License Plate Recognition

11
A Fast Algorithm for License
Plate Detection (LPD(
Prof. Mohy E. Abou El-Soud, Dr. Mohamed Abdel-
Azim, and Eng. Amr E. Rashid
Faculty of Engineering- Mansoura University, Mansoura, Egypt

22
Outlines
 IntroductionIntroduction
 Motivations.Motivations.
 Constraints and Data CollectionConstraints and Data Collection
 Problem DefinitionProblem Definition
 Previous WorkPrevious Work
 The Proposed TechniqueThe Proposed Technique
 Results and ConclusionResults and Conclusion
 Future WorkFuture Work

33
Introduction
 Monitoring vehicles for law enforcement and
security purposes is a difficult problem because of
the number of automobiles on the road today.
 An example is this lies in border patrol:
 It is time consuming for an officer to physically check
the license plate of every car.
 Additionally, it is not feasible to employ a number of
police officers to act as full-time license plate inspectors.
 Police patrols cannot just drive in their cars staring
at the plates of other cars.
 There must exist a way for detecting and identifying
license plates without constant human intervention.
 As a solution, we have implemented a system that can
extract the license plate number of a vehicle from
an image given a set of constraints.

44
Introduction (Cont’d(
In any object recognition system, there are
two major problems that need to be solved
Detecting an object in a scene, and
Recognizing that object.
In our system, the quality of the license
plate detector (LPD) is doubly
important since the make and model
recognition (MMR) subsystem uses the
location of the license plate as a reference
point when querying the car database.

55
License Plate Recognition (LPR) may
also be referenced as:
Automatic Vehicle Identification (AVI).
Car Plate Recognition (CPR).
Automatic Number Plate Recognition
(ANPR).
Car Plate Reader (CPR)
Optical Character Recognition (OCR) for
Cars.
Introduction (Cont’d(

66
Motivations
This area is challenging because it requires
an integration of many computer vision
problem solvers, which include:
Object detection (LPD).
Character recognition (OCR).
LPR is very important in:
Private transport applications.
Monitoring vehicles for law enforcement and
security purposes is a difficult problem because
of the number of automobiles on the road today

77
WhatWhat’’s LPRs LPR??
License Plate Recognition:License Plate Recognition:
LPR is an image-processing based-LPR is an image-processing based-
technology used to identify vehiclestechnology used to identify vehicles
by their license plates.by their license plates.
This technology is used in variousThis technology is used in various
security and traffic applications.security and traffic applications.

99
 LPR is one of the most important types of
intelligent transport system and is of considerable
interest because of its potential applications to
many areas such as:
 highway electronic toll collection,
 traffic monitoring systems and
 ...
 The technology concept assumes that all vehicles
already have the identity displayed (the plate!) so
no additional transmitter or responder is
required to be installed on the car.
Technology HighlightsTechnology Highlights

1010
Problem DefinitionProblem Definition
License plates come in:
Different sizes,
Different Width-Height ratios,
Different color,
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 LPR a
good candidate for testing Pattern Recognition
techniques.

1111
Constraints…
 Use a digital camera,
 Image of the vehicle taken with variable
angles,
 Image of the vehicle taken from fixed
distance (about 1-2 m),
 Vehicle is stationary when the image
was taken,
 Only Egyptian license plates will be
processed.

1212
Data collectionData collection
 All images of vehicles database were taken with a
benq digital camera, with three different
resolutions: (i) 3 M-Pixels, (ii) 4 M-Pixels, and (iii) 5
M-Pixels.
 On average, the images were taken (1-2m) away
from the vehicle.
 They were stored in color JPEG format on the
camera.
 The colored JPEG images were converted into gray
scale raw format on the PC.
 There 30 images dataset.

