Car Plate Recognition by Neural Networks and
Image Processing Using Integration of Wavelets
Jamal Fathi Abu Hasna
Abstract-This paper presents an experimental system for the
recognition of North Cyprus-style car license plates. Images are
usually taken from a camera at a toll gate and preprocessed by a fast
and robust 1-D DFT scheme to find the plate and character
positions. We examine a new wavelet video method of processing
signals for continuous-wave.
Keywords- Feedforward, Denoising, WaveNet, and Discrete
Fourier Transform.
I. INTRODUCTION
C ar license recognition is important in several fields of
application:
traffic control in restricted areas;
automatic payment of tolls on highways or bridges;
general security systems wherever there is the need of
identifying vehicles.
Some approaches exist and have been described in literature.
They are mainly based on pattern matching and normalized
correlation with a large database of stored templates.
In this paper we describe an experimental system for the
recognition of Cyprus-style car license plates. The system is
based on the use of a feedforward neural network (FNN). This
learning approach has been shown to guarantee high rates of
convergence and properties of stability and robustness of the
solution. The data at hand consist of digitized images of cars,
acquired by a high-resolution x4 photo camera and collected
in a Photo CD. The processed images (see Fig.1) are 390 by
480 pixels wide. The distance and the angle of view simulate
a car passing through a toll gate.
Manuscript received June 26, 2006. This work was supported in part by the
Near East University, Electrical & Electronics Engineering, North Cyprus,
Turkey via Mersin-10, KKTC.
Jamal Fathi Abu Hasna was born in Tulkarm on November 19, 1964. He came
to the North Cyprus in 1997, and graduated with a
Bachelor of Science degree in Electrical & Electronics
Engineering from Near East University in 1998-99. He
continued to the graduate program at Near East
University and completed the Master of Science in 2000-
2001 and continued the Doctor of Philosophy degrees in
2001 and now in process.
His current interests include neural computing, adaptive signal processing,
cellular communications, control systems, and wavelets,
Phone: 00905338658472, mail: jamalfathi2004@yahoo.com,
The recognition process starts with the search and the
extraction of the portion of the original image containing the
car plate, or even any part of the car. The characters
contained in the plate are localized by a robust processing
using a non-traditional Discrete Fourier Transform (DFT),
and subsequently isolated and classified by the neural
network.
The scores are validated by a post-processor which takes into
account the syntax of Cyprus-style plates.
The processing of video data in real time is considered to be
somewhat impractical given the current state of technology.
The utility of such processing in real-world applications
would therefore seem to be limited. However, recent
developments at Trident Systems, Incorporated have made
available real-time wavelet processing of video, in the form of
the WaveNet technology [2]. Also, in the future a variety of
fast architectures for computing wavelet transforms will
surely be developed. With respect to a recently published
work, our approach is able to reduce the complexity of the
learning phase (no feature extraction and pattern matching
are required). The character recognition has been speeded up
by the parallel architecture of the FNN. The algorithm has
been tested on a workstation and the Matlab software.
Fig. 1 Original Image
II. PREPROCESSING WITH WAVELET DENOISING
To improve performance for noisy Doppler signals, we apply
Donoho’s O (n) wavelet denoising algorithm [3]. The
algorithm first does the discrete wavelet transform with
Mallat’s pyramid algorithm [4].
The pyramid algorithm computes the transform for some J
dyadic levels of scale, resulting in vectors of detail and
smooth wavelet coefficients d1, d2,…,dJ-1,sJ.
The algorithm then shrinks the detail coefficients for scales
j  J  1 to obtain d1 , d2 ,..dJ 1 . Here the d j are
d   (d ) ,
j  j j j
where   ( x) is a nonlinear threshold shrinkage function III. ARTIFICIAL NEURAL NETWORKS FOR PATTERN
RECOGNITION
given by
The continuous wavelet transform correlates a Doppler signal
with time-localized wavelets at various scales and shifts. It
0 if | x |  gives the change in local signal scale over time, which in this
  ( x )  
case is the Doppler period or inverse frequency. When a
 sign( x)(| x |  ) if | x | 
moving window is placed on the incoming Doppler signal
This threshold shrinkage function is shown in Fig 2.
and the windowed signal is wavelet transformed, the
δλσ(x) corresponding time-varying transform imagery constitutes
video. Samples of this wavelet-generated video over time
then form signal features for pattern recognition neural
networks. These networks are then trained to extract the
Doppler frequency shift over time. This frequency shift is
-λσ
critical information for proximity sensing.
x The continuous wavelet transform constitutes a frame rather
λσ than a basis. Such a redundant representation allows more
flexibility in the selection of signal features. In terms of the
most efficient signal representation, these features should be
orthogonal. However, such a representation in which the
features are completely independent is less robust with respect
to noise immunity and fault tolerance. The search for the best
Fig. 2 Nonlinear threshold shrinkage function for wavelet denoising representation is therefore a tradeoff between redundancy and
robustness [5].
The threshold shrinkage function δλσ(x) is parameterized by a
threshold λand an estimate of the standard deviation of the
noise σ. We use a universal threshold  j  2log( N ) ,

