AbstractFingerprints are the oldest and most widely

used form of biometric identication. However, their

image contrast is poor due to skin conditions and
application of incorrect finger pressure. Fingerprint
enhancement is necessary to ensure the performance
and robustness of the algorithms for fingerprint
recognition. In this work, we present a fast fingerprint
enhancement method based on the diffusion tensor
which allows a better performance than a simple
gradient. The performance and efficiency of the
algorithm are estimated by calculating various quality
metrics and compared with the advanced met.
KeywordsFingerprint, Image enhancement, Tensor
Diffusion.
I. INTRODUCTION
Fingerprint is the most widely used today for
purposes of recognition, identification and
authentication. A fingerprint is an impression left
by the lines of the skin of the fingers or papillary on
an object. It is unique for each individual, formed
during the fetal period and retains the same form
throughout life. It is also immutable; it does not
change over time (except by accident). The
Automatic Fingerprint Identification System (AFIS)
depends on the comparison of these local ridge
characteristics and their relationships to make a
personal identification [10]. Finger prints have
distinctiveness and persistence, which are highly
desirable qualities for biometric applications.
However, finger print images are generally of low
contrast, due to skin conditions and application of
incorrect finger pressure. Also, this is why the
enhancement step is required before performing the
following steps in order to ensure robust
performance. Recently, many researchers are very
interested in the enhancement of images containing
locally oriented structures. Among these
researchers, we denote the nonlinear PDEs methods
involving anisotropic diffusion which has
significantly grown and become an important tool
in contemporary image processing. They will allow
an adaptive smoothness in constraint in the
denoising process by making a strong diffusion in
homogeneous zones and low diffusion across
boundaries. This eliminates the noise while
preserving the image discontinuities. Several
methods have been proposed by Perona and Malik
[1], which has since been expanded and improved
upon [11][12]. Also the method based on structure
tensor moreover, the method based on structure
tensor describes effectively the local variation of
the structures present in the image and provides a
fairly robust anisotropic diffusion.
This paper is organized as follows. Section 2,
addresses the main review for the nonlinear
diffusion EDPs. Section 3, devoted to the
presentation of the nonlinear diffusion tensor
formalism and its mathematical concept. Section 4,
describes the proposed approach. Section 5, gives
the experimental results that will illustrate the
performance of our method and finally section 6
concludes the work.

II. NONLINEAR DIFFUSION PDES
In this section, we propose a main review for some
basic mathematical concepts of the nonlinear
diffusion PDEs and their application for noise
removal and image enhancement. Consider a 2D
dimensional image domain
2
R and let a
grayscale intensity image with a diffusion time t be
represented by a bounded mapping
R t y x u : ) , , ( . The nonlinear PDE equation is
given by [1]:
) ) ( ( ) , , ( u u c div t y x
t
u
=

(1)

Where div: denotes the divergence operator, the
gradient operator and c(.) is the diffusivity function
that weights the gradient to control the diffusion
force and edge enhancement. This equation makes a
strong diffusion in areas with low gradient (low
probability of a contour) and low diffusion in areas
FINGERPRINT IMAGES ENHANCEMENT
USING DIFFUSION TENSOR
Feriel Romdhane, Faouzi Benzarti, Hamid Amiri
Signal, Image and Information Technology Laboratory (SITI), ENIT
[email protected], [email protected], [email protected]


U.S. Government work not protected by U.S. copyright


with high gradient (high probability of a contour)
where the potentials interesting discontinuities can
be found. Among the diffusion function c(.) used
we denote [1]:

2
) (

=
k
u
e u c (2)

This function introduces a parameter k which is
used to fix the boundary between the strong
gradients corresponding to the maintained
transitions and low gradients corresponding to the
noise. To better understand the behavior of Perona
and Malik approach (1) , it would be appropriate to
see this model in a local set ( ) , whose basis
vectors are respectively normal and tangential of
the contour (figure 1) [2] [3] [5].

