PAPER PRESENTATION ON

NEURAL NETWORKS IMPLEMENTATION

IN FINGER PRINT IDENTIFICATION
USING OPTICAL WAVELET PREPROCESSOR

PSNA COLLEGE OF ENGINEERING AND

TECHNOLOGY

Presented by,
Naresh.B,
([email protected])
9865778869
(Final CSE)

Praveen Kumar.G,
([email protected])
9789402500
(Final CSE)

of wavelet functions may be successfully applied.

ABSTRACT:
Neural network is a massively

The recognition is performed by the artificial neural

parallel distributed processor made up of simple

network with built-in position invariance. Proper

processing units which has a natural property for

choice of the learning method additionally provides

storing experimental knowledge and making it

rotation-, scale, and intensity-invariance. To some

available for use. In this paper, a system consisting

extent also occluded images can be properly

of an optical wavelet preprocessor and an electronic

identified.

artificial neural network module has been examined.

The aim was to check whether such system could be

Neural network:

successfully employed to identify and discriminate

It is a massively parallel

highly complex biometric images, such as finger

distributed processor made up of simple processing

prints. The required capabilities included shift-,

units which has a natural property for storing

rotation,

experimental knowledge and making it available for

scale-

and

intensity-invariance.

The

additional constrain was the limited size of the

use.

training set. In fact, only one print of each finger

Types of neural networks

was used as a training pattern, which was a

Biological neural network

significant impediment for the learning process.

Artificial neural network

Since, as it is shown above, performance of ANN

working on raw data is not satisfactory, the
preprocessing based on edges-enhancement and
wavelet transform is proposed to enhance the

Types of neural networks regarding data

flow

generalization ability. Behaviour of various versions

Rummerhart type

of the described set-up is presented. In the optimal

Hopfield network

configuration proposed in this paper, the edgesenhancement filter first improves the quality of an
input image. Owing to this procedure, the ridges in a
fingerprint are more distinct. Subsequently, the
optical wavelet processor is employed. The Haas
wavelet of a certain scale is used to extract specific
features of an image. In future research other types

Applications of neural networks

Financial Analysis -- stock predictions.

Signature Analysis -- the banks in America

have taken to NNs to compare signatures
with what is stored.

Process Control Oversight -- NNs are used to

recognition based on neural networks architecture

advise aircraft pilots of engine problems.

with multi wavelet transforms as a preprocessor.

Direct Marketing -- NNs can monitor results

This system has in certain range invariance against

from a test mailing and determine the most

rotation, shift, and intensity and scale changes. The

successful areas.

achieved

Pen PC's.

satisfactory performance of the system working with

level

of

tolerance

should

provide

the real-life distorted data. Special focus is given on

the advantages provided by the preprocessing

Finger

print

identification

by

using extracting features that are most important to the

artificial neural network with optical process of pattern identification.

Artificial Neural networks with optical

wavelet preprocessor

wavelet preprocessor Optical wavelet preprocessor

The advantages of optical wavelet transform used as

a preprocessor for an artificial neural network are
investigated. We show by digital simulation that this
set-up can successfully identify and discriminate
complex biometric images, such as fingerprints. The
achieved capabilities include limited shift-, rotation-,
scale- and intensity-invariance.

based on VanderLugt correlator The value of optical

wavelet transform (OWT) for the given parameters
can be calculated as the cross-correlation function of
a signal and particular wavelet function .In the
spectrum plane of correlator we obtain product of
the Fourier spectrum of input image and the
conjugate Fourier transform (FT) of the wavelet
Opto-Electron.

Introduction
The advantage of using optical
approach, optical correlators or artificial neural
networks, instead of minutiae analysis is that a
global operation on the images is less sensitive to
local distortions that typically occur during the
finger print reading procedure. The aim of
simulations presented here is to show that the
wavelet preprocessor can improve the ability of an

The wavelet processor is a classical VanderLugt

ANN to identify and discriminate the highly

correlator with a computer generated hologram

complex data such as finger prints. They proposed

(CGH) placed in the Fourier plane. Because the DC

multi

component of the wavelet spectrum is equal to zero,

channel

hybrid

system

for

character

coding the spectra on the CGH requires only small

number of quantization levels, and therefore the

Fingerprint

quantization related error is less than 1%. The

neural network using optical wavelet

optical wavelet transform can be performed with

many other set-ups, but CGH

identification

in

artificial

preprocessing

offers superior flexibility. It does not only provide

an easy way to change the wavelet, its scale or
shape, but also allows parallel realization of multiple
wavelets transforms. Depending on the required
number of wavelets, spatially separated images of
wavelet transforms can be obtained by adding
appropriate spectra.
The larger scales of wavelets correspond to