where N is the number of data samples [56]. For  we use

the median absolute deviation, which is a robust estimation of
standard deviation. Fig. 4 Noisy Image
Finally, the wavelet denoising algorithm computes the inverse
discrete wavelet transform using the new coefficients
d1 , d2 ,..dJ 1 , d J , s J
This results in a non-parametric estimate of the signal
without the noise. The entire wavelet denoising algorithm is
shown in Fig. 3. Coefficient Fig. 5 Denoised Image
Shrinkage

d1
We extract the Doppler shift with feedforward multilayer
neural networks, known as multilayer perceptrons [6]. After
computing the continuous wavelet transform of the denoised
Denoised Doppler signal, we sample the transform coefficients to
Noisy d2
Image provide inputs for the multilayer perceptrons. The networks
Image
are trained with the Levenberg-Marquardt [7][8] rule to
Inverse provide the Doppler shift at a given time. This rule is a
DWT d3 DWT powerful generalization of gradient descent that employs an
approximation of Newton’s method. It is much faster than
standard gradient descent algorithms such as
d4 backpropagation, although it does require more memory.

IV. PROPOSED MODEL

s4
In our scheme we place a window of fixed width over the
Wavelet incoming signal, so as to localize the processing near the
Coefficients present time. We then perform a continuous wavelet
Fig. 3 Wavelet denoising algorithm, where noisy Image and the transform on the signal within the window, resulting in an
denoised one are shown in fig 4, and fig 5. image of the transform. As the signal window then moves
forward in time, the corresponding sequence of transform
 images forms a video. From this wavelet transform video, we less useful, smaller scale signal structures can therefore be
then extract features as input to pattern recognition disregarded. Neglecting unnecessary details allows a
algorithms such as artificial neural networks. This is shown reduction in the amount of data to be processed. This in turn
in fig 6. reduces the complexity of the processing, leading to
Our model has the generation of wavelet transform video improvements in processing time, system size, and system
from time varying camera proximity fuze signal. Temporary cost.
expansion of dimensionality allows us to extract salient This reduction of data through the explicit use of scale is a
features, leading to reduced computational complexity powerful form of data compression. While there are several
other strategies for data compression, this one has the
advantage of being based on the extraction of signal features.
One- Through wavelet transform time integration, a single
Dimensional coefficient provides the correlation between the signal and a
window wavelet at a particular scale and time shift. Wavelets are
Sensor Time known to provide good signal features for pattern recognition
signal algorithms such as artificial neural networks. Indeed, natural
Time-Varying Moving In Time
sensors such as eyes and ears carry out wavelet-type
Continuous wavelet signal in processing.
transform window The continuous wavelet transform effectively increases the
Two-Dimensional
Shift window dimensionality of the signal representation from one to two.
While this might cause some concern at first glance, it is
really not a problem. The reason is that the wavelet
Scale representation will be used to extract signal features only.
Thus the pattern recognition neural networks need not suffer
Wavelet Video from the “curse of dimensionality.” After all, the extracted
features Moving In Pattern features are of a single dimension only, so that the increase in
Time Recognition dimensionality is only temporary. Indeed, because of the high
Video From time- Neural Net quality of wavelet features, it is quite possible that fewer
varying Image in features will be needed, and that recognition performance will
window be improved.
Mallat’s multiresolution analysis [9] leads to discrete
Fig. 6 Proposed Model orthogonal wavelets at dyadic scales and shifts, implemented
via the efficient pyramid algorithm. These discrete wavelets
Proximity sensing is widely applied in manufacturing have been successful in many applications, particularly data
automation and robotics. More recently, there has also been a compression. However, discrete wavelets have limited utility
strong interest in proximity sensing for automobile collision for pattern recognition problems. This is because interesting
avoidance. Our proposed method is also applicable to signal structures are not constrained to follow such power-of-
processing signals in sonar sensors. two patterns. In particular, discrete wavelet transform
coefficients are shift-variant, which in general causes
V. ADVANTAGES OF OUR APPROACH problems for pattern recognition.
Besides, in our proposed scheme, wavelet video processing is In contrast, the continuous wavelet transform has coefficients
not a particular computational hindrance, but rather allows at all scales and shifts, not just dyadic ones. The continuous
salient features to be extracted via the wavelet coefficients. transform therefore has the desirable property of shift
Because of the quality of the wavelet video features, it is invariance. Another advantage of continuous wavelets is that
likely that fewer numbers of inputs will be needed for pattern they have less stringent requirements for admissibility, which
recognition. In this sense our scheme could be considered to allows a wider choice of basis functions. They also have the
be a form a data compression. possibility of being basis functions for adaptive wavelet
In particular, it seems to be a form of data compression that is networks.
ideal for pattern recognition. Through the inclusion of all scales and shifts, the continuous
The multiresolution nature of wavelets also allows us to wavelet transform effectively increases the dimensionality of
explore the tolerance of imprecision in the processing of the signal representation. That is, the representation is made
signals. This provides the freedom to tailor the design of the to be a function of two variables rather than one. We note
sensor to the resolution requirements of the signals being that the discrete wavelet transform introduces no such
processed. This tolerance of imprecision is in the spirit of increase in dimensionality, since the number of transform
fuzzy logic, but in this case the imprecision is in the scale of coefficients is the same as the number of signal sample
the signal structures rather than in the membership of sets. points. This is because the discrete wavelet transform
The important idea is that useful information in signals is employs an orthonormal basis rather than an over complete
generally found at the larger scales (lower frequencies). The frame.
However, the fact that we are using the continuous wavelet
transform coefficients merely for feature extraction means
that we need not be plagued by the curse of increased
dimensionality. In particular, the goal is to use only the
relatively few coefficients that provide the best features. In
fact, the use of such high quality features may well mean that (e)
fewer numbers of inputs will be needed for the pattern
recognition neural networks. Of course, these high quality
features are also likely to improve the performance of the
neural networks.
Our scheme could therefore be considered a form a data
compression. The temporary increase in dimensionality
could then improve compression quality, at least when
measured with respect to pattern recognition performance.
If we disregard the issue of dimensionality, it might still be
argued that computation of the discrete wavelet transform is
faster, which has complexity O (n) . However, a continuous
wavelet transform implemented via the fast Fourier transform (f)
has complexity O (n log n) , which is still quite acceptable for
many applications. Also, a continuous wavelet transform has
the potential for massive parallelism, and allows the
possibility of adaptive wavelet bases [10]

