Contributed paper

OPTO-ELECTRONICS REVIEW 11(4), 327–337 (2003)

Fingerprint identification by using artificial neural network

with optical wavelet preprocessing
M. ANTOWIAK and K. CHA£ASIÑSKA-MACUKOW*

Division of Information Optics, Institute of Geophysics, Faculty of Physics, Warsaw University,

7 Pasteura Str., 02-093 Warsaw, Poland

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 fin-
gerprints. The achieved capabilities include limited shift-, rotation-, scale- and intensity-invariance. We also show that the
edges-enhancement filter, applied before the wavelet transform, significantly improves abilities of the system.

Keywords: optical fingerprint identification, biometrics, artificial neural network, optical wavelet transform.

1. Introduction with multiwavelet transforms as a preprocessor. This sys-

tem has in certain range invariance against rotation-, shift-,
Biometrics is formed from the person’s selected unique intensity- and scale- changes [8]. The achieved level of tol-
physical attributes which may be applied for the purpose of erance should provide satisfactory performance of the sys-
automated personal identification [1]. Typically, biometrics tem working with the real-life distorted data. Special focus
are based on features such as fingerprints, geometry of is given on the advantages provided by the preprocessing
hand and vein patterns, iris or retinal patterns, image of a extracting features that are most important to the process of
face, and characteristic features of the person’s voice. The pattern identification.
inherent feature of biometric data is its great complexity. The results shown here are achievable using discussed
Essential in this kind of images is high spatial-frequency set-up with various techniques of preprocessing and ANN
details. Moreover, the differences among diverse classes of learning. The drawbacks and advantages of the examined
patterns are small which requires high discrimination capa- system are investigated as well as the prospective direc-
bility of the system. tions of development.
This paper discusses the capabilities of the hybrid sys-
tem based on the optical wavelet preprocessor (OWP) and 2. ANN with optical wavelet preprocessor
electronic artificial neural network (ANN) applied to fin-
gerprints identification. Most digital processing methods A general scheme of presented recognition system is
for fingerprint identification are based on the extraction of shown in Fig. 1. This system is a modification of the multi-
minutiae, which are the most unique fingerprint features channel hybrid set-up described in Ref. 7. In this paper
and they are defined as the fingerprint ridge endings and bi- one-channel set-up with only one neural network and one
furcations [2]. The advantage of using optical approach, wavelet transformed image is examined. Main core con-
i.e., optical correlators [3–6] or artificial neural networks, sists of an electronic ANN module and optical wavelet pre-
instead of minutiae analysis is that a global operation on processor based on VanderLugt correlator [9] (Fig. 2).
the images is less sensitive to local distortions that typically
occur during the fingerprint reading procedure.
The aim of simulations presented here is to show that
2.1. Optical wavelet preprocessor based on
the wavelet preprocessor can improve the ability of an VanderLugt correlator
ANN to identify and discriminate the highly complex data
The value of optical wavelet transform (OWT) for the
such as fingerprints. This research is inspired by the results
given parameters can be calculated as the cross-correlation
obtained by the Japanese team of Hirokawa, Itoh, and
function of a signal and particular wavelet function [10].
Ichioka [7]. They proposed multichannel hybrid system for
Using the optical correlator is natural and effective method
character recognition based on neural networks architecture
to obtain it [11]. In the spectrum plane of correlator we ob-
tain product of the Fourier spectrum of input image s(x,y)
*e-mail:
kmacukow@mimuw.edu.pl and the conjugate Fourier transform (FT) of the wavelet

Opto-Electron. Rev., 11, no. 4, 2003 M. Antkowiak 327

Fingerprint identification by using artificial neural network with optical wavelet preprocessing

Fig. 1. Flow chart of the recognition system.

Fig. 2. Electronic ANN module with Optical Wavelet Preprocessor based on VanderLugt correlator.

h(x,y) (which is normally coded on the CGH) which is scales of wavelets correspond to lower spatial-frequencies
equivalent to the Fourier Transform of the ws(.). So, the ba- and therefore the sampling is preformed with the bigger in-
sic formula takes the form terval. For small wavelet scales the sampling is more
denser. This operation reduces the redundancy in the object
FT[ w S ( a, b, c, d ;x, y)] = abH * ( an 1, bn 2 ) S(n 1, n 2 ), (1) in order to improve the generalization ability of ANN.

