International Journal of Image Processing (IJIP) Volume (1) Issue (2)

Editor in Chief Professor Hu, Yu-Chen

International Journal of Image Processing
(IJIP)
Book: 2007 Volume 1, Issue 2
Publishing Date: 30-08-2007
Proceedings
ISSN (Online): 1985-2304

This work is subjected to copyright. All rights are reserved whether the whole or
part of the material is concerned, specifically the rights of translation, reprinting,
re-use of illusions, recitation, broadcasting, reproduction on microfilms or in any
other way, and storage in data banks. Duplication of this publication of parts
thereof is permitted only under the provision of the copyright law 1965, in its
current version, and permission of use must always be obtained from CSC
Publishers. Violations are liable to prosecution under the copyright law.

IJIP Journal is a part of CSC Publishers
http://www.cscjournals.org

© IJIP Journal
Published in Malaysia

Typesetting: Camera-ready by author, data conversation by CSC Publishing
Services – CSC Journals, Malaysia

CSC Publishers

Table of Contents

Volume 1, Issue 2, July/August 2007.

Pages
1 - 10 Multi Local Feature Selection Using Genetic Algorithm For Face
Identification
Dzulkifli Mohamad

11 - 20
Quality - Security Uncompromised and Plausible Watermarking for
Patent Infringement
Yamuna Govindarajan, Sivakumar Dakshinamurthi

21 - 28 Reconstructing Vehicle License Plate Image from Low Resolution
Images using Nonuniform Interpolation Method
Shih-Chieh Lin, Chih-Ting Chen

.

International Journal of Image Processing (IJIP) Volume (1) : Issue (2)

Dzulkifli Mohammad

MULTI LOCAL FEATURE SELECTION USING GENETIC ALGORITHM FOR
FACE IDENTIFICATION

Dzulkifli Mohamad dzulkifli@utm.my
Falkuti Sains Komputer dan Sistem Maklumat
Universiti Teknologi Malaysia
Skudai,81310, Johor, Malaysia
________________________________________________________________________

Abstract

Face recognition is a biometric authentication method that has become more
significant and relevant in recent years. It is becoming a more mature technology
that has been employed in many large scale systems such as Visa Information
System, surveillance access control and multimedia search engine. Generally,
there are three categories of approaches for recognition, namely global facial
feature, local facial feature and hybrid feature. Although the global facial-based
feature approach is the most researched area, this approach is still plagued with
many difficulties and drawbacks due to factors such as face orientation,
illumination, and the presence of foreign objects. This paper presents an
improved offline face recognition algorithm based on a multi-local feature
selection approach for grayscale images. The approach taken in this work
consists of five stages, namely face detection, facial feature (eyes, nose and
mouth) extraction, moment generation, facial feature classification and face
identification. Subsequently, these stages were applied to 3065 images from
three distinct facial databases, namely ORL, Yale and AR. The experimental
results obtained have shown that recognition rates of more than 89% have been
achieved as compared to other global-based features and local facial-based
feature approaches. The results also revealed that the technique is robust and
invariant to translation, orientation, and scaling.
Keywords: Face Recognition, Facial Feature Extraction, Localization, Neural Network, Genetic Algorithm
(GA)

______________________________________________________________________________________

1. INTRODUCTION
Face recognition is one of the physiological biometric technologies which exploit the unique
features on the human face. Although face recognition may seem an easy task for human, but
machine recognition is a much more daunting task [1]. The difficulties due to pose, present or
absent of structural components, occlusion, image orientation, facial expression and imaging
conditions [2]. For the last two decades, there has been growing interest in machine recognition of
faces due to its potential applications, such as film processing, user authentication, access control
system, law enforcement, etc. Typically face recognition system should include four stages. The
first stage involves detecting human face area from images, i.e. detect and locate face. The
second stage requires extraction of a suitable representation of the face region. The third stage
classifies the facial image based on the representation obtained in the previous stage. Finally,
compares facial image against database (gallery) and reports a match.

International Journal of Image Processing, Volume (1) : Issue (2) 1

Dzulkifli Mohammad

To design a high accuracy recognition system, the choice of feature extractor is very crucial. In
general, feature extraction methods can be divided into two categories: face based and
constituent based. The face based approach uses raw pixel information or features extracted from
the whole image which as a representation of face. Therefore face based method uses global
information instead of local information. Principal Component Analysis (PCA) is a typical and
successful face based method. Turk and Pentland developed a face recognition system using
PCA in 1991 [3]. In 1997, Belhumeur et. al. proposed Fisherface technique based on Linear
Discriminant Analysis (LDA) to overcome the difficulty cause by illumination variation [4].
Haddadnia et. al. introduced a new method for face recognition using Pseudo Zernike Moment
Invariants (PZMI) as features and Radial Basis Function (RBF) neural network as the classifier [5],
[6], [7]. Since the global information of an image are used to determine the feature elements,
information that are irrelevant to facial region such as shoulders, hair and background may
contribute to creation of erroneous feature vectors that can affect the face recognition results.
Furthermore, due to the variation of facial expression, orientation and illumination direction, single
feature is usually not enough to represent human face. So the performance of this approach is
quite limited.

The second one is the constituent based approaches are based on relationship between
extracting structural facial features, such as eyes, mouth, nose, etc. The constituent approaches
deal with local information instead of global information. Therefore constituent based method can
provides flexibility in dealing facial features, such as eyes and mouth and not affected by irrelevant
information in an image. Yuille et. al. use Deformable Templates to extract facial features [8].
These are flexible templates constructed with a priori knowledge of the shape and size of the
different features [9]. The templates can change their size and shape so that they can match
properly. These methods work well in detection of the eyes and mouth, despite variations in tilt,
scale and rotation of head. However modeling of the nose and eyebrow was always a difficult task
[8], [9]. Additionally it cannot deals with complicated background settings. Moreover the
computation of template matching is very time consuming. In 1999, Lin et. al. presented an
automatic facial feature extraction using Genetic Algorithm (GA) [10]. In 2002, Yen et. al.
proposed a novel method using GA to detect human facial features from images with a complex
background without imposing any constraints [11]. The normal process of searching for the
features is computationally expensive; therefore GA is used as a search algorithm [11]. Genetic
algorithm possesses the following feature that make them better suited that traditional search
algorithm [12]. Comparing to face based approach, constituent based approach provide flexibility
in dealing facial features, such as eyes and mouth and not affected by irrelevant information in an
image; therefore constituent based approach is selected as a solution in this paper.

In the literature [13] and [14], the combination of an ensemble of classifiers has been proposed to
achieve image classification systems with higher performance in comparison with the best
performance achievable employing a single classifier. In Multiple Classifier System [15], different
structures for combining classifier systems can be grouped in three configurations. In the first
group, the classifier systems are connected in cascade to create pipeline structure. In the second
group, the classifier systems are used in parallel and their outputs are combined named it parallel
structure. Lastly the hybrid structure is a combination of the pipeline and parallel structure.

So, this paper proposes a human face recognition system that can be designed based on hybrid
structural classifier system. The intended scheme actually is designed to have evolutionary
recognition results by gathering available information and extracting facial features from input
images. In this paper, Pseudo Zernike Moment Invariant (PZMI) has been used as a feature
domain to extract features from facial parts. Radial Basis Function (RBF) neural network is used
as the classifier in the proposed method. RBF neural network is chosen due to their simple
topological structure, their locally tuned neurons and their ability to have a fast learning algorithm
in comparison with the multi-layer feed forward neural network [16], [19].

The organization of the paper is structured as follow. Face parts localization using GA, moment
generation using PZMI, facial feature classification using RBF, multi local feature selection,
experimental results and conclusion.


