You are on page 1of 14

“Neural Networks - New Trend Net works”

Presented by,
V.V.N.S.Kishore ( kishorevelagapudi@gmail.com)
S.Chandra Harish (chandraharish_485@yahoo.co.in)

Q.I.S. College of Engineering & Technology (II/IV E.C.E.)


Ongole.
FINGERPRINT RECOGNITION BY
NEURAL NETWORKS
“Neural Networks - New Trend Networks”

Abstract
This paper presents a Fingerprint recognition
system developed using neural network approach. Neural networks are fascinating due to
their outstanding ability in identification and classification of images. They outperform
the traditional approaches due to their ability of nonlinear mapping of real-time events.
Due to its massive parallel structure, it takes less computational time, to recognize
fingerprints, than the traditional approaches. The only time it takes is for its training. The
main reason why a common man today is not using neural networks and relying only on
conventional techniques is the huge size of the network, because of the large number of
inputs. Hence to reduce the dimensionality a new technique is presented which makes use
of cross-correlation of two signals. A known set of bifurcations are taken and they are
used in the cross-correlation with the fingerprint image. The Fourier transform the
resulting image gives the details of the location and distribution of the bifurcations in the
fingerprint. This new approach makes not only biometrics but also neural networks take
new steps in its application areas. A three layer feedforward back-propagation network,
due to its high performance and wide popularity is used for recognition.

Index Terms
Bifurcations, cross-correlation, feedforward back-propagation neural network, Fourier
Transform, Edge detection.
I. INTRODUCTION
Artificial Neural Networks are biological emulations of human nervous systems. They are
best known as universal approximators in approximating a function. It is their highly non-
linear mapping ability which makes them find applications in many areas like pattern
recognition and classification, speech recognition, function approximation, prediction of
currency exchange rates, sonar signal processing, maximizing the yields of chemical
processes and in medical diagnosis.
Fingerprint recognition and classification is another new application for
neural networks. Unlike the traditional methods there is no need of extracting features
from the image explicitly. The network itself does this job. The hidden layers of a
feedforward backpropagation network project the input data into different clusters in a
way these different clusters can be aggregated into different classes. This is how the non-
linear mapping of input data to desired outputs is performed.
Though the network can be easily designed to work with, there is always
the problem of huge size of the network. To reduce the number of inputs to the network
a new technique is presented which makes use of cross-correlation of two images or
signals. It is a technique to find the known target in an image. The correlation of input
dataset with each bifurcation gives us a set of images in which the aerial location or
distribution of the respective bifurcation in the image.
Fingerprints are used for personal identity from the ancient days. The use
of computers in fingerprint recognition is desirable in many applications, such as
building or area security and police work to identify criminals. Recent advances in
automated finger print classification techniques have been investigated. But this is a
different method from conventional approach. Fingerprints are unique, permanent and
unchanging in nature. The probability that any two people in the world can have their
fingerprints alike is 1 in 1.9x1015. Hence they are used from a long time. A fingerprint
is the representation of the epidermis of a finger. At a macroscopic analysis, a
fingerprint is composed of a set of ridge lines which often flow parallel and sometimes
produce local macro-singularities called whorl, loop and delta. At a finer analysis other
very important features can be discovered in the fingerprint patterns. These are called
minutiae, are essentially determined by the termination or the bifurcation of the ridge
lines. A total of 16 bifurcations are selected in this method. These bifurcations play an
important role in fingerprint recognition because the in the next section their use to
reduce the network size is discussed. Unlike traditional approaches we need have to
train the network and test it. The neural network approaches don’t have template
matching things. Hence this approach can best be used in many access control systems
as well as in police services to find out criminals using their latent fingerprints.

Figure 1.Bifurcation and terminate Figure 2.Different bifurcations observed and applied.