where (n1,n2) are the coordinates in the Fourier plane, 2.2. Electronic ANN module
H(n1,n2) = FT[h(x,y)], S(n1,n2) = FT[s(x,y)], and a, b, c, d
The architecture of ANN used in the simulated system is
are the parameters of the wavelet dauther, respectively.
based on the FELSI (feed forward model with local
Next, by calculating the inverse Fourier transform, the
space-invariant interconnections) model introduced and de-
wavelet transform ws(.) of the input signal is obtained.
veloped by the group of professor Ichioka [7,8]. The net-
The wavelet processor is a classical VanderLugt
work, as illustrated in Fig. 3, consists of layers of clusters.
correlator with a computer generated hologram (CGH)
Each cluster is composed of 64´64 neural units. The units
placed in the Fourier plane [11]. Because the DC compo-
in the input layer have linear activation function, whereas
nent of the wavelet spectrum is equal to zero, coding the
units in all other layers have bipolar sigmoid activation
spectra on the CGH requires only small number of
function Y(x)
quantization levels, and therefore the quantization related
error is less than 1% [7]. The optical wavelet transform can
be performed with many other set-ups [10–15], but CGH 1 - e -x
y (x ) = . (2)
offers superior flexibility. It does not only provide an easy 1 + e -x
way to change the wavelet, its scale or shape, but also al-
lows parallel realization of multiple wavelet transforms. All the tested networks have two hidden layers, but
Depending on the required number of wavelets, spatially number of clusters in these layers varies. The structure of a
separated images of wavelet transforms can be obtained by FELSI unit can be symbolically described as: 1-X-Y-2,
adding appropriate spectra. which means that there is one cluster in the input layer, X
In this experiment, the obtained OWT images were clusters in the first hidden layer, Y clusters in the second
down-sampled according to the scale of wavelet. The larger hidden layer and two clusters in the output layer. For exam-

328 Opto-Electron. Rev., 11, no. 4, 2003 © 2003 COSiW SEP, Warsaw
 Contributed paper

Fig. 5. Edges-enhancement filter.

pling interval is inversely proportional to the scale of the

wavelet

w(x, y) = w S ( ax, ay). (3)

After that, the image is centrally placed on the 64´64

black background and normalized within the range [–1,1].
Fig. 3. Architecture of FELSI model (after Refs. 7 and 8). The initial values of weights and biases are random.
Contrary to Ref. 7, where initial values were “extremely
ple, the network presented in Fig. 3 can be described as small”, the fastest and most certain convergence is
1-3-3-2. In all experiments the size of weight matrices was achieved for initial weights randomly distributed in range
11´11 between the input and the first hidden layer and [ - 2.38 n , 2.38 n ], where n is the fan-in of a unit (the
21´21 between other layers. The formulas of backpro- number of units which are fed forward to this unit) [5 after
pagation algorithm for the FELSI network can be found in 39]. The smaller initial values tend to make the learning
Refs. 7–9. process vulnerable to undesirable local minima.
In the process of supervised learning, the resilient
2.2.1. Learning procedure of ANN backpropagation (RPROP) algorithm [17] and the batch
update mode is applied. In each epoch, both training pat-
The samples of fingerprints were taken by the Internet from terns are fed to the network with the corresponding pairs of
the database of Bologna University [16]. The size of sam- target patterns in the output clusters. The parameters of
ples was 90´128 pixels. Two fingerprints were chosen ar- RPROP are set to: Dmin = 10–6, Dmax = 10–1, h- = 0.5,
bitrarily. Next, the negatives were taken and placed cen- h+ = 1.2. Dmax is significantly smaller than normally used,
trally in the black 128´128 scene. As a result the 128´128 in order to support the weight-decay effect and avoid too
black (zero pixel value) input scenes with white (pixel val- fast growth of weights. The decay term varied from 10–6 to
ues in the [0,255] range) ridges are obtained [see Figs. 4(a) 10–4.
and 4(b)]. As the optional stage of preprocessing, the In most cases, the additional noise is introduced in the
edges-enhancement filter presented in Fig. 5 was applied. input plane. In such a situation, for every epoch, random
To provide comparability of the results, the same two uniformly distributed noise is generated and added to the
fingerprints [Figs. 4(a) and 4(b)] were used as training pat- training patterns. This modification of the learning algo-
terns for all simulations. In the following sections they are rithm, although improves the generalization ability, also
referred to as 1_0 and 4_0, respectively. significantly slows down the learning process and, in many
Before the signal is fed to the ANN, the down-sampling cases, it does not allow the process to converge at the de-
is performed. In the experiment, the wavelet dilation pa- sired error level.
rameters for both directions were a = b. The down-sam- The learning error function is defined as

C S X Y
1
error =
CSXY
å å å å Tc,s (x, y) - Oc,s (x, y), (4)
c s x y

where (x,y) are the coordinates in a cluster, S is the number

of training patterns, and C is the number of clusters in the
output layer. In all experiments, X = Y = 64, C = 2, and
S = 2.
The first step in the process of learning was to choose
the desired value of error function. If the error is too high,
the discrimination ability may be not satisfactory. On the
other hand, the risk of overfitting causing the deterioration
Fig. 4. Training patterns used for all systems: (a) 1_0 and (b) 4_0. of the generalization ability, must be taken into account. To