Dzulkifli Mohammad

2. PROPOSED METHOD
2.1 Facial Parts Localization using GA
This is a face segmentation and facial feature extraction process [11], which gathers the sub-
regions of right eye, left eye, mouth and nose using GA. All the images captured were head and
shoulder images and in a frontal view.

2.1.1 Genetic Algorithm
GA is a powerful search and optimization algorithm, which are based on the theory of natural
evolution. In GA, each solution for the problem is called a chromosome and consists of a linear list
of codes. The GA sets up a group of imaginary lives having a string of codes for a chromosome
on the computer. The GA evolves the group of imaginary lives (referred to as population), and
gets and almost optimum solution for the problem. The GA uses three basic operators to evolve
the population: selection, crossover, and mutation.

2.1.2 Face Segmentation
The face segmentation process is proceeded under the assumption that human face region can
be approximated by an ellipsoid [17]. Therefore each chromosome in the population during the
evolutionary search has five parameters genes, the centre of the ellipse (x and y), x directional
radius (rx), y directional radius (ry) and the angle ( ). Figure 1 shows the chromosome for face
segmentation.

x-8bits y-8bits rx-8bits ry-8bits -7bits

FIGURE 1: Chromosome for Face Segmentation

The fitness of the chromosome is defined by the number of edge pixels in the approximated
ellipse like face to the actual number of pixels in the actual ellipse. The ratio is large when both
ellipses overlap perfectly.

2.1.3 Facial Feature Extraction
After the process of face segmentation, segmented image is fed into facial feature extraction
process. The facial feature extraction is based on horizontal edge density distribution [11]. The
horizontal edge map of the image from segmented image is obtained in order to extract facial
features. In this method, rectangle templates of different sizes for different facial features are used.
The sizes of the templates for different features are decided according to general knowledge of
the size of the features. Here, both the eye and eyebrow are contained in the same rectangle
template.

In order to make the search process less computational expensive, face is divided into sub-
regions as shown in Figure 2. The right eye is in the region Er, left eye in the region El, and region
M contains the mouth. The nose region N can be obtained once the eyes and mouth are located.

Er El
.
N

M

FIGURE 2: Sub-regions of the face

GA is used in the process of facial feature extraction to search for the global maximum point when
the template best matches the feature. The chromosome for face feature extraction shown in
Figure 3.


Dzulkifli Mohammad

x-direction y-direction
(7 bits) (7 bits)
FIGURE 3: Chromosome for face feature extraction

The chromosome represents the position of the feature in the x and y direction. The fitness is
evaluated in terms of the density of the template. The best template is selected when the fitness is
maximized. The fitness, F is shown below,
1 m n
F= ∑∑ T ( x, y)
m × n x =1 y =1
(2.1)
 T ( x, y ) = 1 if the pixel is white
where  ,
T ( x, y ) = 0 if the pixel is black
and T is the template, (x,y) are the coordinates of the template, and m × n is the size of the
template.

2.2 Moment Generation using PZMI
PZMI is an orthogonal moment that is shift, rotation and scale invariant and very robust in the
presence of noise. PZMI is been used for generating feature vector elements. Pseudo Zernike
polynomials are well known and widely used in the analysis of optical systems. Pseudo Zernike
polynomials are orthogonal set of complex-valued polynomials, Vnm defined as [5], [6], [7], [16]:
y
V nm (x, y) = R nm (x, y) exp (jm tan -1 ( ))
x
(2.2)
2 2
where x + y ≤ 1, n ≥ 0, m ≤ n and Radial polynomial R nm are defined as:
n− m n −s
R nm (x, y) = ∑D
s=0
n, m , s
(x 2 + y 2 ) 2

(2.3)
where:
(2n + 1 − s)!
D n, m , s
= (−1) s
s! (n − m − s)! (n − m − s + 1)!
(2.4)

The PZMI can be computed by the scale invariant central moments CM p,q and the radial
geometric moments RM p,q as follows:
 k  m 
n− m k m
n +1
PZMI nm = ∑ D n, m , s ∑ ∑  a  b 
  
π (n − m − s)even, s = 0 a =0 b=0   
b
(− j) CM 2k + m − 2a − b,2a + b

n +1
n− m d m
 d  m 
+
π
∑ D n, m , s ∑∑  a  b 
  
(n − m − s)odd, s = 0 a =0 b =0   
b
(− j) RM 2d + m − 2a − b,2a + b
(2.5)
where k =(n-s-m)/2, d=(n-s-m)/2, CM p,q is the central moments and RM p,q is the Radial moments
are as follow:


Dzulkifli Mohammad

µ pq
CM p, q = (p + q + 2)/2
M 00
(2.6)
)2 ) 1/2 ) )
∑ ∑ f(x, y) (x
x y
+ y2 ) x p yq
RM p, q = (p + q + 2)/2
M 00
(2.7)
) )
where x = x − x 0, y = y - y 0 and M pq, µ pq and x 0, y 0 are defined as follow:

M pq = ∑∑ x y
f(x, y) x p y q

(2.8)
µ pq = ∑∑ x y
f(x, y) (x − x 0 ) p (y − y 0 ) q

(2.9)
x 0= M 10 /M 00
(2.10)
y 0= M 01 /M 00
(2.11)

2.3 Facial Feature Classification Using RBF
RBF neural network has been found to be very attractive for many engineering problem because
[18], [19]:
(i)
They are universal approximators, (ii)They have a very compact topology and (iii) Their
learning speed is very fast because of their locally tuned neurons.

Therefore the RBF neural network serve as an excellent candidate for pattern applications and
attempts have been carried out to make the learning process in this type of classification faster
then normally required for the multi-layer feed forward neural networks [19]. In this paper, RBF
neural network is used as classifier in face recognition system.

2.3.1 RBF Neural Network Structure
Figure 4 shows the basic structure of RBF neural networks.

1 1 1

2 2 2

n r s

Input layer RBF layer Output layer

FIGURE 4: RBF Neural Network Structure

The input layer of the neural network is a set of n unit, which accept the elements of an n-
dimensional input feature vector. The input units are fully connected to the hidden layer r hidden
units. Connections between the input and hidden layers have unit weights and, as a result, do not
have to be trained. The goal of the hidden layer is to cluster the data and reduce its dimensionality.
In this structure the hidden units are referred to as the RBF units. The RBF units are also fully
connected to the output layer. The output layer supplies the response of the neural network to


Dzulkifli Mohammad

activation pattern applied to the input layer. The transformation from the input space to the RBF-
unit space is nonlinear (nonlinear activation function), whereas the transformation from the RBF-
unit space to the output space is linear (linear activation function). The RBF neural network is a
class of neural network where the activation function of the hidden units is determined by the
distance between the input vector and a prototype vector. The activation function of the RBF units
is expressed as follow [7], [18], [20]:
 x − ci 
Ri ( x) = Ri 
 σ
 ,
 i = 1,2,..., r
 i 
(2.12)
where x is an n-dimensional input feature vector, ci is an n-dimensional vector called the centre of
the RBF unit, σi is the width of the RBF unit, and r is the number of the RBF units. Typically the
activation function of the RBF units is chosen as a Gaussian function with mean vector ci and
variance vector σi as follow:
 x − ci 2

Ri ( x) = exp − 
 σ i2 
 
(2.13)
Note that σi2 represents the diagonal entries of the covariance matrix of the Gaussian function.
The output units are linear and the response of the jth output unit for input x is:
r
y j ( x) = b( j ) + ∑ Ri ( x) w2 (i, j )
i =1
(2.14)
where w2(i, j) is the connection weight of the ith RBF unit to the jth output node, and b(j) is the bias
of the jth output. The bias is omitted in this network in order to reduce the neural network
complexity [5], [19], [20]. Therefore:
r
y j ( x) = ∑ Ri ( x) × w2 (i, j )
i =1
(2.15)
2.4 Multi Local Feature Selection
The layout of multi local feature selection has been shown in Figure 5. In the first step, facial parts
localization process is done, so the exact location of the facial parts regions is localized. Secondly,
sub-image of each facial parts will be created, which contain only relevant information of facial
parts, such as eyes, nose, mouth, etc. Next in third stage, each of the facial parts is extracted in
parallel from the derived sub-image. The fourth stage is the process of classification, which
classify the facial features. Finally the last stage combines the outputs of each neural network
classifier to construct the recognition.