II. Preprocessing of Fingerprints

One of the main reasons why a common man is not using neural networks for his applications
is the huge size of neural network. The feedforward networks don’t need any feature extraction
but they need to get sufficient input data depending on the various factors such as number of
patterns, noisy patterns, rotated and shifted patterns. Hence to reduce the input size sufficient
enough for classification of specific data in hand, the image has undergone following processes
and finally gave a small input dataset. There are several methods for to accomplish this task[3],
[2], [1]. But after experimenting with each method it is found that those methods are not
suitable for this specific application because those methods are used in different applications
like one-dimensional object recognition where Fourier descriptors of boundary position data
are taken [3], paper currency recognition where random masks are used [2] and a few other
methods listed in [1]. But a careful study by applying Fourier transform (2-D) to
fingerprintsled to the following method. The segmented image or edge detected image is then
subjected to cross correlation with a known set of bifurcations. For that purpose the Fourier
transform of fingerprint and that of 1800 rotated bifurcation are taken, multiplied and subjected
to inverse Fourier transformation. An image is obtained in which the positions where a pattern
similar to a bifurcation, are not changed and others are attenuated. This image has information
about the distribution of the bifurcations in the original image. The Fourier transformation of
this image gives the global information about how often the bifurcation repeats itself in the
entire area of the image [4]. Here the Fourier transform means two-dimensional Fourier
transform only, because here it is dealing with images which are two-dimensional. The starting
values of the Fourier transform give the low frequency details of the image i.e., information[3].
But as we are taking the Fourier transform of cross-correlated image and not of the image
itself. We need only overall distribution of bifurcations. Also the frequency part is susceptible
to noise. Hence only a few starting values are sufficient. Depending on the number of
bifurcations different number of values can be taken. Hence the first 7x7 matrix values are
chosen after experimenting with various number of values. For given fingerprint, the feature
vector is selected adding all the matrices produced by cross-correlation with each bifurcation.
Thus the dimensionality is reduced to a low value depending on the specificity of the problem.
These are used as inputs for the network.
III. STRUCTURE OF THE NETWORK AND TRAINING METHOD
A three-layer feedforward network is chosen for this application. The input layer consisted of
48 neurons, hidden layer consisted of 16 neurons and output layer has 4 neurons which
correspond to recognition patterns. The structure of the network is shown in the following
figure.

The input data is obtained from the fingerprint image by the procedure cited in the previous
section. The network is trained using gradient descent and momentum back-propagation
method. In this method the error in the output is back propagated and hence called back-
propagation method. The training rule [5] is summarized below:
We define output neuron error
ejp = tjp - ojp
where tjp=desired output and ojp=actual output of neuron
A measure of output error based on the pth training sample is denoted Ep and defined as
Ep = ( ∑ | ejp |2 ) /2
This error is minimized through the change of weights by propagating it backwards to the
input layer. Let the activation function is represented for the jth output neuron a
ojp = f (wji, ijp)
where f is sigmoidal activation function and is same for neurons in the network,
f(netj) = 1 / (1 + e-netj)
wji is the weight or strength of connection from neuron i in previous layer to neuron j in the
present layer. The equations of this method are summarized as:
∆pwji = ε δjp oip
where ε is called the learning rate, oip is output of ith neuron and δjp ‘sensitivity of the pattern
error’ is given by
for output neuron
δjp = (ejp) f’(netjp)

f’(netjp) = ojp (1 – ojp)


for hidden layer neuron
δjp = f’(netjp) ∑n δnp wnj
δnp are from the previous layer and wnj are weights in its next layer.
Adding momentum to the training procedure
∆pwji(n + 1) = ε δjp oip + α ∆pwji(n)
It is momentum which is making the network not to trap in a local minima. α is called
momentum coefficient similar to learning rate. Since the error is propagating from output
layer to input layer it is called back-propagation method.
IV. OVERALL STRUCTURE OF THE PROCESS AND DESIGN ISSUES
Design issues:
Correlation can be used to locate features within an
image. The correlation of bifurcations with a fingerprint gives location of these minute
features in the fingerprint. Cross-correlation of the bifurcation with a fingerprint is done by
first rotating the bifurcation by 1800 and applying convolution with respect to Fourier
transform to the fingerprint and rotated bifurcation. This gives an image in which the places
where that bifurcation is present are intensified and other places attenuated. Hence a
thresholded image is used to get exact location of the bifurcation. This process for a
fingerprint is depicted below:
The thresholded image is now undergone Fourier transform
to give information about the areal distribution and location of the bifurcation in the
fingerprint. Since low frequency values give overall information only the first 7x7 matrix is
taken from this Fourier transform. All matrices obtained like for all bifurcations for a given
fingerprint are summed to get reduced data without loss of intrinsic information. This reduces
the data dimension from 128x128 to 7x7. The elements of this matrix are given row after row
as a vector input to the network. Thus this approach has an advantage of reducing data input
to the network.
The fingerprints are taken by ink method. These images are
converted to gray scale. Then to get correct results for patterns with thick ridges, the pattern
is subjected to edge detection using ‘canny’ operator [4]. The edge operator gives the local
information using pixels only a short distance from the pixel to which the operator is applied.