Opto-Electron. Rev., 11, no. 4, 2003 M. Antkowiak 329

Fingerprint identification by using artificial neural network with optical wavelet preprocessing

find the proper level of an error, as the preliminary tests, patterns were fed to the system and the signal in the output
the learning was stopped at different stages and the proper- clusters of the ANN was analysed.
ties of the obtained networks were examined. As a result, To describe quantitatively the output signal, the param-
the range [2´10–3, 3´10–3] was found to give the best re- eter Q was defined
sults. To compare, when all output units have the values
exactly reverse to the target patterns, the error function ì æM ö
ï min ç 1 ,10÷ , if M1 > M 2 and M1 > 0.0208
equals 2. When only one unit has reverse value the error è M ø
ï 2
equals 1.2´10–4. ïï æ M2 ö
Occasionally, even though the RPROP algorithm was Q = í ( -1) min ç ,10÷ , if M1 < M 2 and M 2 > 0.0208 , (5)
ï è M1 ø
used, the learning process was not able to escape an unde- 0 , otherwise
ï
sirable local minimum. In such a situation, the “jog of ï
weights” method was applied [17]. It was observed that îï
this technique was usually successful when the range of
random changes was higher than 75% of weights value. Of where M1 and M2 are obtained as follows. First, the central
course, when the first try did not bring the expected result, parts of the output signals are normalized into the [0,1]
the operation could be repeated with the bigger range of range. Then, convolution with the 6´6 square of value +1
changes. Regrettably, in many cases, especially when the is performed. Finally, the maximum values M1 and M2 in
high level of additional noise was used, this technique was both products of convolution are found.
not able to help the learning algorithm to avoid getting Modulus of Q informs how certain the discrimination
stuck in very broad minima. As a result, some networks de- is, while sign indicates which pattern has been identified
scribed below have the learning error higher than recom- (positive value means the answer “yes” in cluster no.1 –
mended one. fingerprint no.1 is recognized, negative value means the
opposite). In ideal case Q is infinite, hence the hard-clip-
2.2.2. Postprocessing of an output signal ping at the value of 10 (–10) is applied.
This algorithm of post-processing has numerous advan-
The ANN is supposed to discriminate between two classes tages:
of patterns, so two clusters in the output layers are used. • only centrally located output peaks are taken into ac-
When the network is fed with a sample belonging to the count. This reduces the probability of a false,
class no. 1, it is expected to produce the “yes” pattern in the • the square-like output patterns are more important than
cluster no. 1 and the “no” pattern in the other cluster. When isolated excited pixels,
a sample from the class no. 2 is presented, the response • small peaks are ignored. The threshold is supposed to
should be inverse. The output signal corresponding to eliminate false alarms caused by output noise related to
“yes” pattern is the 6´6 matrix of excited (+1 value) pixels the fact that the network’s learning process was stopped
on the background composed from pixels with –1 value with a learning error higher than zero.
[see Fig. 6(a)]. The “no” pattern is simply the background In many cases, when in both clusters the maximum val-
with the output of all pixels equal to –1 [see Fig. 6(b)]. ues are equal and the method proposed in Ref. 7 fails, it
Defining the “yes” pattern as a 6´6 matrix of pixels, still gives proper answer. Two examples of proper identifi-
rather than only one pixel output, brings two benefits. First, cation in uncertain situation are presented in Fig 7. In the
the learning process is more stable, what means that it is upper pair of clusters both maxima equal +1, but in
faster and less likely to get stuck in a local minimum. Sec- Fig. 7(a) the maximum lays outside the detection field and
ond, the obtained output can be better interpreted. As it is therefore is ignored. In this pair, Fig. 7(b) is interpreted as
explained below, it is crucial to understand properly the “yes” and Fig. 7(a) as “no” and the input sample is properly
output signal and the analysis of its deformation greatly identified. In the lower pair of clusters, although both max-
contributes to correct identification. ima are equal and lay in the detection field, the output
When the learning process had been successfully fin- Fig. 7(c) contains square-like pattern whereas in Fig. 7(d)
ished, the systems performance was tested. Different input the maximal unit is separated. Here, cluster in Fig. 7(c) is
read as “yes” and in Fig. 7(d) as “no”. Of course, the com-
parison of both output clusters is crucial and, in other cir-
cumstances, pattern in Fig. 7(d) could also be interpreted as
“yes”. It is important to choose proper level of certainty of
identification.