All the tested networks have two hidden layers, but

lower spatial-frequencies and therefore the sampling

number of clusters in these layers varies. The

is preformed with the bigger interval. For small

structure of a FELSI unit can be symbolically

wavelet scales the sampling is denser. This operation

described as: 1-X-Y-2,

reduces the redundancy in the object in order to

which means that there is one cluster in the input

improve the generalization ability of ANN

layer, X clusters in the first hidden layer, Y clusters

Electronic ANN module

in the second hidden layer and two clusters in the

The architecture of ANN used in the

output layer.

simulated system is based on the FELSI (feed

forward

model

with

local

space-invariant

Learning procedure of ANN

interconnections) model. The network consists of

The size of samples was 90128 pixels. Two

layers of clusters. Each cluster is composed of 6464

fingerprints were chosen arbitrarily. Next, the

neural units. The units in the input layer have linear

negatives were taken and placed centrally in the

activation function, whereas units in all other layers

black 128128 scene. As a result the 128128 black

have bipolar sigmoid activation function Y(x).

(zero pixel value) input scenes with white (pixel

values in the [0,255] range) ridges are obtained. As
the optional stage of preprocessing, the edgesenhancement filter was applied.

Input layer

hidden layers

output

patterns

are

fed

to

the

network

with

the

corresponding pairs of target patterns in the output

clusters.
In most cases, the additional noise is introduced in
the input plane. In such a situation, for every epoch,
random uniformly distributed noise is generated and
added to the training patterns. This modification of
the learning algorithm, although improves the
generalization ability, also significantly slows down
the learning process and, in many
cases, it does not allow the process to converge at
the desired error level. The learning error function is
defined as error. Exactly reverse to the target
patterns, the error function
equals. When only one unit has reverse value the
error equals.
To provide comparability of the results, the same
two Fingerprints were used as training patterns for
all simulations. Before the signal is fed to the ANN,
the down-sampling is performed. In the experiment,
the wavelet dilation parameters for both directions
were a = b. The down-sampling interval is inversely
proportional to the scale of the wavelet.

Occasionally, even though the RPROP algorithm

was used, the learning process was not able to
escape an undesirable local minimum. In such a
situation, the jog of weights method was applied.
It was observed that this technique was usually
successful when the range of random changes was
higher than 75% of weights value. Of course, when
the first try did not bring the expected result, the

After that, the image is centrally placed on the 6464

operation could be repeated with the bigger range of

black background and normalized within the range

changes.

[1, 1]. The smaller initial values tend to make the

learning process vulnerable to undesirable local

Postprocessing of an output signal

minima.
In the process of supervised learning, the resilient

The ANN is supposed to discriminate

Back propagation (RPROP) algorithm and the batch

between two classes of patterns, so two clusters in

update mode is applied. In each epoch, both training

the output layers are used. When the network is fed

with a sample belonging to the class no. 1, it is

consisted of patterns obtained from the training

expected to produce the yes pattern in the cluster

patterns by various types of deformation such as

no. 1 and the no pattern in the other cluster. When

shift, rotation, scaling, irregular intensity changes,

a sample from the class no. 2 is presented, the

occlusion, pinch and punch operation. Since the

response should be inverse. The output signal

main objective of this research was to check whether

corresponding to yes pattern is the 66 matrix of

the examined system could be applied to identify the

excited (+1 value) pixels on the background

real-life

composed from pixels with 1 value. The no

requirements should match a probable distortion

pattern is simply the background with the output of

caused by the inevitable inaccuracies of the finger

all pixels equal to 1.

print acquirement process. Accordingly, the required

distorted

fingerprints,

the

invariance

range of tolerance was: for rotation 10, scale 5%,

Defining the yes pattern as a 66

shift 10% of the input area, pinch and punch

matrix of pixels, rather than only one pixel output,

10%.The second testing set enclosed the real life

brings two benefits. First, the learning process is

Data which contained the different prints of the

more stable, what means that it is faster and less

fingers used as training patterns.

likely to get stuck in a local minimum. Second, the

obtained output can be better interpreted. As it is
explained below, it is crucial to understand properly
the output signal and the analysis of its deformation
greatly contributes to correct identification. When
the learning process had been successfully finished,
the systems performance was tested. Different input
patterns were fed to the system and the signal in the
output clusters of the ANN was analysed.
Modulus of Q informs how certain
the discrimination is, while sign indicates which
pattern has been identified in ideal case Q is infinite,
hence the hard-clipping at the value of 10 (10) is
applied.

