105

Fig. 11: Results of segmentation with different threshold.

the ridges and valleys are now well and suitably positioned. This shifted and improved
positioning lead to images with a very high contrast shown in Fig. 12(a).
The orientation fields for the real fingerprint images were obtained around their singular
points since they are prominent features used by any AFIS for fingerprint classification and
matching. Good quality images are shown in Fig. 13(a), 13(b) and 13(c). Their orientation
estimates are shown in Fig. 13(e), 13(f) and 13(g) respectively. At the singular points,
the orientation field is discontinuous and unlike the normal ridge flow pattern, the ridge
orientation varies significantly. From these results, it is observed that there exists no devi-
ation between the actual fingerprint ridge orientation and the estimated orientation of the

168
 A Multi-Level Model for Fingerprint Image Enhancement

(a) Normalized (b) Histogram of original image (c) Histogram of Normalized image
image

Fig. 12: Normalized image and the histogram plots.

vectors. In both cases, the algorithm produces accurate estimates of the orientation vectors
such that they flow smoothly and consistently with the direction of the ridge structures in
the images. In the superimposed version of image in Fig. 13(h), the contrast of the original
image is lowered. This was done to improve the visibility of the orientation vectors against
the background so that misleading regions can be easily identified. The ridge orientation
estimate for poor quality image shown in Fig. 13(d) is presented in Fig. 13(h). The esti-
mate indicates a fairly smooth orientation field in some well defined regions while it gives
misleading results in areas of very poor quality as evident in the bottom regions of the
estimate.
The orientation estimates resulting from the pixel processing approach in [8] and the block
processing approach of the current study for the image shown in Fig. 14(a) are presented
in Fig. 14(b) and 14(c) respectively. Visual inspection of these results reveals that the two
methods did well in the ridge orientation estimation. However, the orientation is observed
to be closer to the actual orientation in block processing than pixel processing in some
regions especially the core areas represented by the inserted circles. The reason adduced to
this is that assigning equal orientation estimate for pixels in a block rather than maintaining
different values is better and able to take the estimates closer to their actual values.
The ridge frequency estimates for real fingerprint images shown in Fig. 13(a), 13(b), 13(c)
and 13(d) are shown in Fig. 15(a), 15(b), 15(c) and 15(d) respectively. The mean ridge
frequency (MRF), which is the average of the image ridge frequencies, is also presented for
each image. It is noted that the MRF values differ for all the images. This difference is
attributed to the fact that fingerprints exhibit variation in their average ridge frequency
characteristics and contrast levels. The intensities of frequency differ for blocks or regions
within same image. Some blocks or regions exhibit high contrast while others exhibit low
contrast. Based on these, the synthetic images are more appropriate for the evaluation of
the accuracy of the ridge frequency estimation algorithm.
Fig. 16 reveals the extent to which the filtering algorithm was able to remove noise
from the images shown in Fig. 13 for different values of kx and ky. The results shown in
Fig. 16(a), 16(b), 16(c) and 16(d) were obtained using parameter values of kx = 0.45 and
ky = 0.45. With these values, it is shown that the contrast level between ridges and valleys

169
 Iwasokun, Akinyokun, Angaye, Olabode

(a) Right Loop image (b) Arch image (c) Whorl image (d) Poor quality image

(e) Orientation esti- (f) Orientation esti- (g) Orientation esti- (h) Orientation esti-
mate of (a) mate of (b) mate of (c) mate of (d)
Fig. 13: Fingerprint images and their orientation estimates.