3. Testing pattern
We used two test patterns to evaluate the performance of
system. First, testing set consisted of patterns obtained
Fig. 6. Output signal corresponding to (a) “yes” and (b) “no”. from the training patterns [see Figs. 4(a) and 4(b)] by vari-

330 Opto-Electron. Rev., 11, no. 4, 2003 © 2003 COSiW SEP, Warsaw
 Contributed paper
This test was supposed to check the “real-life” perfor-
mance of the system. Every print of a given finger is the
result of all possible distortions. In many cases the degree
of deformation is much higher than in samples from the
first group of tests. For example, the intensity distortion
of the images (b) and (c) in Fig. 11 is evidently more se-
vere than in Fig. 8(d). It is also significant that these sam-
ples often contain only a fragment of a training fingerprint
[e.g., Fig. 10(g)] or even a different part of a finger. In
such cases, the system cannot be expected to be able to
give proper answer. It seems that the only solution is to
collect the fingerprints more accurately. Since the influ-
ence of a distinct deformation cannot be measured or con-
trolled, it is impossible to precisely evaluate the level of
similarity between the training pattern and a given testing
sample. Hence, only qualitative analysis can be per-
formed.

4. Results of recognition – digital simulation

The procedure of numerical simulations explained above,
Fig. 7. Two pairs of output clusters. In the upper pair (a) is in some cases with minor adjustments, has been applied to
interpreted as “no” and (b) as “yes”. In the lower pair (c) is examine the performance of various configurations of the
interpreted as “yes” and (d) as “no”. White square limits the field of described system.

Fig. 8. Examples of deformations: (a) shift, (b) rotation, (c) scaling, (d) intensity change, and (e) occlusion. The edges-enhancement filter
was applied after deformation.

ous types of deformation such as shift, rotation, scaling, ir- 4.1. ANN without preprocessing
regular intensity changes, occlusion, “pinch” and “punch”
First, as a starting point, we examined the capabilities of
operation. Figures 8 and 9 present some examples of these
the system without any preprocessing. This means that pat-
deformations. Since the main objective of this research was
terns 1_0 and 4_0 (Fig. 4) were directly fed to the ANN
to check whether the examined system could be applied to
identify the real-life distorted fingerprints, the invariance
requirements should match a probable distortion caused by
the inevitable inaccuracies of the fingerprint acquirement
process. Accordingly, the required range of tolerance was:
for rotation – 10°, scale 5%, shift – 10% of the input area,
“pinch” and “punch” – 10%.
The “pinch” and “punch” effects are offered by the
Paint Shop Pro 6 application and closely simulate real de-
formations (Fig. 9).
The second testing set enclosed the “real life” data Fig. 9. “Pinch” and “punch” deformations, (a) pinch – degree 15,
which contained the different prints of the fingers used as (b) original image, (c) punch – degree 15. Deformations were made
training patterns. For both training fingerprints seven other with Paint Shop Pro 6. The edge-enhancement filter was applied
samples were used (Figs. 10 and 11). after deformation.

Opto-Electron. Rev., 11, no. 4, 2003 M. Antkowiak 331

Fingerprint identification by using artificial neural network with optical wavelet preprocessing

Fig. 10. Different prints of the finger no. 1. (a) training pattern 1_0 and (b)–(h) testing patterns.

Fig. 11. Different prints of the finger no. 4: (a) training pattern 4_0 and (b)–(h) testing patterns.

module. A variety of learning parameters was used to ob- Table 1. Performance of the system with 1-3-3-2 FELSI, noise
tain the best possible results. Two configurations of the range [–0.5,0.5] and learning error equal to 3.0´10–3.
FELSI unit: 1-3-3-2 and 1-4-4-2 have been tested. The
smaller network proved to give better results, especially % of correct % of wrong
Type of distortion
recognitions recognitions
when noise was added. Performance of the set-up with
1-3-3-2 FELSI, noise range [–0.5,0.5] and learning error Shift (13 pixels) 100.0 0.0
equal to 3.0´10–3 is presented in Table 1.
Rotation (1°–10°) 77.5 10.0
These numbers show that, although the system is shift
invariant, it does not perform well when distorted images Intensity 23.0 0.0
are presented. In the rotation range of 10° as much as 10%
Other – scale, pinch, punch, 20.0 5.0
of samples was incorrectly identified and for another occlusion
12.5% the answer was “undecided”. For intensity distor-
tions only 23% of testing samples was identified correctly. “Rreal-life” samples 7.1 7.1

332 Opto-Electron. Rev., 11, no. 4, 2003 © 2003 COSiW SEP, Warsaw
 Contributed paper

Fig. 12. Fourier spectra of: (a) Haar’s wavelet of scale 2, (b) Haar’s wavelet of scale 4, and (c) a fingerprint - logarithmic scale is used and
the DC region is blocked to show higher frequencies.