3. EXPERIMENTAL RESULTS
To validate the effectiveness of the algorithm, a simple experiment was carried out. The human
face images were taken using a monochrome CCD camera with a resolution of 768 by 576 pixels.
There are also some images from international face database is been used, such as face image
from ORL, Yale and AR Database. The total number of 3065 images have been selected from all
the database as a test and train images. The GA parameters setting used for both face
segmentation and facial feature extraction in the simulation process are shown in Table 1.

Face Feature
segmentation extraction
Population 100 50
Crossover 0.8 0.8
Mutation 0.001 0.001
TABLE 1: GA Parameters


Dzulkifli Mohammad

Figure 7 displays the head and shoulder original image before the process of facial parts
localization. Figure 8 shows the result after the process of facial parts localization.

FIGURE 7: Head and shoulder original image

FIGURE 8: Result of Facial Parts Localization

Table 2 shows some of the features extracted by PZMI. Though it may argued that there exists a
similar value (or closed to) among different facial features but it never happens for the entire
complete set. To investigate the effect of the method of learning on the RBF neural network, three
categories of feature vector based on the order (n) of the PZMI have been set (Table 3). The
neural network classifier was trained in each category based on the training images.

Eye Feature RBF Neural
Extractor 1 Network
(PZMI) Classifier 1

Mouth Feature RBF Neural
Extractor 2 Network
(PZMI) Classifier 2
Facial Parts Facial Parts Decision
Localization Sub-image Strategy
creation Nose Feature RBF Neural
Extractor 3 Network
(PZMI) Classifier 3

Feature RBF Neural
Extractor N Network
(PZMI) Classifier N

FIGURE 5 : The layout of multi local feature selection


Dzulkifli Mohammad

Person A Person B
Left eye Right eye Left eye Mouth
PZMI 9,1 0.025769 0.030027 0.027727 0.010727
PZMI 9,2 0.017139 0.011290 0.012600 0.000254
PZMI 9,3 0.021621 0.017175 0.024139 0.008444

PZMI 9,4 0.002486 0.003062 0.006773 0.027121
PZMI 9,5 0.036770 0.035310 0.030024 0.020046
PZMI 9,6 0.090679 0.092341 0.091703 0.003933

PZMI 9,7 0.062495 0.070282 0.075366 0.011679
PZMI 9,8 0.082933 0.080637 0.083488 0.058776

PZMI 9,9 0.020375 0.014172 0.022936 0.016866

TABLE 2: Features extracted by PZMI

Category No. PZMI feature elements
n=1, m=0,1
n=2, m=0,1,2
n=3, m=0,1,2,3
1 n=4, m=0,1,2,3,4
n=5, m=0,1,2,3,4,5
n=6, m=0,1,2,3,4,5,6
n=6, m=0,1,2,3,4,5,6
2 n=7, m=0,1,2,3,4,5,6,7
n=8, m=0,1,2,3,4,5,6,7,8
3 n=9, m=0,1,2,3,4,5,6,7,8,9
n=10, m=0,1,2,3,4,5,6,7,8,9,10

TABLE 3: Feature Vectors Elements based on PZM

The experimental results and the comparison between the previous research works using the
same dataset from three distinct facial databases are shown in Table 4. It shows that the overall
recognition rate of more than 89% has been achieved by the proposed method. The results also
reveal that the proposed technique is robust and invariant to translation, orientation, and scaling.

Database Eigen Fisher EGM SVM NN Proposed
method
ORL 80.3% 93.8% 81.5% 95.5% 91.5% 96.5%
Yale 66.7% 77.6% 82.4% 78.2% 74.5% 83.0%
AR 28.7% 89.2% 58.7% 59.7% 76.4% 89.2%

Average 58.6% 86.9% 74.2% 77.8% 80.8% 89.6%

TABLE 4: Recognition rate of experiment


Dzulkifli Mohammad

4. CONCLUSION
This paper presented a method for the recognition of human faces in 2-Dimentional digital images
using a localization of facial parts information. The combination of an ensemble of classifiers has
been used to achieve image classification systems with higher performance in comparison with
the best performance achievable employing a single classifier.

5. REFERENCES
1. Chellappa, R.; Wilson, C. L.; and Sirohey, S. “Human and Machine Recognition of Faces: A
Survey”. Proceedings of the IEEE, 83(5): 705-740, 1995
2. Yang, M. H.; Kriegman, D. J.; and Ahuja, N. “Detecting Faces in Images: A Survey”. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 24(1): 34-58, 2002
3. Turk, M.; and Pentland, A. “Face Recognition using Eigenfaces”. In Proceedings of the IEEE
Computer Society Conference on Computer Vision and Pattern Recognition, 586-591. Maui,
USA, 1991
4. Belhumeur, P.N.; Hespanha, J.P.; and Kriegman, D. J “Eigenfaces vs. Fisherfaces:
Recognition using Class Specific Linear Projection”. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 19(7): 711-720, 1997
5. Haddadnia, J.; Ahmadi, M.; and Faez,. “A Hybrid Learning RBF Neural Network for Human
Face Recognition with Pseudo Zernike Moment Invariant”. In Proceedings of the IEEE 2002
International Joint Conference on Neural Networks, 1: 11-16. Honolulu, USA, 2002
6. Haddadnia, J.; Ahmadi, M.; and Faez. “An Efficient Method for Recognition of Human Faces
using Higher Orders Pseudo Zernike Moment Invariant”. In Proceedings of the Fifth IEEE
International Conference on Automatic Face and Gesture Recognition, 315-320. Washington,
USA. 2002
7. Haddadnia, J.; Faez, K.; and Ahmadi, M. “An Efficient Human Face Recognition System
Using Pseudo Zernike Moment Invariant and Radial Basis Function Neural Network”.
International Journal of Pattern Recognition and Artificial Intelligence 17(1): 41-62, 2003
8. Yullie, A. L.; Cohen, D. S.; and Hallinan, P. W. “Feature Extraction from Faces using
Deformable Templates”. In Proceeding of the IEEE International Conference on Pattern
Recognition, 104-109. San Diego, USA, 1989
9. Huang, L. C.; and Chen, C. W. “Human Face Feature Extraction for Face Interpretation and
Recognition”. In Proceedings of the IEEE International Conference on Pattern Recognition,
204-207. Hague Netherlands, 1996
10. Lin, C. H.; and Wu, J. L. “Automatic Facial Feature Extraction by Genetic Algorithms”. IEEE
Transactions on Image Processing, 8(6): 834-845, 1999
11. Yen, G. G.; and Nithianandan, N. “Facial Feature Extraction Using Genetic Algorithm”. In
Proceedings of the IEEE 2002 Congress on Evolutionary Computation, 2: 1895-1900.
Honolulu, USA, 2002.
12. Tang, K. S.; Man, K. F.; Kwong, S.; and He, Q. “Genetic Algorithms and Their Applications”.
IEEE Signal Processing Magazine 13(6): 22-37, 1996
13. Giacinto, G.; Roli, F.; and Fumera. “Unsupervised Learning of Neural Network Ensembles for
Image Classification”. In Proceedings of the IEEE-INNS-ENNS International Joint Conference
on Neural Networks, 3: 155-159. Como Italy, 2000
14. Kittler, J.; Hatef, M.; Duin, R. P. W.; and Matas, J. “On Combining Classifier”. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 20: 226-239, 1998
15. Ho, T. K.; Hull, J. J.; and Srihari, S. N. “Decision Combination in Multiple Classifier System”.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(1): 66-75. 1994
16. Haddadnia, J.; Faez, K.; and Moallem, P. “Neural Network Based Face Recognition with
Moment Invariants”, In Proceedings of the IEEE International Conference on Image
Processing, 1: 1018-1021. Thessaloniki Greece, 2001
17. Yokoo, Y.; and Hagiwara, M. “Human Face Detection Method using Genetic Algorithm”. In
Proceedings of the IEEE Congress on Evolutionary Computation, 113-118. Nagoya Japan,
1996