Structure of the process:


The block diagram of the whole process is shown in the preceding page which briefly
explains the application of Fourier Transforms to the correlated images , summing and
normalizing the images to apply to the Neural Networks and lastly the out specifying the
identity of the person,
The brief explanation by using a block diagram is as
shown starting with the 32 patterns,
Starting with
32 patterns

Block diagram of the overall process.


The correlated due to binary nature not showing the exact intensities of the various points.

A bifurcation and a fingerprint. The edge detected finger


The correlated image. Thresholded image.

The invariance properties of this technique are discussed below:


For any pattern recognition application invariance is an important issue because the input
patterns may be slightly shifted or scaled or rotated. A good recognition system should work
better even in the presence of these errors. The proposed technique is able to give not only
dimension reduction of data but also some invariance properties within some limits. These
properties are explained as:
 Invariance to shift: When the Fourier transform of an image, in which there a shift
in the position of fingerprint, is taken the magnitude, remains constant but the phase
changes. Hence the magnitude of Fourier transform is taken to account for this
problem.
 Invariance to scaling: When the fingerprint is small, its Fourier transforms values
change. But all the values will be multiplied by a constant. Hence normalization is
done to overcome this problem by dividing all data values with the first or last value.
 Invariance to rotation: Invariance to noise can be accomplished by using polar
coordinates and not rectangular coordinates. After converting to polar coordinates
the Fourier transform of any image becomes shift invariant because the center is
located and the Fourier transform is applied. This accounts for rotation invariance
problem. But the implementation is not done in polar coordinates. So the rotation
invariance may not accomplished by this rectangular coordinates.
 Invariance to noise and other effects: The problem of invariance to noise and other
effects like smoothing etc is easily solved by the neural network approach. The
advantage of neural approach compared to other approaches is its immunity to noise.
Hence this is problem is also solved.
V. APPLICATIONS
The fingerprint recognition system is nowadays applied in several fields. In fact, the
increased needs of security in different applications and the marked reduction in costs of
these technologies gave a strong impulse to diffuse from military and forensic applications to
network security systems, access control systems etc. These applications are briefly
discussed:
 Logon to a PC: Fingerprint-based logon to a PC. It avoids using unsafe or
difficult-to-remember passwords.
 Logon to a network: Fingerprint-based logon to a network and other remote
disks, printers, internet etc. Each PC Client will be equipped with a
fingerprint scanner and a user’s authentication is checked by his fingerprints
and the server then decides whether let him access the network or not.
 Document protection and Secure exchange of documents:
This application consists of storing documents in a protected archive and
preventing, by using fingerprint recognition, any unauthorized access. This
kind of protection is particularly useful when the document has to be
extracted from the PC and stored elsewhere (i.e. on a CDROM or a Tape) or
has to be sent by e-mail over an insecure net like Internet.
 Remote Authentication on Internet: The boom of world wide web has
intensified the interest in electronic commerce, but at the same it is attracted
those people who pursue illicit activities. What is needed is a secure way to
remotely authenticate users, before giving them access to services, resources
or allowing them to carry out an electronic transaction. Passwords are not a
valid solution to deal with secure authentications, since they can be stolen or
shared between more users. Fingerprint recognition is a natural solution to
remote authentication.
 Access Control: The physical access to a room (office, lab, home ...) is
controlled through fingerprints. This application has the double advantage of
ensuring a great security (thus preventing illicit accesses) and of avoiding
users to carry keys or badges with them.
 Government and Large-scale Applications:This is the field where the
fingerprint technologies were first applied. The main applications are:
In Police Services for criminal identification,
In Civil Services to avoid multiple concession of benefits under false identities
• For person identification in customs, airports, prisons,
• For identification of local criminals.
Since the present fingerprint recognition system is flexible in that it can be used in small
scale as well as large scale systems, it can be applied to all the aforementioned applications.
VI. RESULTS AND DISCUSSIONS
TABLE-I
RESULTS FOR TRAINED PATTERNS
Type Actual pattern Rotated by 100 Shifted Scale reduced Noise
Percentage 99.91 99.9 99.9 99.9 99.9
performance
The network was trained in 2208 epochs and performance reached to
1.17x10-6. Training took 14 seconds. The network is trained with 32 patterns with 8 patterns
for each fingerprint. Of the 8 patterns first one is good pattern, 2,3 are rotated by100 (one
clockwise, other anticlockwise), 4,5 are shifted (one horizontally, other vertically), 6 is
scaled pattern, 7,8 are noisy patterns. The performance for those patterns is shown in the
above table-I. The performance vs no. of epochs while training is plotted and shown below:
The network was tested with 32 fingerprints with 8 patterns for each fingerprint. This time
the patterns presented are more smoothed, more rotated, more shifted, more scaled, noisy
with more amplitude. The performance for those patterns is shown in table-II.