Fig.1 Image contour in local set ( ) ,

The model can be written as :

( )
( )

u c u c
u u u
u
u
t
u

' '
'
+ =
+

(3)

It follows that for a contour of the image, the vector
is perpendicular (e.g. Gradient direction), and the
vector is oriented in the direction tangential to
the contour. The set ( , ) is an orthonormal
mobile basis. Near the contour, the image shows a
strong gradient. It is therefore preferable to
encourage the diffusion in the direction parallel to
the latter and not in the direction of the gradient, to
better preserve the discontinuity. This is equivalent
analytically to cancel the weight (e.g.

c = 0), and
to assume that the constant (e.g.

c = constant). On
the other hand, in the homogeneous regions, the
image has a low gradient. It is, therefore, preferable
to spread in all directions, which is equivalent to
setting weights

u and

u into a constant (e.g.

c =

c = constant). Despite the anisotropic

diffusion process in (1) has a good resistance
against the low noise variance but in large
irregularities the diffusion can no longer smooth.
And for the local structures of the image, this
process is not very strong, particularly in the
presence of local complexes orientations. In this
case, it is desirable using the diffusion tensor
formalism.

III. DIFFUSION TENSOR

A. Structure Tensor
In order to describe local variations structures
present in the image and to provide a fairly robust
anisotropic diffusion, we use the diffusion tensor.
Weickert [6] was interested in a definition of the
diffusivity tensor by using a structure tensor which
takes into account the orientation of the gradient
and the flow towards the orientation of interesting
features. It is in the form of a symmetric positive
semi-definite matrix constructed in the following
way [7]:
) (
T
u u K J

= (4)

With ) (
y x
u u K u =

: the smoothed version of
the gradient which is obtained by a convolution
with a Gaussian kernel

K .Where

K : a Gaussian
kernel with standard deviation .
So the structure tensor can be written as:

=
22 12
12 11
2
2
J J
J J
u K u u K
u u K u K
J
y y x
y x x

(5)

Its eigenvalues can be calculated as [6]:

+ + =
2
12
2
22 11 22 11 2 , 1
4 ) ( J J J J J

(6)

Where
2 1
and the corresponding orthonormal
set of eigenvectors ) , (
2 1
v v is given by [6]:

+ + +
+ +
+ + +
2
12
2 2
12
2
22 11 11 22
2
12
2
22 11 11 22
2
12
2 2
12
2
22 11 11 22
12
1
4 ) ) 4 ) ( (
4 ) (
4 ) ) 4 ) ( (
2
j j j j j j
j j j j j
j j j j j j
j
v

(7)
And
2 1
v v .
Its eigenvectors are parallel and orthogonal to

u
respectively. The eigenvector
1
v corresponds to the
larger eigenvalues
1
which defines the direction of
largest spatial change (e.g. The gradient
direction). The eigenvalues
1
and
2
provide also
useful information on the coherence of a structure.
The coherence can be measured by the following
equation:

2
12
2
22 11 2 1
4 ) ( J J J + = (8)

The homogeneous areas are characterized by
0
2 1
= = , and for 0
2 1
= >> we have a
straight edges and corners.

B. Coherence Enhancement Diffusion (CED)
The concept coherence enhancement diffusion has
the same basic model as (1), but its diffusion tensor
is defined according to a structure tensor, the
diffusion model becomes [7]:

( ) [ ] u J D div
t
u
=

(9)

Where D(.) is the diffusion tensor which is a
positive definite symmetric 2x2 matrix constructed
from the eigenvectors ) , (
2 1
v v of the structure
tensor which defines the directions and the
diffusion functions associated ) , (
2 1
presenting
the intensity of the action given by [7]:

( )

+
=
=
=

else e
if
k
1
0
2
2 1
2 1
2
1

(10)

With 0 > k and a small parameter ) 1 , 0 ( . The
diffusion tensor D is defined as:

( )
T
v
v
v v D

=
2
1
2
1
2 1

0
0

(11)

The method achieves a smoothing with constant
intensity set by a parameter in the direction of
the gradient. In the orthogonal direction, the
intensity of smoothing is controlled by the
coherence: the smoothing is more important in
areas where it is higher ) (
2 1
k > which
corresponds to the strongly oriented areas, such as
the edges of the structures. The behavior
highlighted is isotropic smoothing in areas of low
coherence and smoothing performed mostly along
structures when the coherence is stronger.