Dzulkifli Mohammad

18. Yingwei, L.; Sundarajan, N.; and Saratchandran, P. “Performance Evaluation of A Sequential
Minimal Radial Basis Function (RBF) Neural Network Learning Algorithm”. IEEE Transactions
on Neural Network, 9(2): 308-318, 1998
19. Zhou, W. “Verification of the Nonparametric Characteristics of Backpropagation Neural
Networks for Image Classification”. IEEE Transactions Geoscience and Remote Sensing,
37(2): 771-779. 1999.
20. Jang, J. -S. R. ANFIS: “Adaptive-Network-Based Fuzzy Inference System”. IEEE Transactions
Systems, Man and Cybernetics, 23(3): 665-684, 1993



Quality - Security uncompromised and Plausible Watermarking
for Patent Infringement

Yamuna Govindarajan yamunaphd@yahoo.com
Reader /Department of Electrical and Electronics Engg
Annamalai University
Annamalai Nagar -608002,India

Sivakumar Dakshinamurthi
Professor / Department of Electronics and Instrumentation Engg dsk2k5@gmail.com
Annamalai University
Annamalai Nagar -608002,India

Abstract

The most quoted applications for digital watermarking is in the context of
copyright-protection of digital (multi-)media. In this paper we offer a new digital
watermarking technique, which pledges both Security and Quality for the image
for the Patent protection. This methodology uses tale techniques like Shuffling,
Composition & Decomposition, and Encryption & Decryption to record the
information of a protected primary image and the allied watermarks. The
quadtree can aid the processing of watermark and AES provides added security
to information. Besides that, we intend a novel architecture for Patent Protection
that holds promise for a better compromise between practicality and security for
emerging digital rights management application. Security solutions must seize a
suspicious version of the application-dependent restrictions and competing
objectives.

Keywords: Digital watermarking, Patent protection, Shuffling, Quadtree, Advanced Encryption Standard.

1. INTRODUCTION
Digital watermarking is a technique which allows an individual to add hidden Patent notices or
other verification messages to digital audio, video, or image signals and documents. Such a
message is a group of bits describing information pertaining to the signal or to the author of the
signal (name, place, etc.). The technique takes its name from watermarking of the paper or
money as a security measure [1]. According to the human perception, the digital watermarks can
be divided into two different types as follows: visible and invisible. Visible watermarks change the
signal altogether such that the watermarked signal is totally different from the actual signal, e.g.,
adding an image as a watermark to another image. Invisible watermarks do not change the signal
to a perceptually great extent, i.e., there are only minor variations in the output signal. An
example of an invisible watermark is when some bits are added to an image modifying only its
least significant bits (LSB).

Patent protection for multimedia information has been a key concern of multimedia industry. The
electronic representation and transfer of digitized multimedia information (text, video, and audio)



have increased the potential for misuse and theft of such information, and significantly increases
the problems associated with enforcing Patents on multimedia information. Digital watermarking
technology opens a new door to authors, producers, publishers, and service providers for
protecting their rights and interests in multimedia documents [2].

In order to protect the patent of a digital image, a matured digital image watermarking technique
must have to meet the following properties [3]:

– Perceptual transparency: The algorithm must embed data without affecting the
perceptual quality of the underlying host signal.

– Security: A secure data embedding procedure cannot be broken unless the unauthorized
user access to a secret key that controls the insertion of the data in the host signal.

– Robustness: The digital watermark must survive after being attacked by lossy data
compression and image manipulation and processing operations, e.g. cut and paste,
filtering, etc.

– Unambiguous: Retrieval of the watermark should unambiguously identify the owner.

– Universal: The same watermarking algorithm should be applicable to all multimedia under
consideration.

– Imperceptibility: The watermark itself should not be visible by the human visual system
(HVS) and should not degrade the image quality.

– Reliability: To ensure that the project application returns the correct watermark each time.
In spite of the loss of watermarking information by the optimizer, we should always be
able to obtain correct and accurate results from the project.

Today two technologies are applied when protecting image data in Digital Rights Management
(DRM) environments: Encryption and Digital watermarking. Encryption renders the data
unreadable for those not in the possession of a key enabling decryption. This is especially of
interest for access control, as usage of the image data is restricted to those owning a key.

Digital watermarking adds additional information into an image file without influencing quality of
file size. This additional information can be used for inserting Patent information or a customer
identity into the image file. The latter method is of special interest for DRM as it is the only
protection mechanism enabling tracing illegal usage to a certain customer even after the image
data has escaped the secure DRM environment.

In general these two mechanisms show a certain antagonism with respect to the transparency
requirements of the encrypted, respectively watermarked data. Both mechanisms apply small
media type specific changes on the cover data. But whereas transparent watermark embedding
should keep the auditory quality of the marked data unaffected, partial encryption is targeted on
maximum effect on the quality of the digital media

1.1 Related Work

Rahul Shukla, Pier Luigi Dragotti, Minh Do_, and Martin Vetterli [5] have proposed a novel coding
algorithm based on the tree structured segmentation, which achieves the oracle like exponentially
decaying rate-distortion (R-D) behavior for a simple class of signals, namely piecewise
polynomials in the high bit rate regime. Raphael Finkel and J.L.Bentley have proposed [6] an



optimized tree and an algorithm to accomplish optimization in n log n time. They discuss the
specific case of two-dimensional retrieval, although the structure is easily generalized to arbitrary
dimensions. P. Strobach proposes [7] the concept of recursive plane decomposition (RPD) is
embedded in a quadtree data structure to obtain a new variable block size image coding
algorithm that offers a high performance at a low computational cost.

H. K. C. Chang, P. M. Chen, and L. L. Cheng have proposed [10] an efficient data structure,
called the common-component binary tree (CCBT), to hold the linear quadtrees corresponding to
a set of similar binary images. T. W. Lin has come out with [11] a new approach for storing a
sequence of similar binary images Based on linear quadtree structures and overlapping concepts.
F. Boland, J. O. Ruanaidh, and C. Dautzenberg have proposed[12] an overview of watermarking
techniques and a solution to one of the key problems in image watermarking, namely how to hide
robust invisible labels inside grey scale or colour digital images. Ji-Hong Chang and Long-Wen
Chang have proposed [13] a new digital watermarking algorithm for images with bar code and
digital signature and their simulation shows that the proposed algorithm gets satisfactory results
for various attacks. Hans Georg Schaathun have proposed[14] to prevent un authorised copying
of copyrighted material, by tracing at least one guilty user when illegal copies appear and they
describe how error-correction makes it possible to relax this assumption, and they modify two
existing schemes to enable error-correction. Soroosh Rezazadeh, and Mehran Yazdi [15] have
discussed a robust digital image watermarking scheme for copyright protection applications using
the singular value decomposition (SVD).

