nor hamidah abdul ghafar 05 24

7/30/2019 Nor Hamidah Abdul Ghafar 05 24

http://slidepdf.com/reader/full/nor-hamidah-abdul-ghafar-05-24 1/24

<llo (^-OVOO 2>f ^ o ^

MARA UNIVERSITY OFT E CH NO L O G Y

LOGO RECOGNITION USING

ARTIFICIAL NEURAL NETWORK

(ANN)

NOR HAMIDAH BINTI ABD UL GHA FAR

2003283152

Thesis submitted in fulfillment of the requirements for

Bachelor of Science (Hons) Information Technology

Faculty of Information Technology And

Quantitative Science

APRIL 2005



S U P E R V I S O R ' S A P P R O V A L

NA ME : Assoc. Prof. Dr . M azan i M an a f

S I G N A T U R E

D A T E



DECLARATION

I certify that this thesis and the research to which it refers are the product of my own

work and that any ideas or quotation from the work of other people, published or

otherwise are fully acknowledged in accordance with the standard referring practices of

the discipline

APRIL 18 , 2005 NO R HAM IDAH BINTI ABD UL GHA FAR

2003283152



A C K N O W L E D G E M E N T

Firstly, thanks to God because give me a strength and good health that m ake m e possible

to complete this research.

My appreciation also dedicated for my parents, Abdul Ghafar Ahmad and Saniah

Dahlan, because gives me a moral and material support. Thank you for your

understanding to the research that I am doing.

I am also would like to forward my highest gratitude to my lecturer, Prof. Madya Dr.

Mazani bin Manaf and Pn. Zaidah binti Ibrahim for their paramount patience and

tolerance in guidance me in doing this research. Thanks a lot for all your information

and brilliant idea for my research.

Not to forgot to all my friend, thank you for your encouragement and support. Without

their belns, mv research will not be as perfect as this. Thank you.

11



A B S T R A C T

This project is about logo recognition using Artificial Neural Network (ANN). In order

to recognize the logo, a training phase using back propagation technique was

implemented. Based on study of existing research, many image and pattern recognition

has been done by using Artificial Neural Network and back propagation technique. Logo

was scanned or captured trough the Internet. The re are five logo an d eac h logo ha ve four

different size or fi-om different so urce. Logo mus t firstly don e proc ess of p re-p roc ess ing

by using MatLab 6.5 in order to normalize the logo to a specific size and for noise

removal. In addition, the edge detection for the logo also used MatLab 6.5 to get the

logo parameter and transform the logo into binary representation. The binary

representation was used for the input node of neural network for back propagation

training algorithm. To ensure a good performance of logo recognition prototype,

numbers of experiments are done by adjusting the parameters of back propagation

training algorithm. Finally, this research found that Artificial Neural Network and back

propagation algorithm is suitable for image and pattern recog nition.