IV. PROPOSED APPROACH
The proposed approach is mainly based on the non
linear diffusion tensor inspired from the CED
approach in order to eliminate noise in
homogeneous areas and to enhance oriented
structures in the areas of contours. The diffusion
functions are as follows:

( ) ( )

=
=
=
else k
if
. tanh
0
2 1
2 1
2
1

(12)

With 0 > k and a small parameter ) 1 , 0 ( . The
hyperbolic tangent function guarantees that the
smoothness of the structure tensor carries over to
the diffusion tensor, and that
2
does not exceed 1
which is a typical scaling convention in nonlinear
diffusion filtering. Moreover, the positivity of
keeps the diffusion tensor ) (

J D uniformly
positive definite. Furthermore, the diffusion
equation (9) can be discretized by using the
formalism of finite differences approximation [13]
to the following iterative scheme [14]:
( ) [ ]
n n n
u J D div dt t j i u t j i u + = +
+
. ) ; , ( ) 1 ; , (
1

(13)
Our diffusion model allows for more respectful
smoothing for images with fine and discontinued
structures. Figure 2 summarizes the algorithmic
procedure of our model.





Fig. 2 Algorithm of the proposed method
In the following section,we will test and simulate
this model on synthetic images and on a fingerprint
images.

V. EXPERIMENTAL RESULTS

In order to evaluate the performance of the
proposed method, we strive first to degrade the
image tested by two types of noise. First, we
degraded a synthetic image with many curves
oriented (Fig.3-(a)) with a high Gaussian noise such
that the signal to noise ratio SNR = 19.32 dB
(Fig.3-(b)) and for 50 iterations. Then we degraded
the same image with a high speckle noise such that
SNR= 20.15 dB (Fig.4-(b)) and for 80 iterations.


(a) (b) (c)

(d) (e) (f)

(g) (h)
Fig.3 Simulation in the case of Gaussian noise, (a) Original
image, (b) Degraded image, (c) Median filter method, (d)Perona
and Malik method, (e) Catte and Al. method, (f) Alvarez and Al.
method, (g) Coherence enhancement diffusion method , (h)
Proposed method.


(a) (b) (c)

(d) (e) (f)

(g) (h)
Fig.4 Simulation in the case of Speckle noise, (a) Original
image, (b) Degraded image, (c) Median filter method, (d)Perona
and Malik method, (e) Catte and Al. method, (f) Alvarez and Al.
method, (g) Coherence enhancement diffusion method , (h)
Proposed method.
Yes
Input Image
Initialization of the parameters:
Niter , , et t
Creation of the Gaussian kernel:

=
2
2 2
2
2
2
1

y x
e K

Calculation of the structure
tensor:
) (
T
u u K J

=
Calculation of the eigenvalues:
1
and
2

Calculation of the diffusion
functions:
1
and
2

The diffusion equation:
( ) ( ) ( ) [ ]
n n n
u J D div t t j i u t j i u + = +
+
. ; , 1 ; ,
1

n Niter
End
No


Compared to the other methods, the results in figure
Fig.3-(h) and figure Fig.4-(h) shows a better picture
quality, particularly in terms of noise smoothing
and edge preservation. Diffusion follows
adequately the direction of the lines of curvature. It
should also be noted that our approach and the CED
method have similar performance.
To evaluate the performance of our method we used
different quantitative measures such as the peak
signal-to-noise-ratio (PSNR), the mean structural
similarity (MSSIM) index [8] and
Jpeg_quality_score index [9].
The PSN defined by:

=
MSE
N
PSNR
max
10
log 10 (14)
Where
max
N the maximum fluctuation in the input
image and MSE denote the mean square error given
by:
( ) ( ) j i f j i f
MN
MSE
r
N
j
M
i
, ,
1
0
1
0
1
0
=

=

=
(15)

Where ) , (
0
j i f the original image and ) , ( j i f
r
the
restored image.

Tab.1 Comparative table in the case of Gaussian noise
Method PSNR MSSIM Jpeg_quality_score
Median 11.09 0.45 6.41
Perona 11.67 0.52 5.18
Catte 12.05 0.53 6.35
Alvarez 10.78 0.39 6.48
CED 12.85 0.53 6.68
Proposed 12.98 0.54 6.87

Tab.1 Comparative table in the case of Speckle noise
Method PSNR MSSIM Jpeg_quality_score
Median 9.99 0.43 5.47
Perona 11.31 0.61 5.60
Catte 11.65 0.58 5.50
Alvarez 11.53 0.46 6.05
CED 12.47 0.63 6.13
Proposed 12.67 0.64 6.25

We note that the MSSIM approximates the
perceived visual quality of an image better than
PSNR. The Jpeg_quality_score is an index that
adopts an objective method without reference to
predict the perceived quality of images in a
deterministic way. The score is usually a value
between 1 and 10 (10 is the best quality, 1 poor
quality). The comparative tables (Tab.1 and Tab.2)
show that our method has the best scores of PSNR,
MSSIM and Jpeg_quality_score. We note that our
method is very robust to any kind of noise.
In the next experiment, we apply our method on a
real fingerprint image sized 235x182 pixels shown
in figure Fig.5-(a). To quantify the results, we use
only the index jpeg_quality_score (Jqs) since we
have no idea about the image reference. We note
that our method slightly exceeds the performance of
the CED method for enhancing oriented structures.
The features of the fingerprint are best preserved
and the quantitative measures show that a higher
value of the index jpeg_quality_score is in favor of
our method shown in the figure Fig.5-(g).