JUAN R. HERNA´ NDEZ, FERNANDO PE´REZ-GONZA´ LEZ have found [16] a statistical
approach to obtain models that can serve as a basis for the application of the decision theory to
the design of efficient detector structures. M. Barnia, F. Bartolinib, V. Cappellinib, E. Maglic, G.
Olmo have conversed [17] near-lossless digital watermarking for copyright protection of remote
sensing images and they show that, by forcing a maximum absolute difference between the
original and watermarked scene, the near-lossless paradigm makes it possible to decrease the
effect of watermarking on remote sensing applications to be carried out on the images. Md.
Mahfuzur Rahman and Koichi Harada have proposed [18] a parity enhanced topology based spot
area watermarking method to embed information in objects with layered 3D triangular meshes
such as those reconstructed from CT or MRI data.

Nizar Sakr, Nicolas Georganas, and Jiying Zhao have proposed [19] an adaptive watermarking
algorithm which exploits a biorthogonal wavelets-based human visual system (HVS) and a Fuzzy
Inference System (FIS) to protect the copyright of images in learning object repositories. Jacob
Löfvenberg, Niclas Wiberg have mentioned [20] the performance of random fingerprinting in
conjunction with a specific testing method.

1.2 Objective of the Work

In our paper, We propose a new algorithm which for Patent Protection which will neither affect the
quality of the Patent Image, nor compromise the security also.

This method is to shuffle the digital watermark into a primary image, then projected on to the
spatial domain, combined with the information of digital watermark and primary image, and
creates a Patent secret code vector (PSCV). This PSCV can be authorized and becomes a
critical message for the future image patent argument. When certification authority generates the
256 bit secret key, Encrypted PSCV and argued image will extract watermark from the image for
the purpose of image patent loyalty judgment. In our method, the primary image will not be
modified nor created watermarked image. The proposed technique is suitable for non-modifiable
image; for instance, medical image. This paper is composed of following sections: Section 2
details the proposed methodology, Section 2.1 briefly introduces Shuffling and Projection
Technique, Section 2.2 briefs quadtree structure, Section 2.3 briefs Encryption technique. Section
3 demonstrates the experimental results and the discussions. Finally, the conclusions are
presented in Section 4.



2. THE PROPOSED METHODOLOGY
Patent protection has become a hot issue in current digitized world due to the prevailing usage of
Internet and an accumulation of multimedia data distribution, for instance, audio, image, and
video. Digital watermarking techniques can protect images intellectual property right efficiently.
We are proposing a new digital watermarking technique for the Patent protection. Figure: 1
represents the Block Diagram of the proposed methodology. The technique is discussed in 2.1,
also Shuffling technique, Quadtree, AES algorithm is discussed in 2.2, 2.3, and 2.4 respectively.

2.1 The Algorithm

2.1.1 Generation of Patent Secured Secret Code Vector (SSCV)
Input: Primary Image (Ip), Patent Image(Ic), Key seed (Ks)
Sp ← Size of (Ip)
Sc ← Size of (Ic)
1. First the Ic is shuffled to get Shuffled Image (Is) of size Sc
2. A key(Ku) is generated from a random number using Ks
3. Arrive X position (XIy) , Y Position (YIy) from Ku
4. Based on the position XIy, YIy a new image is yielded from the Ip , called sub primary
image/Yielded Image (Iy)
5. Is is projected on Iy to arrive projected image Ipr
6. Ipr is decomposed to get the Patent Secret Code Vector (PSCV)
7. Generate 256 bit key (Ak) from the random seed Ks
8. Encrypt the PSCV using the key Ak to get the SSCV

2.1.2. Generation of Patent Image / Watermark Image (Ic)
1. Input: SSCV, Ip, Ks
2. Generate 256 bit key (Ak) from the random seed Ks
3. SSCV is decrypted using (Ak) to get PSCV
4. PSCV is composed to get Ipr
5. A key(Ku) is generated from a random number using Ks
6. Arrive X position (XIy) , Y Position (YIy) from Ku
7. Based on the position XIy, YIy a new image is yielded from the primary image (Ip).
8. Iy is projected on Ipr to arrive Is
9. Is will be reshuffled to get Ic which is Patent image

2.1.3. Verification Process
1. The Secured Secret Code Vector of the user (SSCVu) is claimed to get an image Icu.
2. Compare the Claimed Image Icu with the Patent Image Ic.
3. If Icu and Ic are equal, then claimed user is the Owner of the Patent Image, besides the
claimed user is not the owner of the Patent image.



Patent Image (Ic) Primary Image (Ip)

Generation of Generation of
Secret Key (Ku) Secret Key (Ku)

Shuffled Patent Image (Is) Sub Primary Image
(Iy)

Projected Image (Ipr)

Patent Secret Code Vector
(PSCV)

256 bit Key (Ak)

Secured Secret Code Vector
(SSCV)

FIGURE: 1 Block diagram of the proposed methodology

2.2 Shuffling and Projection Technique
Shuffling is a linear-time algorithm (as opposed to the previous O(n log n) algorithm if using
efficient sorting such as merge sort or heap sort) which involves moving through the pack from
top to bottom, swapping each in turn with another from a random position in the part of the pack
that has not yet been passed through (including itself). Providing that the random numbers are
unbiased, this will always generate a random permutation [4].

2.2.1. Shuffling Technique
We have proposed a Shuffling methodology here. Icv is the Vector representation of copyright
image .The locations for the shuffled image(SL) is generated as follows:

SL={ xi : p(xi) , xi ∉ {xk}, i=0,..N-1,k=0,. i-1} (1)

Where N is size of Copyright image

p(xi) = mod (PRNG(rs) , N) , i=0 , … , N-1 (2)

SIvi = { yi : p(yi) , i=0,..N-1) (3)

p(yi) = Icv[SL[i]] i=0,…N-1 (4)

The shuffled image is then reconstructed from the shuffled vector and is represented as Is, Where
PRNG is the pseudo random number generation function.

SIv is the Shuffled Image Vector. rs is the seed value for PRNG and is defined as follows

rsn = rn-1, n=1,…N-1 (5)



When n =0, rs0 = key seed value
SIv is converted to sqrt(N) x sqrt(N) matrix which is ls (Shuffled Patent Image).

2.2.2. Projection Technique

T T -1
During Composition Is is projected on (Iy ) to get Ipr
T T -1
Ipr = (Is * (Iy ) ) (6)
T
During Decomposition Iy is projected on Ipr to retrieve Is
T
Is = Iy * Ipr (7)

2.3. Quad Tree

A quadtree is a tree data structure in which each internal node has up to four children. Quadtrees
are most often used to partition a two dimensional space by recursively subdividing it into four
quadrants or regions. The regions may be square or rectangular, or may have arbitrary shapes. A
node of a point quadtree is similar to a node of a binary tree, with the major difference being that
it has four pointers (one for each quadrant) instead of two ("left" and "right") as in an ordinary
binary tree. Also a key is usually decomposed into two parts, referring to x and y coordinates.
Therefore a node contains following information:

• 4 Pointers: quad[‘NW’], quad[‘NE’], quad[‘SW’], and quad[‘SE’]
• point; which in turn contains: key; usually expressed as x, y coordinates

The leaf node in the quadtree represents a quadrant with identical pixels. The color of the leaf
node is the same as its corresponding quadrant. On the other hand, the internal node represents
a quadrant mixed by black pixels and white pixels. The quadrant corresponding to an internal
node is still needed to be recursively subdivided [5, 6, 7].

2.4. Advanced Encryption Standard

Encryption renders the data unreadable for those not in the possession of a key enabling
decryption. This is especially of interest for access control, as usage of the image data is
restricted to those owning a key. For this Encryption we use an AES algorithm, which is more
secured compare to DES. AES supports key sizes of 128 bits, 192 bits and 256 bits and will
serve as a replacement for the Data Encryption Standard which has a key size of 56 bits.

AES stands for Advanced Encryption Standard. AES is a symmetric key encryption technique
which will replace the commonly used Data Encryption Standard (DES). AES is secure enough to
protect classified information up to the TOP SECRET level, which is the highest security level and
defined as information which would cause "exceptionally grave damage" to national security if
disclosed to the public [8].