I l l



TABLE OF CONTENT

TOPIC PAGE

A CK N O W L E D G E M E N T

ABSTRACT

TABLE OF CONTENT

LIST OF TABLES

LIST OF FIGURES

11

ii i

iv

viii

ix

CHAPTER 1 INTRODUCTION

1.0 INTRODUCTION

1.1 PRO BLEM STATEM ENT

1.2 AIM OF THE PRO JECT

1.3 OBJE CTIVE OF THE PRO JECT

i .4 SCOPE OF THE PROJECTSIGNIFICANCE OF THE PRO JECTC

1.6 CONCLUSION

1

1

2

2

23

3

CHAPTER 2 LITERATURE REVIEW

2.0 INTRODUCTION 4

2.1 ARTIFICIAL NEUR AL NETW ORK (ANN) 4

2 .1 .1 REAL WORLD APPLICATION OF NEUR AL NETW ORK 52.2 BACK PROPAGATION 6

2.3 FRAMEWORK

2.3.1 CAM ERA FRAM EWO RK 7

2.3.2 MAXIMU M ENTROPY AND GAUSSIAN MO DEL 7

IV



2.4 IMAGE ENHAN CEME NT 7

2.4.1 EDG E DETE CTION 8

2.5 IMPRO VING RECOGN ITION 9

2.6 TRAD EMA RK RECOGN ITION 10

2 .7 LOGO RECOGNITION USING RECURSIVE NEUR AL NET WO RK 11

2.8 CON CLU SION 11

CH A P T ER 3 RE S E ARCH A P P RO A CH A N D M E T H O D O L O G Y

3.0 INTRO DUC TION 12

3.1 FRAM EW ORK 13

3.1.1 LOGO COLL ECTION 14

3.1.2 PRE-PROC ESSING 14

3.1.3 LOG O DETEC TION 14

3.1.4 SEGM ENTA TION 14

3.1.5 LOGO RECO GNITION 15

3.2 METH ODOLO GY 16

3.2.1 PROJEC T DEFINITION 18

3.2.2 DATA AND LOGO COLL ECTION 18

3.2.3 DATA ANA LYSIS 18

3.2.4 LOGO PRE-PR OCE SSING 19

3.2.5 LOGO SEGM ENTA TION 20

3.2.6 PROJE CT DESIGN 22

3.2.7 PROJECT IMPLEME NTATION 22

3.2.8 PROJECT TRAINING AND TESTING 22

3.2.9 PROJECT DOCU MEN TATION 25

3.3 CON CLU SION 26

CHAPTER 4 CONSTRUCTION

4.0 IN TROD UCTION 27

4.1 NEURAL NETWORK PROGRAM 27



4.2 IMAGE REPRESENTATION 28

4.3 PROTOT YPE INTERFACE 29

4.4 SYSTE M FILE 29

4 .5 NETW ORK TOPOLOGY AND PARAMET ERS 30

4.6 IMAG E MAP PING 31

4 .7 HARDWA RE AND SOFTWARE REQUIREMEN T

4.7.1 HAR DW ARE 32

4.7.2 SOFTW ARE 32

4.8 CONC LUSION 32

CHAPTER 5 RESULT AND ANALYSIS

5.0 INTRO DUC TION 33

5.1 NETW ORK CONVER GENCE 33

5.2 NETW ORK PARAM ETERS 34

5.2.1 MOM ENTU M RAT E 34

5.2.2 LEARN ING RATE 35

5.2.3 INPUT NO DE 35

5.3 NEURAL NETW ORK TOPOLOG Y 35

5.4 RECO GNITION PERCE NTA GE 36

5.5 CON CLUSION 37

CHAPTER 6 CONCLUSION

6.0 INTROD UCTION 38

6.1 BEN EFIT 38

6.2 RECOMM ENDATION 396.3 CON CLUSION 39

REFERENCES 41

VI



APPENDICES

A: INTERFACE FOR LOGO RECOGNITION USING ARTIFICIAL 46

NEURAL NETWORK

B : IMAGE OF LOGO 49

C : BINARY REPRESENTATION OF LOGO 50

D : BACK PROPAGATION ALGORITHM TRAINING PROGRAM 55

E : TESTING PROGRAM 66

V l l



L I S T O F T A B L E S

T A B L E P A G E

Table 3.1 Neu ral network architecture

Table 4.1 Network topology and parameter

Table 4.2 Ne twork output for logo image ma pping

Table 5.1 Netw ork topology and param eter

Table 5.2 Netw ork topolog y and param eter

22

31

31

34

36

V l l l



L I S T O F F I G U R E S

F I G U R E P A G E

Figure 3.1 Fram ework for logo recogn ition 13

Figure 3.2 Method ology for logo recog nition 16

Figure 3.3 Result of pre-proc essing 19

Figure 3.4 Result of segme ntation 20

Figure 3.5 Image representation in binary 21Figure 3.6 The training phase 23

Figure 3.7 Back propag ation flow chart 24

Figure 3.8 The prediction or recogn ition pha se 25

Figure 4.1 Image representation 28

IX



C H A P T E R 1

I N T R O D U C T I O N

1.0 I NTRODUCTI ON

Humans are very good at image and pattern recognition. They will know it by their

experience and learning process. By using this idea and concept numerous product have

been developed based on computer vision. Computer vision researchers aim at

reproducing the capability of recognition in machines. The machine or computer

program s that have the ability of recognizing are ve ry useful.

Recognition of graphical object is one of image and pattern recognition. Much attention

has been paid on this recognition of graphical object, like company logos and

trademarks in the field of documen t image pro cessing .

Nowadays, there are many product use a logo that is not genuine in order to sell their

product by cheating the customer. So, customer can use this application to make sure the

logo is legal like the logo of PETRONAS and EON (refer Appendix B).

1.1 P R O B L E M S T A T E M E N T

By looking at the logo, human can know and told what logo is it. Machine also can

perform it by learning process. To make the application able to recognize the different

logo, this project must know the pattern and the characteristic of each logo. The

application m ight be able to learn the shape and the ch aracteristic of that log o.