1313
Data Collection (ContData Collection (Cont’’dd((
Original Color Image

1414
Data Collection (ContData Collection (Cont’’dd((
Gray Scale Image

1515
Previous WorkPrevious Work
The current LPD techniques can be
classified into four main algorithms:
Corner template matching,
Hough transforms combined with
various histograms based methods,
color based filter, and
Vertical edge detection followed by
size and shape filtering

1616
Algorithm-1: Vertical Edge
Detection
Candidate selection:Candidate selection:
Histogram equalization,Histogram equalization,
Binarization,Binarization,
Sobel edge detection, andSobel edge detection, and
List possible licensesList possible licenses
For each candidate:For each candidate:
Localized histogram,Localized histogram,
Binarization, andBinarization, and
Elimination by 2-D correlationElimination by 2-D correlation
16

1717
Algorithm-2:Algorithm-2: Vertical Edge
Detection
This algorithm used a recognition
algorithm based on width to height ratio:
Vertical edge detection,
Size and shape filtering,
Vertical edge matching, and
Compute the Black to white ratio and then
perform plate extraction.

1818
Drawbacks
There are many problems in these two
algorithms:
Width to height ratio differs from a car to another
depending on the distance between the camera
and the car,
Small vertical edges will difficult the recognition
problem because it change the width between
edges,
When we use different view this will remove
desired vertical edges, and
There are many objects in the image achieves
equal width to height ratio

2020
Algorithm-3: AdaBoost AlgorithmAlgorithm-3: AdaBoost Algorithm
 Since license plates contain a form of text, we
decided to face the detection task as a text
extraction problem.
 Window search over the entire frame.
 Use three different sized windows.
 Independent Classifier for Each Size
 Strong Classifier Constructed from Weak
 Classifiers Via AdaBoost algorithm .
 Computationally Simple.
 Draw backs: Regions contain character except
license plate.

2121
Drawbacks of AdaBoost Algorithm

2222
The Proposed Algorithm
 The proposed algorithm was divided into
four main parts:
 Histogram Equalization,
 Removal of Border and Background,
 Image Segmentation, and
 License Plate Detection.

2323
A. Histogram Equalization
Is an image transformation that computes
a histogram of every intensity level in a
given image and stretches it to obtain a
more sparse range of intensities.
This manipulation yields an image with
higher Contrast than the original.

2424
Original image
Remove
this
partition
Remove this
partition

2525
The Output Image of Histogram
Equalization

2626
B. Removal of Border and
Background
Sobel Vertical edges

2727
B. After Removing SmallB. After Removing Small
ElementsElements

2929
Horizontal Edges After Removing
Small Elements

3030
Car after removing border and
back ground

3131
C. Image Segmentation
Often the license plate will be in the lowerOften the license plate will be in the lower
half of the image so we will remove upperhalf of the image so we will remove upper
half of the image.half of the image.

3232
D. License Plate Detection
Feature extraction.Feature extraction.
Principal component analysis.Principal component analysis.
Artificial neural networks.Artificial neural networks.

3333
D.1 Feature Extraction
Feature extraction is the transformation of
the original data (using all variables) to a
dataset with a reduced number of
variables.
In the problem of feature selection, the aim
is to select those variables that contain the
most discriminatory information.

3434
D.1 Feature Extraction (Cont’d(
There are several reasons for performing
feature extraction:
To reduce the bandwidth of the input data.
To provide a relevant set of features for a
classifier.
To reduce redundancy.
To recover new meaningful underlying
variables or features that the data may easily be
viewed.

3535
D.1 Feature Extraction (Cont’d(
Wavelets have been demonstrated to give
quality representations of images.
This DWT representation can be thought of
as a form of “feature extraction” on the
original image
We will use Haar-like features, where sums
of pixel intensities are computed over
rectangular sub-windows.

3636
D.2 Principal component
analysis (PCA(
In some situations, the dimension of the
input vector is large, but the components of
the vectors are highly correlated
(redundant).
It is useful in this situation to reduce the
dimension of the input vectors.
An effective procedure for performing this
operation is PCA.