In addition to the increased security that comes with larger key sizes, AES can encrypt data much
faster than Triple-DES, a DES enhancement that which essentially encrypts a message or
document three times. According to NIST's "The AES algorithm is a symmetric block cipher that
can encrypt (encipher) and decrypt (decipher) information"[9]. Table 1 compares the advantages
of AES with DES



3. EXPERIMENTAL RESULTS AND DISCUSSION
In our experiments, for a given grey-valued primary image Ip, its image size is 512 x 512 pixels. In
that, there is a 64 x 64 grey Patent image Ic which is to be shuffled to get the shuffled image Is.
First, in our method, a secret key Ku is chosen. The Patent is shuffled by using Random number
seed (Ks) to select 64 x 64 permuted coordinates. Then, a sub primary image/yielded image Iy
with 64 x 64 pixels is selected from “Barbara” by applying Ks to get the top-left coordinate of Iy in
Ip. The next step is to generate a projected image Ipr for Iy and Is. The algorithm then recursively
divides each Ipr image into equal size quadrants, respectively. A secret code vector PSCV is built
by using a quadtree decomposer. Besides that we do Encryption by using AES algorithm which is
more secured in that a 256 bit key Ak is used to get the Secured secret code vector SSCV. The
watermark process is finished after the SSCV is built. Finally, the recovered watermark image is
obtained by reversal of the above process. Figure 2 are sample images of PSCV Construction and
Figure 3 are example of reconstruction.

Table 1: Comparing DES and AES
DES AES
Key Length 56 bits 128, 192, or 256 bits
Cipher Type Symmetric block cipher Symmetric block cipher
Block Size 64 bits 128, 192, or 256 bits
Developed 1977 2000
Cryptanalysis resistance Vulnerable to differential and Strong against differential,
linear cryptanalysis; weak truncated differential, linear,
substitution tables interpolation and Square attacks
Security Proven inadequate Considered secure
56 128 192 256
Possible Keys 2 2 ,2 , or 2
7 16 24 32
Possible ASCII printable 95 95 , 95 , or 95
character keys*
21
Time required to check all For a 56-bit key: 400 days For a 128-bit key: 5 x 10 years
possible keys at 50 billion
keys per second**
* When a text password input by a user is used for encryption (there are 95 printable characters in ASCII).

**In theory, the key may be found after checking 1/2 of the key space. The time shown is 100% of the key
space.

(A) (B) (C) (D) (E)



(A) (B) (C) (D) (E)

(A) (B) (C) (D) (E)

FIGURE: 2 Intermediate results of PSCV construction
A - Primary Image(512 * 512) , B - Patent Image(64 * 64) , C -Sub primary Image(64 * 64)
D - Shuffled Patent Image (IS) (64 * 64) , E - Projected Image(64 * 64).

(A) (B) (C) (D) (E)

(A) (B) (C) (D) (E)



(A) (B) (C) (D) (E)

FIGURE: 3 Intermediate results of Patent Image Reconstruction
A - Primary Image (512 * 512) , B - Projected Image(64 * 64)., C -Sub primary Image(64 * 64)
D - Shuffled Patent Image (IS) (64 * 64) , E - Patent Image(64 * 64)

4. CONCLUSION

In this paper, in part, we overviewed scores of issues that have been addressed for data
encryption and watermarking in a DRM given the thrust toward security for emerging resource
constrained DRM applications. We proposed a new solution that provides a better compromise
between security and quality of an image. In this current solution we proposed the five level
securities, which can defend the data from hack. This is resulting in a paradigm shift in the area of
information protection, in which ideas from areas such as media processing are often
incorporated to provide more lightweight solutions.

5. REFERENCES

1. “Digital watermarking” definition from http://en.wikipedia.org/wiki/ Digital_ watermarking

2. “A Digital Watermarking System for Multimedia Copyright Protection” - .Jian Zhao
Fraunhofer Center for Research in Computer Graphics, Eckhard Koch Fraunhofer
Institute for Computer Graphics, Wilhelminenstr. 7, D-64283 Darmstadt, Germany.

3. Power Point Presentation in “Watermarking of Digital Images “– by Assoc Prof. Dr. Aziza
A. Manaf & Akram M. Zeki , 1st ENGAGE European Union - Southeast Asia ICT ,
Research Collaboration Conference, March 29–31, 2006

4. “Shuffling Technique” from - http://en.wikipedia.org/wiki/Shuffling.

5. Rahul Shukla, Pier Luigi Dragotti, Minh Do_, and Martin Vetterli “Improved Quadtree
Algorithm Based on Joint Coding for piecewise Smooth Image Compression”

6. Raphael Finkel and J.L.Bentley (1974). "Quad Trees: A Data Structure for Retrieval on
Composite Keys". Acta Informatica 4 (1): 1-9.

7. P. Strobach “Quadtree structured recursive plane decomposition coding of images”, IEEE
Trans. Signal Proc., vol. 39, pp. 1380-1397, June 1991.

8. “AES Encryption Information” - http://www.bitzipper.com/aes-encryption.html.

9. Announcing the “ADVANCED ENCRYPTION STANDARD (AES)” - Federal Information,
Processing Standards Publication 197,November 26, 2001.



10. H. K. C. Chang, P. M. Chen, and L. L. Cheng, “Overlapping Representation of Similar
Images Using Linear Quadtrees Codes,“ In 8th IPPR Conference on Computer Vision,
Graphics and Image Processing, Tao-Yuang, Taiwan, R.O.C., Aug. 1995.

11. T. W. Lin, “Compressed Quadtree Representations for Storing Similar Images,” Image
and Vision Computing, Vol. 15, 1997, pp. 883-843.

12. F. Boland, J. O. Ruanaidh, and C. Dautzenberg, “Watermarking digital images for
copyright protection,” in Proc. IEEE Int. Conf. Image Proc. Applicat., 1995, pp. 321–326.

13. Ji-Hong Chang and Long-Wen Chang,” A New Image Copyright Protection Algorithm
Using Digital Signature of Trading Message and Bar Code watermark” palmerston North,
November 2003

14. Hans Georg Schaathun,” On watermarking/fingerprinting for copyright protection”.

15. Soroosh Rezazadeh, and Mehran Yazdi,” A Nonoblivious Image Watermarking System
Based on Singular Value Decomposition and Texture Segmentation” PROCEEDINGS
OF WORLD ACADEMY OF SCIENCE, ENGINEERING AND TECHNOLOGY VOLUME
13 MAY 2006 ISSN 1307-6884

16. JUAN R. HERNA´ NDEZ, FERNANDO PE´REZ-GONZA´ LEZ,” Statistical Analysis of
Watermarking Schemes for Copyright Protection of Images” PROCEEDINGS OF THE
IEEE, VOL. 87, NO. 7, JULY 1999.

17. M. Barnia, F. Bartolinib, V. Cappellinib, E. Maglic, G. Olmo,” Near-lossless digital
watermarking for copyright protection of remote sensing images”.

18. Md. Mahfuzur Rahman and Koichi Harada,” Parity enhanced topology based spot area
watermarking method for copyright protection of layered 3D triangular mesh data”
IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.2A,
February 2006.