(a) Original Image (b) Orientation estimate (c) Orientation estimate

obtained in [8] for current study

Fig. 14: Orientation estimates for pixel and block processing approaches.

for each of the images is neither too high nor too low. Appropriate level of smoothing
is also applied to the ridges along the local orientation. With lower parameter values of
kx = 0.40 and ky = 0.40, it is shown in Fig. 16(e), 16(f), 16(g) and 16(h) that the contrast
level between ridges and valleys is too low and this explains why there is a number of dark
regions. The degree of smoothing is also poor as there is a good number of overlapping
ridges in the filtered images. When kx = 0.5 and ky = 0.5 were used, the results of the
filtering experiments shown in Fig. 16(i), 16(j), 16(k) and 16(l) reveal that there is no

170
 A Multi-Level Model for Fingerprint Image Enhancement

(a) Ridge frequency estimate (b) Ridge frequency estimate

of Fig. 13(a) with mean ridge of Fig. 13(b) with mean ridge
frequency of 0.0148 for all the frequency of 0.0951 for all the
32×32 blocks 32×32 blocks

(c) Ridge frequency estimate (d) Ridge frequency estimate

of Fig. 11(c) with mean ridge of Fig. 13(d) with mean ridge
frequency of 0.0685 for all the frequency of 0.0881 for all the
32×32 blocks 32×32 blocks

Fig. 15: Ridge frequency estimate of selected images.

significant difference from the results obtained for kx = 0.45 and ky = 0.45 except that
some regions described by circles appear to be excessively smoothened.
It is therefore stated that based on the modified algorithm, parameter values of kx = 0.45
and ky = 0.45 are most appropriate for image filtering as against kx = 0.50 and ky = 0.50
proposed in [8, 9]. This reduction in parameter values is due to the better performance of
the block processing approach in the orientation and ridge frequency estimations as attested
to by the results in Table 1 and Table 2. Results presented in Fig. 16 show that the filtering
algorithm is able to smoothen to a fine level with appropriate parameter values for good
quality fingerprint images. When the quality degrades like the one shown in Fig. 13(d),
the performance of the algorithm diminishes as it produces images with inappropriately
filtered regions as shown in Fig. 16(d). This is also corroborated with reasons adduced to
the values presented in Table 1 and Table 2.
The results of the binarization experiments for the images shown in Fig. 13 are presented
in Fig. 17. Visual inspection of the results shows that the binarization algorithm perfectly
separated the ridges (black pixels) from the valleys (white pixels). To obtain these results,
the grey-level value of each pixel in the filtered image is examined and, if the value is greater
than the threshold value 1, then the pixel value is set to a binary value one; otherwise, it
is set to zero. The threshold value successfully made each cluster as tight as possible and
also eliminated all overlaps. The threshold value of 1 was taken after a careful selection

171
 Iwasokun, Akinyokun, Angaye, Olabode

(a) (b) (c) (d)

Filtered images for kx = 0.45 and ky = 0.45

(e) (f) (g) (h)

Filtered images for kx = 0.40 and ky = 0.40

(i) (j) (k) (l)

Filtered images for kx = 0.50 and ky = 0.50

Fig. 16: Filtered images with different values for kx and ky .

(a) (b) (c) (d)

Fig. 17: Results of Binarization for images shown in Fig. 13(a), 11(b), 11(c) and 11(d).

from a series of within and between class variance values ranging from 0 to 1 that optimally
supported the maximum separation of the ridges from the valleys. The clear separation of
the ridges from the valleys verifies the correctness of the algorithm as proposed in [9] and
implemented in this research.
The results of the thinning experiment on each of the images shown in Fig. 13 are
presented in Fig. 18(a-d). The MATLAB’s bwmorph operation using the ’thin’ option was
used to generate the thinned images. These results show that the ridge thickness in each of
the images has been reduced to its smallest form or skeleton (one pixel wide). It is also shown

172
 A Multi-Level Model for Fingerprint Image Enhancement

(a) (b) (c) (d)

Thinned images with the enhancement stages

(e) (f) (g) (h)

Binarized images without the enhancement stages

(i) (j) (k) (l)

Thinned images without the enhancement stages
Fig. 18: Thinned and Binarized images.

that the connectivity of the ridge structures is well preserved. Fig. 18(e-h) shows the results
of performing binarization experiments on the raw images without the enhancement stages.
In contrast to Fig. 17(a-d), the binary images in Fig. 18(e-h) are not well connected and
contain significant amount of noise and corrupted elements. Consequently, when thinning is
applied to these binary images, the results in Fig. 18(i-l) show that the accurate extraction
of minutiae would not be possible due to the large number of spurious features produced.
Thus, it is shown that employing the image enhancement stages prior to thinning is effective
for accurate and speedy extraction of minutiae.