Results of recognition digital simulation

ANN without preprocessing
ANN with edges-enhancement
and wavelet transform in the preprocessing stage.
The goal of this section is to find the configuration
of the tested system that could provide good
generalization ability and sufficient discrimination
capability at the same time. The idea of feature
extraction is based on the reduction of information
content to the required minimum. So, the main task
is to define which part of image is crucial and this
should be omitted. In the case of fingerprints, the
high frequency details, such as ridges, are essential

Testing pattern

and the selected method of preprocessing must be

We used two test patterns to

evaluate the performance of system. First, testing set

sensitive to this kind of information. We chose to

use the Haars wavelet of scale 2. This choice can be

confirmed by the analysis of Fourier spectra.. It is

In the systems presented above, both edges-

visible that a significant part of information is

enhancement and wavelet transform were used in

carried by the band of high frequencies. Most

the preprocessing stage. This combined method of

probably, this is the information crucial to the

preprocessing gives good results, so it is worth

recognition process. Obviously, a suitable wavelet

investigating how big the role of each preprocessing

should not eliminate this band of frequencies. The

operation.

Haars wavelet of scale 2 appears to best fit the

spectrum

of

finger

print.

An

additional

ANN

with

only

edges-enhancement

in

the

preprocessing operation, performed before the

preprocessing stage In order to assess the influence

wavelet transform, is proposed. The most basic, in

of edges-enhancement, a system identical with the

tuitive

set-up tested above, but without the wavelet

idea

of

preprocessing

is

the

edges-

enhancement. As such, it does not extract any

preprocessor was examined.

particular kind of features or details, but partially

removes unwanted distortions in an analyzed image.

At first, these results may seem unexpectedly

Therefore the image, before being transformed by

disappointing, but when the real meaning of the

the wavelet preprocessor, was filtered with proper

edges-enhancement operation is fully understood,

edges-enhancement filter edges are enhanced. It is

this poor performance can be fully explained. The

visible

filter

generalization capability of this system is extremely

successfully recovers sharp ridges in the previously

limited. The system does not respond to slightly

blurry areas.

distorted samples. This is most probably caused by

that

the

edges-enhancement

over fitting. These problems are associated with the

Although the performance of both systems is much

way in which the edges-enhancement filter changes

better than of the system using raw data, there are

the image. The filtered image practically does not

still some misidentifications. In all tests, the system

contain pixels of medium intensity.

with 1-3-3-2 FELSI behaved better than the larger

The ridges, which normally have slightly blurred

network. It is known that the generalization ability

boundaries, become sharp. of the systems with

strongly depends on the number of free parameters

preprocessing (edges enhancement and Haar 2), 1-3-

in the neural network. Apparently, the 1-4-4-2

3-2 FELSI, decay parameter 10 ANN with only

FELSI suffers from the excessive number of clusters

wavelet transform in the preprocessing stage As the

in hidden layers. Thus, in following

final test, the same system but with wavelet

simulations the 1-3-3-2 network was used.

preprocessor instead of edges-enhancement filter

was examined The discrimination capability of this

configuration is not sufficient to provide reliable

additional constrain was the limited size of the

identification of distorted samples, especially when

training set. In fact, only one print of each finger

rotation

low

was used as a training pattern, Behaviour of various

generalization ability. The low level of selectivity of

versions of the described set-up is presented in the

the feature extraction process is the probable cause

optimal configuration proposed in this paper the

of this situation. It is visible that only the combined

edges-enhancement filter first improves the quality

method of edges-enhancement and wavelet-based

of an input image. Owing to this procedure, the

feature extraction gives satisfactory results. In

ridges

hardware implementations of the preprocessor, both

Subsequently, the optical wavelet processor is

these operations can be simultaneously performed

employed.

is

considered.

This

indicates

in

fingerprint

are

more

distinct.

by the properly prepared filter.

REFERENCES

Conclusion
In this paper, a system consisting of
an optical wavelet preprocessor and an electronic
artificial neural network module has been examined.
The aim was to check whether such system could be
successfully employed to identify and discriminate
highly

complex

biometric

images,

such

as

fingerprints. The required capabilities included shift,

rotation, scale- and intensity-invariance.
The development of true Neural Networks is a fairly
recent event, which has been met with success. Two
of the different systems (among the many) that have
been developed are: the basic feedforward Network
and

the

Hopfield

Net.

The future of Neural Networks is wide open, and

may lead to many answers and/or questions. Is it
possible to create a conscious machine? What rights
do these computers have? How does the human
mind work? What does it mean to be human? The

H.M.HSONS

FUNDAMENTALS

OF

ANN PRENTICE HALL OF INDIA, 1998.

www.google.com
www.nerogame.org
www.geekswithblogs.net/cbreisch/archiv
e/2006/08/15/88021.aspx
www.freshmeat.net/projects/fann