However this project must consider if we dealing with spot-noisy logo, where strips or

blobs produce a partial obstruction of the logo. This problem might affect the

performance of it. A suitable technique is needed in order to enha nce the ima ge.

1.2 AIM OF THE PR O JEC T

To develop and evaluate back-propagation artificial neural ne twor k in order to re cogn ize

logo.

1.3 OBJECTIVE OF THE PR OJE CT

The ob jectives of this project are :

1. To identify and evaluate the existing technique that has been used in recognition

2. To identify artificial intelligence technique suitable for logo recognition.

3. To identify the criteria that will be used in order to recognize logo and technique

to manag e the problem of noise in logo.

4. To develop back-propagation neural network prototype that can recognize logo.

1.4 SCO PE OF TH E PR OJ EC T

This prototype was recognizing the logo by its unique characteristic. Types of logo that

were considered are logo with image only and logo with image and text.



1.5 S I G N I F IC A N C E O F T H E P R O J E C T

This project was identified and evaluates the existing technique that has been used for

image and pattern recognition. After the process of identifying, this research was found

out the suitable technique used for logo recognition. In addition, this project was

identifying the criteria and features that are needed to recognize the logo.

1 .6 CONCLUSI ON

This chapter described about introduction of problem, problem statement, aim of

research, objective of research, scope of research and significance of the project.

There are many researches on image and pattern recognition that can be used as a

reference and guideline for this project. The existing technique that is back propagation

neural network was widely used for image and pattern recognition.



C H A P T E R 2

L I T E R A T U R E R E V I E W

2.0 INTRODUCTION

Image and pattern recognition is one of the most popular areas that have been conducted

widely among the researches. These researches have been conducted in order to

summarize the best technique to pre-processing the image, tool and best technique use to

recognize the image and framework that can be used in image and pattem recognition.

2 .1 AR TIFI CI AL NEURAL NE TW OR K (ANN)

Artificial Neural Network (ANN) is a very powerful method that can be used to solve

the most difficult problem that cannot be solving by conventional or traditional

computational method. In other words, artificial neural network is very suitable for a

complex system like prediction, classification and recognition problem. However the

most problem that applied neural network successfully is the recog nition problem .

ANN makes it possible for the machine learning to learn from experience and example

of particular problem. Generally, ANN is made of input layer, hidden layer and output

layer. In order to make it leam efficiently, it will updated and adjust the nu m erical

weight of the neural network.

In addition, ANN also used for image processing. ANN is useful in image processing as

either non-parametric classifiers, non-linear regression functions or for unsupervised



feature extraction. It was review that more than 200 applications of neural network in

image processing (Egmont-Petersen, M. and de Ridder, D., 2002).

ANN also has been implemented in a parallel pyramidal technique for 2-D object

recognition. The combination of structured neural networks and a highly efficient

feature extraction mechanism will leads to a lower number of weights and a fixed neural

network structure. It was found that the set learning was stopped after 26 learning

epochs, with 100% of recognition percentage on training and 93.75% on test, while it

was stopped after only seven epochs with 98.61% of recognition on training and 99.31%

on test when the provided the letter pattern set. (Ca ntoni, V. and Petro sino, A ., 200 2).

Other model that has been introduced as a simple model is Pulse Coupled Neural

Networks (PCNN). This model proved to be highly applicable in the field of image

especially image segmentation (Muresan, R.C.,2000).

2.1 .1 R E A L W O R L D A P P L I C A T I O N O F N E U R A L

N E T W O R K

Artificial Neural Network has been used in many areas of application. The areas

are :

1. Industrial inspection : inspect the quality and process control, e.g., the

production of food, steels and textiles (Egmont-Petersen, M. and de Ridder,

D., 2002).

2. Docum ent processing : Retaining docum ent format in O CR (Butler, T.,

1995), Recog nition of fax docu ment (Ricker, J. and W inkler, A. S., 199 5).



3. Identification and authentication : Off-line signature recognition and

verification (Baltzakis, H. and Papam arkos, N., 2000 ).4. Medical diagnosis : ANN contribute to the improvement of imaging

information and spread of intelligent systems in medical imaging (Karkanis,

S. andM agou las, G.D., 1999).

2.2 BACK PRO PAG ATION

Back propagation is the most popular method for learning process. Back propagation is

learning by example. It consists of two phases that is input and output layer. An input

will put into the network and will propagate from layer to layer to get the output. This

output will compare to the desired output and an error will calculate. If there is an error,

the error value is then propagated backwards trough the network and small changes are

made to the weights.