In case of scaling, occlusion and “pinch” and “punch” the cessfully recovers sharp ridges in the previously blurry ar-
performance was even poorer. The second group of tests, eas. This effect can be clearly observed in the central re-
conducted with different prints of the fingers used as train- gion of pattern in Fig. 13(a).
ing patterns, turned out to be more difficult. Only 1 out of All tests presented below have been performed with the
14 identifications was undoubtedly right and there was also Haar’s wavelet of scale 2 coded in the wavelet prepro-
one wrong recognition. cessor. For all samples the edges-enhancement filter was
In this situation, preprocessing is proposed as the first applied. This means that samples in Fig. 4 have been
method to improve generalization and discrimination capa- filtered with the filter presented in Fig. 5 and then the
bility of the system. wavelet transform was performed. Such preprocessed im-
ages were fed to the FELSI unit as the training patterns.
4.2. ANN with edges-enhancement and wavelet First, the influence of FELSI size on the systems’ per-
transform in the preprocessing stage formance was examined. Two networks had been prepared:
1-3-3-2 and 1-4-4-2. In the learning process no additional
The goal of this section is to find the configuration of the noise was used. The decay parameter was set to 10–4. In
tested system that could provide good generalization ability both cases the learning error was equal to 2´10–3. Tables 2
and sufficient discrimination capability at the same time. and 3 illustrate the performance of both systems.
The idea of feature extraction is based on the reduction Although the performance of both systems is much
of information content to the required minimum. So, the better than of the system using “raw” data, there are still
main task is to define which part of image is crucial and some misidentifications. In all tests, the system with
which should be omitted. In the case of fingerprints, the 1-3-3-2 FELSI behaved better than the larger network. It is
high frequency details, such as ridges, are essential and the known that the generalization ability strongly depends on
selected method of preprocessing must be sensitive to this
kind of information. We chose to use the Haar’s wavelet of
scale 2. This choice can be confirmed by the analysis of
Fourier spectra. Figure 12(c) illustrates the spectrum of a
typical fingerprint. It is visible that a significant part of in-
formation is carried by the band of high frequencies. Most
probably, this is the information crucial to the recognition
process. Obviously, a suitable wavelet should not eliminate
this band of frequencies. Figures 12(a) and 12(b) present
the Fourier spectra of Haar’s wavelets. The Haar’s wavelet
of scale 2 appears to best fit the spectrum of a fingerprint.
An additional preprocessing operation, performed be-
fore the wavelet transform, is proposed. The most basic, in-
tuitive idea of preprocessing is the edges-enhancement. As
such, it does not extract any particular kind of features or
details, but partially removes unwanted distortions in an
analysed image. Therefore the image, before being trans-
formed by the wavelet preprocessor, was filtered with
proper edges-enhancement filter (Fig. 5). Figure 13 pres-
ents training patterns before and after the edges are en- Fig. 13. Training patterns: (a) pattern 1_0, (b) pattern 1_0 with enhanced
hanced. It is visible that the edges-enhancement filter suc- edges, (c) pattern 4_0, and (d) pattern 4_0 with enhanced edges.

Opto-Electron. Rev., 11, no. 4, 2003 M. Antkowiak 333

Fingerprint identification by using artificial neural network with optical wavelet preprocessing

Table 2. Performance of the system with preprocessing (edges en- were used. All had the 1-3-3-2 FELSI unit, all used
hancement and Haar 2), 1-3-3-2 FELSI, decay parameter 10–4 and edges-enhancement and Haar’s wavelet of scale 2 in the
learning error equal to 2.0´10–3. preprocessor and for all Weight Decay with parameter 10–5
was employed. The only difference was the range of noise,
% of correct % of wrong
Type of distortion which was [–0.25, 0.25], [–0.5, 0.5] and [–0.75, 0.75]. For
recognitions recognitions
the higher values of noise, the learning process did not
Shift (13 pixels) 100.0 0.0 properly converge. The performance of these systems is
Rotation (1°–10°) 90.0 0.0 presented in Table 4.
These numbers clearly show that good tolerance for ar-
Intensity 95.0 0.0 tificially distorted samples does not necessarily guarantee
satisfactory performance when strongly deformed images
Table 3. Performance of the system with preprocessing (edges en- are used.
hancement and Haar 2), 1-4-4-2 FELSI, decay parameter 10–4 and While for controlled deformations the increased noise
learning error equal to 2.0´10–3. range provides better generalization ability, in case of the
% of correct % of wrong
“real-life” its impact is more complex. The effectiveness of
Type of distortion recognition is at 36% when the noise is in the range [–0.25,
recognitions recognitions
0.25], reaches 50% for [–0.5, 0.5] noise-range and falls to
Shift (13 pixels) 100.0 0.0 29% when noise is increased to [–0.75, 0.75]. The only
Rotation (1°–10°) 75.0 2.5 benefit from the higher learning noise is elimination of
false identifications. A similar effect was reported in
Intensity 65.0 0.0 Ref. 7.
The analysis of these results indicates that the [–0.5,
0.5] noise-range is the optimal trade-off between computa-
the number of free parameters in the neural network. Ap- tional time and system’s performance. Figure 14 gives the
parently, the 1-4-4-2 FELSI suffers from the excessive detailed information about identification of each testing
number of clusters in hidden layers. Thus, in following sample. In virtually all cases the recognition is extremely
simulations the 1-3-3-2 network was used. sure. In Fig 14(c), all three “undecided” answers are for the
Let us to concentrate on the meaning of random noise occluded samples. This is acceptable since, in its nature,
added in the input layer during the process of learning. To the neural network sees the input patterns as a whole, rather
analyse this topic more precisely, three identical systems than concentrate on particular details. Apparently, some