(a) (b)

(c) (d) (e)

(f) (g)


Fig.5 Enhancement fingerprint, (a) Original image (Jqs=7.14),
(b) Median filter method (Jqs=6.52), (c) Perona and Malik
method (Jqs=5.69), (d) Catte and Al. method (Jqs=7.12), (e)
Alvarez and Al. method (Jqs=7.13), (f) Coherence enhancement
diffusion method(Jqs=7.37), (g) Proposed method (Jqs=7.55)




VI. CONCLUSION

In this paper, we have proposed a fingerprint
images enhancement approach based on the
diffusion tensor which provides a fairly robust
anisotropic diffusion. The idea is to allow diffusion
along orientation of greatest coherence in the local
image structures. The experimental results are very
promising in terms of reducing noise while
preserving fingerprint features and structures.

REFERENCES

[1] Perona P., Malik J.Scale-space and edge detection using
anisotropic diffusion, IEEE Trans. Transactions on Pattern
Analysis and Machine Intelligence, PAMI, 1990,Vol
12(7),pp 629639 .
[2] R. Deriche, O. Faugeras,Les EDP en Traitement des Images
et Vision par Ordinateur, INRIA, Rapport de recherche n
2697, Novembre 1995.
[3] P. Kornprobst, R. Deriche, and G. Aubert,Nonlinear
Operators in Image Restoration, Proceedings of the
International Conference on Computer Vision and Pattern
Recognition, Puerto Rico, IEEE Computer Society, June
1997, 325-331.
[4] R. Terebes, O. Lavialle, P. Baylou, M. Borda, Directional
anisotropic diffusion, Proceedings of the XI-th European
Signal Processing Conference EUSIPCO2002, vol 2,
Toulouse, France, 2002, pp 266-269.
[5] Tschumperl D., Deriche R.Vector-valued image
regularization with PDE's, A common framework for
different applications. IEEE Transactions on Pattern
Analysis and Machine Intelligence,2005, Vol 27(4), pp 506-
517.
[6] J. Weickert. Anisotropic diffusion in image processing,
ECMI Series, Teubner-Verlag, Stuttgart, 1998.
[7] Weickert J., Scharr H. A scheme for coherence-enhancing
diffusion filltering with optimized rotation invariance, J.
Visual Comm. Imag. Repres, 2002, Vol 13,pp 103-118 .
[8] Wang Z., Bovik A.C, Sheikh H.R, Simoncelli E.P. Image
quality assessment: from error visibility to structural
similarity, IEEE Transactions on Image Processing, 2004,
Vol 13(4),pp 600612 .
[9] Zhou Wang, Hamid R. Sheikh and Alan C. Bovik, No-
reference perceptual quality assessment of JPEG
compressed images, Laboratory for Image and Video
Engineering (LIVE), The University of Texas at Austin,
Austin, TX 78712-1084, USA.
[10] H.C. Lee and R.E. Gaensslen, Advances in Fingerprint
Technology. New York, NY: Elsevier, 1991.
[11] F.Catt, P.L .Lions, J.M. Morel et T. Coll. Image selective
smoothing and edge detection. SIAM Journal on Numerical
Analysis, Vol. 29, n1, pp. 182-193,1992.
[12] L. Alvarez, P.L. Lions, J.M. Morel. Image selective
smoothing and edge detection by nonlinear diffusion .II.
SIAM Journal on Numerical Analysis, vol. 29, n3, pp. 845-
866, 1992.
[13] Garcia R., Deriche. R, Rousson. M, Lopez. C.A. Tensor
processing for texture and colour segmentation,
Scandinavian conference on image analysis SCIA.2005, Vol
3540, pp 1117-1127.
[14] Faouzi B. and Hamid A., Speckle Noise Reduction in
Medical Ultrasound Images, IJCSI International Journal of
Computer Science Issues, Vol. 9, Issue 2, No 3, March
2012.