The back propagation method is a supervised learning method because it performs

feedback to a previous layer from the output layer, such as the input layer or hidden

layer, to update weights so that we eventually are matchin g o ur expected output.

Back propagation is suitable to use if there is a large amount of input or output data is

available. It is also suitable for problems that appear to have overwhelming complexity,

but there is clearly a solution. In addition, it is suitable for recog nition prob lem.

Lastly, Spot back propagation that updating the classical back propagation was

introduced. Spot back propagation is a leaming algorithm that taking into account the

different error fiinctions. This algorithm is significantly much more robust with respect

to spot-noise than classical back propagation (Ces arini, F. and Gori, M., 19 97).



2 . 3 F R A M E W O R K

2 . 3 . 1 C A M E R A F R A M E W O R K

Camera is an open-source system that allows a user with particular knowledge of

the documents to be recognized. It combines image processing and recognition

tools in an easy to use. (Droettboom, M. an d Fujinaga, I., 200 2).

Camera modules perform several tasks that were pre-processing, document

segmentation and analysis, symbol segmentation and classification and

syntactical or structural analysis.

2.3 .2 M AXI M UM ENTRO PY AND GAUSSIAN M OD EL

Maximum Entropy is a fi-amework that can b e used to estimate class p osterior

probabilities for pattern recognition task. It has been applied to the estimation of

probability distribution. In addition, this fi-amework allows estimating a large

number of parameters, so it is suitable to classification task such a text

classification. Advantage of this framework is that it is easily to include new

feature fiinctions into the classifier (K eysers, D., N ey, H ., 200 0).

2.4 IMAG E ENHA NCEM ENT

Image enhancement operators improve the detect ability of important image details or

objects by man or machine. Example operations include noise reduction, smoo thing,

contrast stretching and edge enhancement (Shapiro, L.C., Stockman, C.C, 2001).



2.4 .1 EDGE DETE CT ION

Edg e detection is the mo st comm on approa ch for detecting meaningful

discontinuities in gray level (Gonzalez, R.C., Woods, R.E., 2002). A good edge

detector will provide accuracy for pattern and im age recognition.

There are many type of edge detector method that can be used to detect the edge

of image like Sobel method, Prewitt method, Roberts m ethod, Laplacian of

Gaussian method. Zero-cross method and Canny m ethod. Am ong all this, theCanny edge method is a very popular and effective o perator (S hapiro , L.G.,

Stockman, G.C., 2001 ). It also has been stated that the C anny edge detection

algorithm is the optimal edge detector (Green , B., 2002). It takes as input a grey

scale image, and produces as output an image showing the positions of tracked

intensity discontinuities (Fisher, B., Perk ins, S., 1994). The C anny detection first

smoothes the intensity image and then produces extended contours segments by

following high gradient magnitudes from one neighborhood to another (Shapiro,

L.G., Stockman, G.C., 2001).

A list of criteria can be used in order to improve the performance of edge

detectors (G reen, B., 2002):

1. Lo w error rate.

Edg e that occurs in image should no t be missed and there be no respon ses to

non-edges.

2. The edge points must be w ell localized

Th e distance between the edge pixels as found by the d etector

and the actual edge is to be at a minim um.



3. Have only one response to a single edge.

This criteria is needed becau se the first two criteria w ere no tsubstantial enough to completely eliminate the possib ility of

multiple responses to an edge.

2.5 I M PRO VI NG REC OG NITION

In order to improve recognition, fast and efficient methods must be applied to the

application of recognition. Image Block Representation (IBR) was introduced in the

feature extraction stage of such a system for the fast implementation of pattern pre

processing techniques. In addition, a Constrained Optimization based approach can be

implementing for efficiently tiaining feed forward neural network of the multilayer

perceptron that can lead to effective pattem recognition system. Usually both concepts

are involved in the design of an improved Optical Character Recognition system

(Mertzios, B.G. and Karras, D.A., 1999).

Image pre-processing is the important stage in recognition process. Recognition

performance is improved by pre-processing the input images so that intensity changes at

different fi-equencies and orientation are enhanced (Viola, P.A., 1996). We must

improve the quality of an image that we want to recognize to get a good result. In order

to make an enhancement of image, the image can have a process of adaptive contrast

enhancement, nonlinear histogram equalization, high boosting, local histogram

equalization and gamma correction (Satrovoitov, V.V. and Samal, D.I., 2003). All this

method can improve the brightness of the image and produce less noise.