19. Nizar Sakr, Nicolas Georganas, and Jiying Zhao,” Copyright Protection of Image
Learning Objects using Wavelet-based Watermarking and Fuzzy Logic” 3rd annual e-
learning conference on Intelligent Interactive Learning Object Repositories Montreal,
Quebec, Canada, 9–11 November, 2006

20. Jacob Löfvenberg, Niclas Wiberg,” Random Codes for Digital Fingerprinting”.



Reconstructing Vehicle License Plate Image from Low
Resolution Images using Nonuniform Interpolation Method

Shih-Chieh Lin sclin@pme.nthu.edu.tw
Department of Power Mechanical Engineering
National Tsing Hua University
Hsin-Chu, 300, Taiwan

Chih-Ting Chen
Department of Power Mechanical Engineering
National Tsing Hua University
Hsin-Chu, 300, Taiwan

Abstract

In this study, non-uniform interpolation method was adopted to reconstruct
license plate image from a series of low resolution vehicle license plate images.
Several image registration methods which were used to estimate the position and
orientation differences between these low resolution images are tested in this
study. It was found that the Fourier method is superior to other methods. The
non-uniform interpolation method is then used to reconstruct vehicle license plate
images from images with a character size as small as 3 × 6 pixels. Results show
that although the number or character is still not easy to read, the reconstructed
image shows a better readability than the original image.

Keywords: image enhancement, image registration, license plate recognition.

1. INTRODUCTION
Frequently, there is a need to identify vehicle license plate images taken from a camera or CCD
that is far away from the vehicle for security. The vehicle plate recognition systems developed by
other researchers [1-5] were generally designed for parking lot management or automatic toll
booth and might not be suitable for such an application.

In order to fulfill this need, techniques used to reconstruct high resolution image from low
resolution image can be adopted. Besides interpolation methods, super resolution techniques
are frequently used for this purpose.

By gathering more subpixels spatial information through multiframes of images the super
resolution method used a set of low resolution image to reconstruct high resolution images.
However, it is of important to know the relative position between these images before
reconstructing the image.

There are various registration methods had been proposed by other researchers [6]. In this
paper, several registration methods are adopted to estimate the relative position between images.
These methods are presented in the following section. The non-uniform interpolation method
used to reconstruct high resolution image is also briefly described. Effects of registration
methods and number of image frame on the reconstructed image are then studied. The best



registration method is then used to reconstruct several low resolution vehicle license plate
images. Finally, conclusions are made based on the test results.

2. REGISTRATION METHODS AND RECONSTRUCTING METHODOLOGY
The registration methods frequently used can be categorized as area based method and feature
based method [6]. In this study, one area based method and three feature based methods are
tested. The area based method tested is the Fourier method [7, 8], and the feature based
methods tested are Li [9], Lapalacian, and Lapalacian of Gaussian [10]. These methods are
briefly described in the following.

Area-based method
There are two area based approaches generally used for image comparison. One is the cross
correlation method and the other one is the Fourier method. The cross correlation method used
correlation coefficient as an estimate to determine the similarity of two images [11]. This method
may be adopted to estimate position error. But the subpixel image alignment problem we facing
here including both position and orientation error estimation and compensation. Since the cross
correlation method can not deal with images with rotation, this approach might not be suitable for
this purpose.

The Fourier method used the Fourier transform to identify the frequency dormant information of
examined images. By examining the orientation and phase difference in the Fourier transform of
these images, the position and orientation difference can be estimated [7, 8]. Therefore, the
Fourier method is tested in this study.

Feature-based methods
Barbara [12] divided the feature based method into four steps: (1) feature detection; (2) feature
matching; (3) transform model estimation; and (4) image resampling and transformation. Feature
detection is searching for region feature, line feature or point feature in the examined images.

In this study, pixel point with maximum gradient in local area or corner point are searched and
used as feature points. By comparing the position differences of these feature points, the location
and orientation differences between two images can be estimated.

Li [9] used Harris and Hessian operator to search for pixel point with maximum gradient in local
area. Lapalacian and Lapalacian of Gaussian operator can also be used to estimate the image
gradient and corner point [10]. In this study, these three feature–based methods are also tested.
After identifying the feature points, the related position change of feature points in images can be
estimated. There are many methods had been proposed for this estimation [11, 14-15]. Since
only translation and rotation is needed for consideration in this case, the method to calculate the
differences is relatively simple and straight forward. The least square error method is used to
calculate the position and orientation differences in this study.

Reconstruction methodology
There are many super resolution approaches had been proposed since Tsai and Huang [16].
These methods can be categorized as [17]: stochastic method, iterative back-projection method
and nonuniform interpolation method. Comparing with other methods the nonuniform
interpolation method is relatively easy to use, therefore this method is adopted here to reconstruct
license plate image.

As indicated in Fig. 1, the relative position between a series of images are determined first,
interpolation methods are then used to reconstruct high resolution image[18-20]. In this study the
bi-cubic interpolation method is used.



LR
Images

HR
Images

Interpolation onto
Registration
HR Grid
FIGURE 1: Schematic Diagram for Non-Uniform Interpolation Method.

3. TEST RESULTS AND DICUSSIONS
Effects of registration on image positioning
In order to study effects of registration methods on image positioning, a high resolution image
(3200 × 2400) as shown in Fig. 2 was used as the referential images. Low resolution (320 × 240)
images were made by averaging each 10 × 10 area of the standard image. A series of low
resolution images with position and orientation difference are made from images with a shift of 0
to 12 pixels to the referential image (corresponding to a shift of 0. to 1.2 pixels in low resolution
o
image) and 0.1 to 2.8 degree rotation in the standard image.

FIGURE 2: The Standard 3200x2400 Image Used In This Test.

Table 1 shows the summary of the comparison of position and orientation errors. Examining this
table, it is observed the area method is superior to the feature based methods tested in this study.
In the feature based methods, the use of Lapalacian mask is the best, and then of use of
Lapalacian-Gaussian. Both the mean and standard deviation in orientation error for the
Lapalacian (0.041, 0.064), and Lapalacian Gaussian (0.047, 0.070)are quite close to that of the
Fourier area method(0.047, 0.051). However, the mean and standard deviation in orientation
error for the Fourier method(0.021, 0.025) is much smaller than the Lapalacian(0.100, 0.131), and
Lapalacian Gaussian method(0.109, 0.140).

TABLE 1: Comparison of position and orientation error resulted from various registration methods.
Li Lapalacian Lap. Of Gau. Fourier
Position Maximum 0.54 0.36 0.36 0.1
errors (pixels)Minimum 0.1 0. 0. 0.
Average 0.129 0.100 0.109 0.021
SD* 0.159 0.131 0.140 0.025
Orientation Maximum 0.3 0.1 0.1 0.1
errors Minimum 0. 0. 0. 0.
(degree) Average 0.094 0.041 0.047 0.047
SD* 0.132 0.064 0.070 0.051
*SD: standard deviation



Effects of registration method and number of image frames on image reconstruction
In order to study effects of registration methods as well as the number of images used for
reconstruction on the reconstructed image, a series of low resolution images with position and
orientation difference are made from images with a shift of 1 to 10 pixels to the referential image
o
(corresponding to a shift of 0.1 to 1.0 pixels in low resolution image) and 0.1 to 1.0 degree
rotation in the standard image.

These images are then randomly picked as the base for image reconstruction. Registration
methods introduced in the previous section are used to estimate the position and orientation
differences between these images.

Fig. 3 shows interaction effects of image number and registration method on the correlation
coefficient of the referential image and the reconstructed image. The correlation coefficient
between the original high resolution image and the reconstructed image are used as the
performance index.
0.9

0.89
Correlation Coefficient

0.88

0.87

0.86 Fourier
Li
0.85 Lap
LoG
0.84
2 4 6 8 10 12 14 16 18
Number of Images

FIGURE 3: Interaction Effects of Image Number and Registration Method on the Correlation Coefficient of
the Referential Image and the Reconstructed Image.

As shown in the figure, it was found that for all registration methods tested, the performance
index increased as the frame number increased in the beginning (<7). However, no further
improvement in the performance index was observed when the frame number further increase.
In most cases the correlation coefficient might slightly decrease as the frame number further
increased. Theoretically, a better image can be reconstructed with more fames of image with the
cost of more computing time. However, it is found that errors are not avoidable when estimating
the position and orientation differences between images, using more frames of image might make
the reconstructed image more blur instead of more clear. It is also expected that the Fourier
method is superior to other methods no matter how many image frames were used for image
reconstruction since the Fourier method had performance in image positioning than other
methods.