TABLE-II
RESULTS FOR TEST PATTERNS
Type Smoothed Rotated by 200 Shifted Scale reduced Noise
Percentage 99.90 92.7 98.89 98.97 98.84
performance

From the results depicted in above tables it can be observed that the network is
giving good results invariant to size, scale and rotation. Invariance with respect to rotation is
not good because the normalization to account for rotation is not precisely done. Also the
initial conditions affect the generalization ability of the network [8]. The rate of fingerprint
recognition depends much on the quality of fingerprints and effectiveness of the
preprocessing system such as the threshold level used in edge detection. Also if there are too
many broken lines are present in the image, the preprocessing system may fail.
VII. CONCLUSIONS
A fingerprint recognition system that uses local features preprocessing techniques are applied
to produce an invariant feature vector. Gradient decent and that momentum back propagation
of used as a training method to train the network. The network gives correct recognition even
in the presence of positioning or rotating errors. The use of neural network with this
technique for fingerprint recognition makes new technological advances in biometrics
VIII. FURTHER RESEARCH
The following avenues can be identified for further research:
 By increasing the number of bifurcations used and also using other features of
fingerprints like loop, whorl, tented arc, delta etc. this can be applied to systems with
large data set.
 In the preprocessing stage by using histogram equalization the performance can be
improved even for fingerprints which are thick ink patterns and seem to contain less
information needed for recognition.
 Training with genetic algorithms can give optimum results for this application.
 A precise normalization should be done to account for the rotation invariance.
 The speed of recognition can be improved by using a fast processor for cross-
correlation of fingerprints with bifurcations.
IX. REFERENCES:
♦ Zhang G.P., ”Neural Networks for Classification: A Survey”, IEEE
Transactions on systems, man and cybernetics—Part C: Applications and
Reviews, Vol. 30, No. 4, pp.451-462, November 2000.
♦ Takeda F. and Omatu S., “High Speed Paper Currency Recognition by Neural
Networks”, IEEE Transactions on neural networks, Vol. 6, No. 1, 73-77,
January 1995.
♦ Kim H. and Nam K., “Object Recognition of One-DOF Tools by a Back-
Propagation Neural Net”, IEEE Transactions on neural networks, Vol. 6, No.
2, 484-487, March 1995.
♦ Gose E., Johnsonbaugh R. and Jost S., “Pattern Recognition and Image
Analysis”, Chapters 7 and 8, pp. 298-315, 379-388, Prentice Hal PTR, 1996.
♦ Google Search Engine http//:www.google.com/nueral techniques/fingerprints

You might also like