Furthermore, deformable template has been shown to significantly improve the

performance of image recognition for difficult tasks such as character recognition, digit

recognition, and trademark recognition. However, it is often time consuming. Grid



feature is a popular feature extraction scheme in image recognition. It was found that by

using deformable grid it can improve the recognition rate of off-line Chinese characterrecognition by 5.8% and 6.0% improvement for trademark recognition (Jikun, Y. and

Jianmin, X., 2003).

Segmentation is also an important process for the image that considers recognizing. The

aim of segmentation is to partition the image into a number of regions. The researchers

must choose the best method of segmentation to get the best performance. There are

several methods that are :

1. Histogram analysis that use thresholding that is simplest segmentation process,

inexpensive and fast.

2. Region growing and split-merge algorithms that successively divide an image into

Ctv »o ll ^- r •ii-**>_i^T7c»«*lo»>rMn r» f£\rrtr^v% i-r o -v*vi i 0«*"* « I O * M + ^ X <-**-»+,ck»«<r» « o * - » J ^ + *-^^t^4-

UXXAiAAmwA A.a.\jrAX V> V ^ / A X l ^ t ^ L y x x x ^ 1. W ^ X V /1 .X AJI. O V ^ l l l . ^ / OJiXXAXXUM XI. y ^ / I X C W A X C V Aid I X \ J l . X X XW k .

3. Colour-space analysis and segmentation that analyze the individual co lor features.

4. Watershed segmentation that can be applied to wound image.

5. Radial search algorithm that simplify the task of edge detection.

6. Boundary following algorithm that have been used to find the edges of cells in

microscopy image.

7. Active contour models that is able to find contours in the image.

2.6 T R A D E M A R K R E C O G N I T I O N

The increasing number of trademark makes it important to develop trademark

recognition. The grid feature was used in experiment represents local direction of

trademark contour. A grid was obtained by partitioning the trademark image

hierarchically in both horizontal and vertical directions in order to balance the number of

pixels inside up and down part (for horizontal partition) or inside left and right part (for

vertical partition). However, the contour of trademark fails to represent all information.

10



Therefore, inner black pixels (contrast to contour) are used too( Jikun, Y. and Jianmin,

X., 2003).

2.7 L O G O R E C O G N I T I O N U S I N G R E C U R S I V E N E U R A L

N E T W O R K

One of logo recognition is implemented by using recursive neural netw ork. Log o images

are converted in a structured representation based on contour trees, where symbolic and

sub-symbolic information coexist. A contour-tree is constructed by associating a node

with an exterior or interior contour extracted from the logo instance. Nodes in the tree

are labeled by a feature vector, which describes the contour by means of its perimeter,

surrounded area, and a synthetic representation of its curvature plot. The contour-tree

representation contains the topological structured information of logo and continuous

values pertaining to each contour node. Afterwards, recursive neural networks are used

to leam and recognize the logo instances represented by contour-trees. Experimental

results are reported on 40 real logos distorted with artificial noise (Francesconi, E. and

Frasc oniP., 1997).

2.8 CONCLUSION

As a conclusion, artificial neural network is suitable for image recognition. Back

propagation technique is efficient for the learning process of neural network. While

doing the recognition, there is many things that must be consider by the researcher.

11



C H A P T E R S

P R OJEC T AP PR OAC H AND M ETH ODOLO GY

3.0 INTRODUCTION

Project approach and methodology is a step by step process that must be completed

successfully to get the project objective. A good project approach and methodology will

guide researcher to done their work more arrange able and on scheduling. This is

important in every project.

This chapter gives an explanation about the framework and a methodology used in order

to implem ent and develop the project or the prototype . The framew ork invo lves steps of

logo collection, pre-processing, logo detection, segm entation and lastly logo recogn ition.

Methodology involves nine steps that are project definition, data and logo collection,

data analysis, logo pre-processing, logo segmentation, project design, project

implementation, project training and testing, and the last one but very important is a

docum entation process. Every steps will discussed and exp lain later.

12



3.1 FRAMEWORK

SCAN/CAPTURED LOGO IMAGE

COLLECTION OF LOG O

PRE-PROCESSING

BLACK AND WHITE LOGO

L O G O D E T E CT I O N

FEATURE OF LOGO

^

C T ? / ^ A y *" !I ?X T T^ A n P T / ^ X T

SEGMENTED LOGO

^

L O G O RE CO G N IT

NAME OF LOGO

TION

Figure 3.1 Framew ork for logo recogn ition

13

nor hamidah abdul ghafar 05 24

Documents