Comparison of reconstructed vehicle license plate
From the previous section, the Fourier method is considered as the best approach for image
registration. This method is then adopted here to reconstruct vehicle license plate image. For
comparison, the results generated by using the Lapalacian method as the registration method
were also presented. 3 sets of vehicle images as shown in Fig 4 are taken using a Pentax Optio
SV digital camera. There are 16 frames taken or each sample for reconstruction.



Sample 1 Sample 2 Sample 3

Original
vehicle
Image

Size 320 x 240 320 x 240 320 x 240

License Plate

size 38 x 17 27 x 13 24 x 11
FIGURE 4: Dimensions of Test Samples.

Figures 5 to 7 shows the images reconstructed using 4, 7 and 16 frames, and two registration
methods. It is obvious that image reconstructed from 7 frames of image had better results than
that reconstructed from 4 frames of image. However, there is no evident that image
reconstructed from 16 frames of image had better results than that reconstructed from 7 frames
of image. These results are consistent with the study shown in the previous section.

FIGURE 5: Comparison Of Images Reconstruct From a 38 X 17 Vehicle License Plate Image (a) Original
Image (b) Images Reconstructed Using 4 Low Resolution Images (c) Images Reconstructed Using 7 Low
Resolution Images (d) Images Reconstructed Using 4 Low Resolution Images(Left: Using Fourier Method;
Right: Lap Method).

In Fig. 5, the size of license plate image is 38 × 17 pixels while the size of each number is around
5 × 10 Pixels. The plate number is readable even without reconstruction. However, the image is



much easy to read after reconstruction. And the images reconstructed by using the Fourier
method as the registration method seems a little clearer than those using the Lapalacian method.

3.5 × 7 Pixels. The plate number is hardly readable without reconstruction. Although, the
reconstructed images are still not quite readable, but it quite easy to identify the first three
characters are “M” “L” and “4”. Although we can not read 3 directly from the fourth character, 3 is
a quite obvious guess for the fourth character. For the last two characters, one might guess they
are 0, 5 or 9 instead of other numbers.

Right: Lap Method).

3 × 6 Pixels. The plate number is hardly readable even after reconstruction. However, there is a
good chance to guess the first four characters as “ML43”. For the last two characters, one might
guess they are 0, 5 or 9 instead of other numbers.

4. CONSLUSIONS
In this paper, non-uniform interpolation method was adopted to reconstruct low resolution license
plate image. This method used a series of low resolution images to reconstruct high resolution
image. It is of importance to know the position differences between these images before the
reconstructing process. Several image registration methods are tested here to estimate the
position and orientation differences between these images. It was shown that the Fourier method
is superior to other methods tested in this study. Vehicle license plate images with a character
size as small as 3 × 6 pixels were tested. Results show that although the character is still not
easy to read, the reconstructed image do improved the readability. The study shows the bottle



neck for this study is the methodology for registration. Without better registration methodology, a
better super resolution method is still useless. It was also found that without further improvement
in the registration technique, not much improvement can be achieved by increasing frame
number.

Right: Lap Method).

5. REFERENCES

1. C. A. Rahman, W. Badawy, A. Radmanesh, ”A real time vehicle's license plate recognition
system,” Proceedings. IEEE Conference on Advanced Video and Signal Based Surveillance,
Miami, Florida 2003.

2. S. Z. Wang and H. J. Lee, “Detection and Recognition of License Plate Characters with
Different Appearances,” Proc. IEEE 6th Intern. Conf. On Intelligent Transportation Systems,
Shanghai, China, 2003.

3. S. L. Chang, L. S. Chen, Y. C. Chung, S. W. Chen, “Automatic License Plate Recognition,”
IEEE Transactions on Intelligent Transportation Systems, 5(1):42-53, 2004.

4. H. F. Zhang, W. J. Jia, X. J. He, Q. Wu, "Learning-Based License Plate Detection Using
Global and Local Features," 18th International Conference on Pattern Recognition, 2006.

5. T. Naito, T. Tsukada, K. Yamada, K. Kozuka, S. Yamamoto, “Robust License-Plate
Recognition Method for Passing Vehicles Under Outside Environment,” IEEE Transactions
on Vehicular Technology, 49(6):2309 – 2319, 2000.



6. B. Zitov´a, J. Flusser, “Image Registration Methods: A Survey,” Image and Vision Computing,
21(11):977-1000, 2003.

7. E. D. Castro, C. Morandi, “Registration of translated and rotated images using finite Fourier
transform,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 9:700–703,
1987.

8. P. Vandewalle, S. SÄusstrunk, M. Vetterli, “A Frequency Domain Approach to Registration of
Aliased Images with Application to Super-Resolution," URASIP Journal on Applied Signal
Processing, Article ID 71459, 2006.

9. X. Li, J. Chen, “An Algorithm for Automatic Registration of Image,” International Conference
on Microwave and Millimeter Wave Technology(ICMMT 2004), Beijing, China, 2004.

10. L. Kitchen, A. Rosenfeld, “Gray-level corner detection,” Pattern Recognition Letters, 1: 95-
102, 1982.

11. R. Berthilsson, “Affine correlation,” Proceedings of the International Conference on Pattern
Recognition ICPR’98, Brisbane, Australia, 1998.

12. H.G. Barrow, J.M. Tenenbaum, R.C. Bolles, H.C. Wolf., “Parametric correspondence and
chamfer matching: Two new techniques for image matching,” Proceedings of the 5th
International Joint Conference on Artificial Intelligence, Cambridge, Massachusetts, 1977.

13. C. Harris, M. stephens, “A Combined Corner and Edge Detector,” Proceedings of the 4th
Alvey Vision Conference, Manchester, UK, 1988.

14. G. Borgefors, “Hierarchical chamfer matching: a parametric edge matching algorithm,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, 10:849–865, 1988.

15. W.H. Wang, Y.C. Chen, “Image registration by control points pairing using the invariant
properties of line segments,” Pattern Recognition Letters, 18:269–281, 1997.

16. R. Y. Tsai, T. S. Huang, “Multiframe image restoration and registration,” Advances in
Computer Vision and Image Processing (R. Y. Tsai, T. S. Huang, Eds.), 1: 317-339, JAI
Press, London, 1984.

17. S. C. Park, M. K. Park, M. G. Kang, “Super-resolution image reconstruction: a technical
overview,” IEEE Signal Processing Magazine, 20(3): 21-36, 2003.

18. N. R. Shah, A. Zakhor, “Resolution Enhancement of Color Video Sequences,” IEEE
Transactions on Image Processing, 8(6): 879-885, 1999.

19. M.S. Alam, J.G. Bognar, R.C. Hardie, and B.J. Yasuda, “Infrared image registration and high-
resolution reconstruction using multiple translationally shifted aliased video frames,” IEEE
Trans. Instrum. Meas., 49: 915-923, 2000.

20. N. Nguyen and P. Milanfar “An efficient wavelet-based algorithm for image superresolution,”
Proceedings of the IEEE International Conference on Image Processing, Vancouver,
Canada, 2000, pp.

Acknowledgements
The authors wish to thank National Science Council, Taiwan, R.O.C. for their financial support
(NSC95-2212-E-007-234).


International Journal of Image Processing (IJIP) Volume (1) Issue (2)

More Related Content

What's hot (17)

Viewers also liked (8)

Similar to International Journal of Image Processing (IJIP) Volume (1) Issue (2) (20)

International Journal of Image Processing (IJIP) Volume (1) Issue (2)