| License
Plate Recognition Using Image Processing Techniques & SVM Classifier |
by Ido Shemesh and David Arieh Fellman Supervised by Dori Peleg |
Abstract
One of the things necessary for
designing a completely automatic system for opening a gate (for the
faculty
parking lot, in our case) without installing a signal transmitter in
each car,
is a system which can read the license plate of an approaching vehicle,
and see
whether or not it is permitted to enter.
In our project we designed such a
system. First we separated each digit
from the license plate using image processing tools.
Then we built an SVM classifier, using a
training set based on digits extracted from approximately 430 license
plates. Finally, we built a Graphical
User Interface for selecting a picture of a license plate, to identify
the
number on it.
 |
The Problem
License plates come in different
sizes and in different Width-Height ratios, the fonts used for digits
on
license plates are not the same for all license plates.
These problems, and the changing weather conditions,
are what make the field of License Plate Recognition a good candidate
for
testing Pattern Recognition techniques, such as SVM.
The system is built to be able to construct a
new training set at the moment, for later use, however we would
recommend
designing a tool that will enlarge an existing set.
The solution
The
method we
used for deciphering the numbers from the images was first to separate
the figures
of the digits from the total image of the license plate.
This is done by first transforming the
grayscale picture of the license plate to a black-white image. The threshold for this transformation was
first determined by using the 'graytresh' Matlab function, and if the
result
was insufficient, various thresholds were tried, until the most
successful of
those is found.
The
separation of the digits is done by first filtering all objects which
are not
likely to be digits, because of their dimensions, their location or
their
orientation.
The
method used for identifying the digits is SVM (Support Vector Machine). This method receives a training set of
labeled feature vectors, and uses it to separate a Hilbert Space into
decision
areas. If a linear separation is applied
the separating spaces will be in the form of 
.
Otherwise, they will be in
the form of 
.
A
new features
vector will be classified
according to which decision area it fits in
 |
| Figure 1 -
Filtering Process of unwanted LP noises |
 |
| Figure 2 -
Gui for building Training set for the SVM classifier |
 |
| Figure 3 -
Gui for identifying number on a license plate using SVM classifier |
Tools
This
project was developed in a Matlab 6.5, on a PC platform environment, using the
OSUSVM
toolbox. The License plates were filmed
using
a Watec WAT-202 video camera,
transferred to a
computer using Adobe
Premiere,
and separated into frames using Virtual Dub.
Conclusions
In this
project, a system for recognizing the number off a license
plate was designed. For this goal, we used 432 photos of license
plates, from which we received 3011 images of digits. In this process,
approximately 0.03% of the digits were lost. The classification has two
parts: Separation of the digits from the license plate, and
classifying each digit using SVM. After the classification, the digits
are linked into one number. The best Classification Rate received was
98.141% for each of the methods using SVM (Linear, Polynomial, or
Gaussian). Therefore, the use of a Linear Classifier is recommended,
since it is the quickest and easiest. This system is
operational.
Acknowledgment
We would like
to thank the following people:
Related documentation