VI. SOME EXPERIMENTS

These practical experiments taken using the proposed model
for not only the plates, also for the body of the car, and these Fig. 7 (a), (b), and (c) Samples of wavelet-generated video for
experiments are as shown in fig 7 below. training and testing inputs to pattern recognition neural networks,
while (d), (e) and (f) are the outputs of the proposed model.

VII. CONCLUSION
(a) We have just demonstrated the effectiveness of features
extracted from wavelet-generated video, and We tested the
pattern recognition performance of such features in the
estimation of time varying Doppler shift from noisy sensor
signals. In particular, we sampled the wavelet video; we saw
that the frequency estimation performance of the neural
(b) networks is overall good.

ACKNOWLEDGMENT
My deepest thanks are to Prof. Dr. Senol Bektas, and to Prof.
Dr. Fakhraddin Mamedov. I would like to express my
gratitude to my collegues in the department. Also I would like
to express my gratitude to my family.
(c)
REFERENCES
[1]. R.Parisi, E.D. Di Claudio, G. Orlandi and B.D. Rao, “A
generalized learning paradigm exploiting the structure of
feedforward neural networks,” IEEE Trans. on Neural
Networks, vol.7, no.6, November 1996.
[2]. P. Comelli, P. Ferragina, M. Notturno Granieri, and F.
Stabile, “Optical recognition of motor vehicle license
(d)
 plates,” IEEE Trans. On Vehicular Technology, Vol.
44, No. 4, November 1995, pp. 790-799.
[3]. S.Haykin, Neural Networks-A Comprehensive
Foundation, IEEE Press, 1994.
[4] A. Papoulis, Probability Random Variables and
Stochastic Processes, Mc Graw-Hill, New York, 3rd
Edition, 1991.
[5] W.A. Gardner, Statistical Spectral Analysis: A Non-
Probablistic Theory, Prentice Hall, 1988.
[6] P.J. Huber, Robust Statistics, John Wiley, New York,
1981. 0-7803-4455-3/98/$10.00 (c) 1998 IEEE
[7]. Antoniadis, A.; G. Oppenheim, Eds.(1995), Wavelets and
statistics, Lecture Notes in Statistics 103, Springer
Verlag.
[8]. Burke Hubbard, B. (1996), The world according to
wavelets, AK Peters, Wellesley. The French original
version is titled Ondes et Ondelettes. La saga d'un outil
mathématique, Pour la Science, (1995).
[9]. Chui, C.K. (1992a), Wavelets: a tutorial in theory and
applications, Academic Press.
[10]. Cohen, A. (1995), Wavelets and multiscale signal
processing, Chapman and Hall.
[11]. Cohen, A.; I. Daubechies, B. Jawerth, P. Vial (1993),
"Multiresolution analysis, wavelets and fast wavelet
transform on an interval," CRAS Paris, Ser. A, t. 316,
pp. 417-421.
[12]. Coifman, R.R.; Y. Meyer, M.V. Wickerhauser (1992),
"Wavelet analysis and signal processing," in Wavelets
and their applications, M.B. Ruskai et al. (Eds.), pp. 153-
178, Jones and Bartlett.
[13]. Teolis, A. (1998), Computational signal processing with
wavelets, Birkhauser.
[14]. Vetterli, M.; J. Kovacevic (1995), Wavelets and subband
coding, Prentice Hall.