4. Conclusion
This paper discusses the results of the modification and verification of the fingerprint en-
hancement algorithm developed and implemented in [8, 9]. Some stages of the algorithm
were slightly modified for improved performance. For instance, block processing approach
was introduced into the orientation estimation algorithm in place of the pixel processing ap-
proach. While the pixel processing approach subjects each pixel in the image to orientation
estimation, the block processing approach firstly divides the image into S×S blocks and
obtains the orientation estimate for the center pixel. This resulted in higher performance
as attested to by the Tables of MSE for ridge orientation and frequency estimates. Param-

173
 Iwasokun, Akinyokun, Angaye, Olabode

eter values of kx = 0.45 and ky = 0.45 were found to perform well in the image filtering
experiments as against kx = 0.5 and ky = 0.5.
The results of the experiments conducted for image segmentation, normalization, ridge
orientation estimation, ridge frequency estimation, Gabor filtering, binarization and thin-
ning on synthetic and real fingerprint images reveal that with free or minimal noise level,
the algorithms perform well. Improved performance is specifically noticed for the modified
ridge orientation estimation algorithm. It is also established that each stage of the enhance-
ment process is important for obtaining a perfectly enhanced image that is acceptable and
presentable to the features extraction stage. The results obtained from the final stage of
thinning show that the connectivity of the image ridge structure has been preserved and
improved at each stage.

References
[1] Roberts C., Biometrics, Accessed: July, 2009
http://www.ccip.govt.nz/newsroom/informoation-notes/2005/biometrics.pdf
[2] Michael C. and Imwinkelried E., “Defence practice tips, a cautionary note about fingerprint
analysis and reliance on digital technology”, Public Defense Backup Centre Report, 2006.
[3] Palmiotto M. J., “Criminal Investigation”, Chicago: Nelson Hall, 1994.
[4] Salter D., “Fingerprint - An Emerging Technology”, Engineering Technology, New Mexico State
University, 2006.
[5] Tsai-Yang Jea, and Venu Govindaraju, “A minutia-based partial fingerprint recognition sys-
tem”, Center for Unified Biometrics and Sensors, University at Buffalo, State University of
New York, Amherst, NY USA 14228, 2006
[6] Akinyokun, Angaye O. C. and Iwasokun G. B., “A Framework for Fingerprint Forensic”, Pro-
ceeding of the First International Conference on Software Engineering and Intelligent System,
organized and sponsored by School of Science and Technology, Covenant University, Ota, Nige-
ria, 2010, pages 183-200.
[7] Akinyokun O. C. and Adegbeyeni E. O., “Scientific Evaluation of the Process of Scanning and
Forensic Analysis of Fingerprints on Ballot Papers”, Proceedings of Academy of Legal, Ethical
and Regulatory Issues, Vol. 13, Numbers 1, New Orleans, 2009.
[8] Hong Lin, Wan Yifei and Anil Jain, “Fingerprint image enhancement: Algorithm and per-
formance evaluation”, Pattern Recognition and Image Processing Laboratory, Department of
Computer Science, Michigan State University, 2006, pp1-30
[9] Raymond Thai, “Fingerprint Image Enhancement and Minutiae Extraction”, PhD Thesis Sub-
mitted to School of Computer Science and Software Engineering, University of Western Aus-
tralia, 2003.
[10] Liang, Xu, “Image Binarization using Otsu Method”, Proceedings of NLPR-PAL Group CASIA
Conference, 2009, pp345-349
[11] Kovesi P., “MATLAB functions for computer vision and image analysis”, School of Computer
Science and Software Engineering, University of Western Australia, Accessed: 20 February 2010
http:/www.cs.uwa.edu.au/~pk/Research/MatlabFns/Index.html

174