Fig. 14. System with preprocessing (edges enhancement and Haar 2). Test of the tolerance of a system with 1-3-3-2 FELSI and [–0.5, 0.5]
noise range against: (a) rotation, (b) intensity distortions, (c) other deformations, and (d) real-life samples. In each chart the first half of
points corresponds with the training pattern 1_0 and for them positive value of the parameter Q is expected. The second half of points
corresponds with the training pattern 4_0 and negative value of the parameter Q is expected.

334 Opto-Electron. Rev., 11, no. 4, 2003 © 2003 COSiW SEP, Warsaw
 Contributed paper

Table 4. Performance of the systems with preprocessing (edges enhancement and Haar 2), 1-3-3-2 FELSI, decay parameter 10–5,
learning error equal to 2.0´10–3 and various noise ranges.
% of correct recognitions/% of wrong recognitions

Type of distortion [–0.25,0.25] noise range [–0.5,0.5] noise range [–0.75,0.75] noise range

Shift (13 pixels) 100.0/0.0 100.0/0.0 100.0/0.0

Rotation (1°–10°) 82.5/0.0 95.0/0.0 95.0/0.0

Intensity 70.0/0.0 85.0/0.0 80.0/0.0

Other – scale, pinch, punch, occlusion 75.0/0.0 85.0/0.0 85.0/0.0

“Real-life” samples 35.7/14.3 50.0/7.1 28.6/0.0

part of crucial information has been destroyed and, conse- cement operation is fully understood, this poor perfor-
quently, the recognition was impossible. mance can be fully explained. The generalization capabil-
In the systems presented above, both edges-enhance- ity of this system is extremely limited. The system does
ment and wavelet transform were used in the preprocessing not respond to slightly distorted samples. This is most
stage. This combined method of preprocessing gives good probably caused by overfitting. Although the learning
results, so it is worth investigating how big the role of each process in this test was stopped significantly earlier than
preprocessing operation is. in previous tests, the network lost its ability to generalize.
The learning could be stopped even earlier, but in such
4.3. ANN with only edges-enhancement in the case the discrimination capability, which already is not
preprocessing stage satisfactory, would suffer. These problems are associated
with the way in which the edges-enhancement filter
In order to assess the influence of edges-enhancement, a changes the image. The filtered image practically does not
system identical with the set-up tested above, but without contain pixels of medium intensity. The ridges, which
the wavelet preprocessor, was examined (Table 5). normally have slightly blurred boundaries, become sharp.
At first, these results may seem unexpectedly disap- This procedure is in fact equal to enhancement of
pointing, but when the real meaning of the edges-enhan- high-frequencies.

Table 5. Performance of the system with preprocessing (only edges enhancement), 1-3-3-2 FELSI, decay parameter 10–5, learning error
equal to 2.0´10–3 and [–0.5,0.5] noise range.
Type of distortion % of correct recognitions % of wrong recognitions

Shift (13 pixels) 100.0 0.0

Rotation (1°–10°) 0.0 0.0

Intensity 5.0 0.0

Other – scale, pinch, punch, occlusion 00.0 0.0

“Real-life” samples 21.4 28.6

Table 6. Performance of the system with preprocessing (only wavelet transform – Haar 2), 1-3-3-2 FELSI, decay parameter 10–5,
learning error equal to 2.0´10–3 and [–0.5,0.5] noise range.