3737
D.3 Recognition Stage UsingD.3 Recognition Stage Using
ANNsANNs

3838
D.3 Recognition Stage UsingD.3 Recognition Stage Using
ANNsANNs
 Apply ANN with adaptive sub-window:Apply ANN with adaptive sub-window:
 See final outputSee final output

Second techniqueSecond technique
(A) Image enhancement(A) Image enhancement
(B) Removal of Border and Background(B) Removal of Border and Background
(C) Image Segmentation(C) Image Segmentation
(D) License Plate Detection using 2D(D) License Plate Detection using 2D
correlation.correlation.
3939

4040

4141

Image enhancement(Wiener filter outputImage enhancement(Wiener filter output((
4242

Removal of boarder and backgroundRemoval of boarder and background
4343

After removing small elementsAfter removing small elements
4545

Horizontal edgesHorizontal edges
4646

Image segmentationImage segmentation
4747

Vertical edge detectionVertical edge detection
It is observed that most of vehicles usuallyIt is observed that most of vehicles usually
have more horizontal lines than verticalhave more horizontal lines than vertical
lines. To reduce the size of the imagelines. To reduce the size of the image
vertical edges are detected.vertical edges are detected.
 this help in extracting the license platethis help in extracting the license plate
exactly from segmented image, even it isexactly from segmented image, even it is
out of shape,out of shape,
4848

Image after applying vertical edge detectionImage after applying vertical edge detection
4949

Matching by correlationMatching by correlation
 The correlation problem is to find all places in the imageThe correlation problem is to find all places in the image
that match a given sub image (also called a mask orthat match a given sub image (also called a mask or
template)template)
 Typically mask image is much smaller thanTypically mask image is much smaller than
Original imageOriginal image
 One approach for finding matches is to treat mask imageOne approach for finding matches is to treat mask image
as spatial filter and compute the sum of products (or aas spatial filter and compute the sum of products (or a
normalized version of it) for each location of mask imagenormalized version of it) for each location of mask image
in . Then the best match (matches) of subimage inin . Then the best match (matches) of subimage in
original image is (are) the location(s) of the maximumoriginal image is (are) the location(s) of the maximum
value(s) in the resulting correlation image.value(s) in the resulting correlation image.
5151

Cont’dCont’d
For prototyping. An alternative approach isFor prototyping. An alternative approach is
to implement correlation in the frequencyto implement correlation in the frequency
domain.domain.
Making use of the correlation theoremMaking use of the correlation theorem
Which like the convolution theorem.Which like the convolution theorem.
 Relates spatial correlation to the product ofRelates spatial correlation to the product of
the image transforms.the image transforms.
5252

5454
Conclusion & Results
Finally, we have built an LPD system that is:Finally, we have built an LPD system that is:
Real-time,Real-time,
Works well with inexpensive cameras, andWorks well with inexpensive cameras, and
Does not require infrared lighting or sensors asDoes not require infrared lighting or sensors as
are normally used in commercial LPR systems.are normally used in commercial LPR systems.
There no database for Egyptian license plateThere no database for Egyptian license plate
and there is no standard license plate inand there is no standard license plate in
Egypt.Egypt.
We achieved 93.33% detection rate for smallWe achieved 93.33% detection rate for small
dataset; i.e., 28 license plate of 30.dataset; i.e., 28 license plate of 30.

5555
Future work
Modern FPGA platforms provide theModern FPGA platforms provide the
hardware and software infrastructure forhardware and software infrastructure for
building a bus-based system on chipbuilding a bus-based system on chip
(SoC) that meet the applications(SoC) that meet the applications
requirements.requirements.
In order to accelerate the system we canIn order to accelerate the system we can
implement ANN classifier using FPGA withimplement ANN classifier using FPGA with
parallel processing instead of using Matlabparallel processing instead of using Matlab
.we expect that we can achieve an overall.we expect that we can achieve an overall
LPR system speed up.LPR system speed up.

License Plate Recognition

More Related Content

What's hot (20)

Viewers also liked (13)

Similar to License Plate Recognition (20)

More from Amr Rashed (20)

Recently uploaded (20)

License Plate Recognition

Editor's Notes