Type of distortion % of correct recognitions % of wrong recognitions

Shift (13 pixels) 100.0 0.0

Rotation (1°–10°) 52.5 10.0

Intensity 20.0 0.0

Other – scale, pinch, punch, occlusion 30.0 0.0

“Real-life” samples 0.0 21.4

Opto-Electron. Rev., 11, no. 4, 2003 M. Antkowiak 335

Fingerprint identification by using artificial neural network with optical wavelet preprocessing

4.4. ANN with only wavelet transform in the The recognition is performed by the artificial neural
preprocessing stage network with built-in position invariance. Proper choice of
the learning method additionally provides rotation-, scale-,
As the final test, the same system but with wavelet prepro- and intensity-invariance. To some extent also occluded im-
cessor instead of edges-enhancement filter was examined ages can be properly identified.
(Table 6). The discrimination capability of this configura- The modified resilient backpropagation algorithm,
tion is not sufficient to provide reliable identification of which proved to be a significant improvement over the pre-
distorted samples, especially when rotation is considered. viously implemented standard backpropagation with mo-
In case of other deformations, the network usually is not mentum, can be successfully used in the learning process.
able to give answer. The techniques of additional noise in the input layer and
This indicates low generalization ability. The low level weight decay, joined with proper architecture of the net-
of selectivity of the feature extraction process is the proba- work, noticeably improve the capabilities of the system.
ble cause of this situation. It is visible that only the com- The presented results confirm the assumption that pre-
bined method of edges-enhancement and wavelet-based processing based on feature extraction enhances the gener-
feature extraction gives satisfactory results. In hardware alization ability of an artificial neural network. Moreover,
implementations of the preprocessor, both these operations the discrimination capability is also significantly improved.
can be simultaneously performed by the properly prepared Table 7 summarizes the best results obtained with various
filter. configurations of the examined set-up when controlled de-
formations were tested.
5. Conclusions Surprisingly low rates of recognitions for system using
only wavelet processor are probably caused by the em-
In this paper, a system consisting of an optical wavelet pre- ployed method of interpreting the output signals. Possibly,
processor and an electronic artificial neural network mod- if some less demanding method had been used instead,
ule has been examined. The aim was to check whether such both correct and wrong recognitions rate would have been
system could be successfully employed to identify and dis- higher. Nevertheless, as it can be seen in the case of the last
criminate highly complex biometric images, such as finger- system, it is achievable to attain a reasonably high rate of
prints. The required capabilities included shift-, rotation-, correct recognitions and, at the same time, significantly re-
scale- and intensity-invariance. The additional constrain duce the probability of misidentification. The system using
was the limited size of the training set. In fact, only one both edges-enhancement and wavelet processor in the pre-
print of each finger was used as a training pattern, which processing stage, provides very high reliability and satis-
was a significant impediment for the learning process. factory percentage of successful recognitions. At present,
Since, as it is shown above, performance of ANN work- the results for the “real-life” data are not as promising as
ing on raw data is not satisfactory, the preprocessing based for the artificially prepared samples.
on edges-enhancement and wavelet transform is proposed In future, the system’s performance can be further im-
to enhance the generalization ability. Behaviour of various proved in several ways. First of all, the accuracy of the data
versions of the described set-up is presented. acquisition can be improved. Moreover, the size of training
In the optimal configuration proposed in this paper, the set can be increased. As it has been already mentioned, in
edges-enhancement filter first improves the quality of an further research other types of wavelets, especially the
input image. Owing to this procedure, the ridges in a fin- Morlet’s wavelet, should be tested. Another improvement
gerprint are more distinct. Subsequently, the optical wave- can be achieved by using the potential of CGH and em-
let processor is employed. The Haa’s wavelet of a certain ploying simultaneously more than one wavelet and several
scale is used to extract specific features of an image. In fu- ANN units. Such parallel system could provide better rec-
ture research other types of wavelet functions may be suc- ognition rate with “real life” data, owing to the approach
cessfully applied. based on various features of the input image.

Table 7. The best obtained results of the set-ups with various methods of preprocessing. Only the controlled distortions (intensity, rota-
tion, scaling, occlusion, pinch, punch) are considered.

Method of preprocessing % of correct recognitions % of wrong recognitions

No preprocessing 50.00 6.25

Wavelet processor 6.25 3.75

Edges-enhancement filter 40.00 8.75

Edges-enhancement + wavelet processor 88.75 0.00

336 Opto-Electron. Rev., 11, no. 4, 2003 © 2003 COSiW SEP, Warsaw
 Contributed paper
References 9. M. Antkowiak, “Application of optical wavelet transform as
a preprocessor for the artificial neural network in
1. S. Prabhakar, S. Pankanti, and A.K. Jain, “Biometrics rec- biometrics”, M.Sc. Thesis, Warsaw University, Warsaw
ognition: security and privacy concerns”, IEEE Secu- 2002.
rity&Privacy Magazine 1, 33–42 (2003). 10. Y. Li and Y. Sheng, “Wavelets, optics, and pattern recogni-
2. A. Ross, A. Jain, and J. Reisman, “A hybrid fingerprint tion”, Chapter 3 in Optical Pattern Recognition, edited by
matcher”, Pattern Recognition 36, 1661–1673 (2003). F.T.S. Yu and S. Jutamulia, Cambridge University Press,
3. J. Rodolfo, H. Rajbenbach, and J.P. Huignard, “Perfor- Cambridge, 1998.
mance of a photorefractive joint transform correlator for 11. K. Hirokawa, K. Itoh, and Y. Ichioka, “Real-time optical
fingerprint identification”, Optical Engineering 34, wavelet-transform with positive and negative signals”, Op-
1166–1171 (1995). tical Review (Japan) 4, 366–369 (1997).
4. Y. Pétillot, L. Guibert, and J.L. de Bougrenet de la Tocnaye, 12. Y. Sheng, D. Roberge and H. Szu, “Optical wavelet trans-
“Fingerprint recognition using a partially rotation invariant form”, Optical Engineering 31, 1840–1845 (1992).
composite filter in a FLC joint transform correlator”, Optics 13. K. Hirokawa, K. Itoh, and Y. Ichioka, “Optical wavelet pro-
Communications 126, 213–219 (1996). cessor by holographic bipolar encoding and joint transform
5. A. Stoianov, C. Soutar, and A. Graham, “High-speed fin- correlation”, Applied Optics 36, 1023–1026 (1997).
gerprint verification using an optical correlator”, Optical 14. K. Hirokawa, K. Itoh, and Y. Ichioka, “Optical wavelet
Engineering 38, 99–107 (1999). transform using an LCTV-based joint transform correlator”,
6. K. Cha³asiñska-Macukow and R. Kotyñski, “Correlation SPIE’s International Technical Group News Letter 11, 6–7
methods in optical encription and security”, in Perspetives (2000).
in Modern Optics & Optical Instrumentation, pp. 197–193, 15. D. Mendlovic and N. Konforti, “Optical realization of the
edited by J.J. Joseph and K. Singh, Anita Publications, New wavelet transform for two-dimensional objects”, Applied
Dehli, 2002. Optics 32, 6542–6546 (1993).
7. K. Hirokawa, K. Itoh, and Y. Ichioka, “Invariant pattern rec- 16. University of Bologna, Biometrics System Lab. Database
ognition using neural networks combined with optical wavelet DB2_B made available as a part of the Fingerprint Verifica-
preprocessor,” Optical Review (Japan) 7, 284–293 (2000). tion Competition, http://bias.csr.unibo.it/fvc2000/Down-
8. W. Zhang, K. Itoh, J. Tanida, and Y. Ichioka, “Parallel dis- loads/DB_B.zip
tributed processing model with local space-invariant inter- 17. A. Cichocki and R. Unbehauen, “Neural Networks for Opti-
connections and its optical architecture”, Applied Optics 29, misation and Signal Processing”, John Wiley & Sons, New
4790–4797 (1990). York, 1993.

Opto-Electron. Rev., 11, no. 4, 2003 M. Antkowiak 337

 Dekker is a digital publisher that offers authoritative scientific, technical, & medical
content accessible at the article level with linked references

Encyclopedia of Optical Engineering

IN 3 VOLUMES
Edited by: Ronald Driggers

Purchase Options
List Price: $1,500.00
FREE ONLINE SUBSCRIPTION for one year with purchase
of print edition!

Institutional orders for Dekker Encyclopedias cannot be placed through

dekker.com. Subscriptions will be governed by an access/site license.
Please contact sitelicenses@dekker.com
or call 1-800-228-1160 (USA, Canada & South America)
or 0041-61-260-63-00 (Europe, Far East, Middle East & Africa)
for ordering information.

Encyclopedia | Print Published: 09/01/2003 | Online Published: 09/01/2003

Hard Cover
2800 pages | Illustrated
ISBN: 0-8247-4258-3

Description

From astronomy to x-ray optics, this Encyclopedia contains more than 230 vivid entries examining the most
intriguing technological advances and perspectives from distinguished professionals around the globe-selecting
topics of utmost importance in areas including digital image enhancement, biological modeling, biomedical
spectroscopy, and ocean optics for thorough coverage of recent applications in this continually expanding field.

Compiled by 300 of the most widely respected names in the electrooptic sciences, the Encyclopedia is
destined to serve as the premiere guide in the field with nearly 2000 figures, 560 photographs, 260 tables,
and 3800 equations.

Color graphics available online!

This one-of-a-kind reference delivers top-notch research on

• target detection
• image restoration
• fiber optic sources
• holography
• lens design
• classification and coding
• optical biopsy
• radiometry

Timely and dynamic, the Encyclopedia is an all-inclusive resource of the most significant and essential
breakthroughs in photonics, imaging, and wireless communication.

338 Opto-Electron. Rev., 11, no. 4, 2003 © 2003 COSiW SEP, Warsaw