Workshop on InformationWorkshop on Information

H t t f

Session Chair: Zoran Skocir, University of Zagreb, Croa

Technologies IIITechnologies III

Human resources management system forIvona Zakarija, University of Dubrovnik, Croatia; ZKrunoslav Zubrinic, University of Dubrovnik, Croa

Recognition of damaged characters – MontZeljko Deljac, Robert Mostak, HT-Croatian telecoHigh School, Croatia

Providing Data Grid Capability Using SocietMeisam Hejazinia, Mohammad Reza Razzazi, Amj , ,

New approach to designing a Zero-overshoamplifiers Bertrand Gerfault, Thales communications R&D, Chahbazian GERAC R&D FranceChahbazian, GERAC R&D, France

n and Communication n and Communication

Hi h Ed ti i tit ti

atia

r Higher Education institutions Zoran Skocir, University of Zagreb, Croatia;atia

te Carlo method mmunications, Croatia; Predrag Brodanac, V.

ty Network on Mobile Phones mirkabir University of Technology, Irany gy,

oot Automatic Level Control for high-power

France; Balwant Godara, ISEP, France; Frank

Workshop on InformationWorkshop on Information

Th F t d Ch ll f S t M

Session Chair: Ivan Marinovic, University of Split, Croa

Technologies IIITechnologies III

The Future and Challenges for Spectrum MShkelzen Cakaj, Muharrem Shefkiu, Post and Tel

Selection of the base set of methods for coIvica Ruzic, University of Split, Croatia; Dalibor M–BICRO, Croatia; Mojca Ciglaric, University of Lju

Comparison of threshold methods for playeVladimir Plestina, Nikola Rozic, Vladan Papic, Un

Influence of unsharp masking on OCR perfoMatko Saric, Dinko Begusic , Hrvoje Dujmic, Univ

Design and Implementation of Smart GuideDiscoveryDiscoveryHyunKyung Yoo, JeongHwan Kim, SangKi Kim, E

n and Communication n and Communication

t A t ti i K

atia

anagement Automation in Kosovo lecommunication of Kosovo (PTK), Kosovo

mputer forensic Marijanovic, Business innovation centre of Croatia

ubljana, Slovenia

ers segmentationniversity of Split, Croatia

ormanceversity of Split, Croatia

e Services with the Semantic Service

ETRI, Republic of KOREA

The Future and Challenges for Spectrum Management Automation in Kosovo

Shkelzen Cakaj1, Muharrem Shefkiu2

1, 2 Post and Telecommunication of Kosovo (PTK), PTK Building, Dardania, nn. Prishtina, Kosovo

E-mail: shkelzen.cakaj@ptkonline.com; muharrem.shefkiu@ptkonline.com Abstract: PTK (Post and Telecommunication of Kosovo) is incumbent telecommunication operator, still as a public owned company providing fixed and mobile services in Kosovo. Mobile operator (VALA) has around 880.000 customers. On the fixed side it is not done too much on penetration increment (currently it is around 6%), but very good steps on technological level are taken. The complete network structure is digitalized, based on TDM and NGN concept. Both these structures today are operational providing services for voice and data through XDSL fixed lines. The transmission infrastructure is based on MW and F.O., always aiming toward building the strong fiber optic based structure. Since late 2007, the second operator (IPKO) for fixed and mobile services is operating, also. Specifically, the present international institutions are using MW links for their purposes and services. There are present ilegal operators, also. Thus, the radiospectrum should be well controlled and managed in order to be used efficiently and to avoid interferences. Through this paper the future aspects toward the spectrum management automation process in Kosovo’s network are presented. Keywords – Radio Spectrum, Management Functions, 1. INTRODUCTION

”The radio frequency spectrum is a limited natural resource and must be used rationally, efficiently and economically, so that countries and groups of countries may have equitable access to it. Radio waves propagate in space with no regard for political frontiers” – by ITU (International Telecommunication Union). International Telecommunication Union is United Nations’ specialized intergovernmental organization in the field of telecommunications which brings together governments and industry to establish and coordinate the operation of global telecommunication networks and services. Consists of Member States & Sector Members representing public and private companies and organizations with an interest in telecommunications [1]. Thus, the ITU has primary worldwide role on spectrum management policy. With the introduction of new technologies and services, it is very important to address the business approach to spectrum management and its impact on public services and social aspects. From this prespective it is considered the spectrum management automation concept for Kosovo. 2. SPECTRUM MANAGEMENT GOALS Traditional MW links, satellite access systems for massive and scientific applications, different mobile

system for public communication and for other military purposes, emergency systems, wireless broadband access services, etc, are present in everyday life worlwide. On the other hand, the trend of emerging wireless services to the customer’s premises not only for communication but also for other practical needs, including ultra wide band services, makes obvious neccesity of frequency allocation and coordination among these services in order to provide massive access, reliable and accurate services. This process is more sensitive in urban areas, because of higher density, specifically in megacities which are becoming reality [2]. The general concept of these services’ presence in high density areas, is depicted in Fig, 1.

Fig. 1. Different wireless services The world is indeed becoming a smaller place for spectrum use related to telecommunications services. Thus, the spectrum management aspects are consdiered from international and national point of view. International aspects of spectrum management consists of a number of interrelated activities leading to inter-government agreements. From the national point of view, the aim of spectrum management is to enable the variety of telecommunication services, which should operate effectivelly and share the radio spectrum without interference. Considering these aspects, yields few main goals of spectrum management, as following [3]: • To ensure that the radio frequency spectrum is utilized and managed in an orderly, efficient and effective manner;

WICT/III - 1569219751- 2509 © SoftCOM 2009

• To reduce congestion in the use of frequencies and to protect frequency users from any interference or other inability to make use of the frequencies assigned to them; • To avoid obstacles to the introduction of new technologies and telecommunication services; • To provide opportunities for the introduction of the widest range of telecommunication services and the maximum number of users, thereof as is practically feasible. From the prespective of achieving these spectrum management goals for Kosovo telecommunication network, main figures of this network structure are further presented. 2. KOSOVO’s NETWORK STRUCTURE The existing transmission network at Post and Telecommunication of Kosovo is build of two systems Micro-Wave (MW) and optical fiber (OF) configured in meshed topology, always aiming toward building the strong fiber optic based structure [4]. MW is used where there is no fiber optic access or as a redundancy. MW links operate at 7GHz, 13GHz and 18GHz. This transmission platform is serving both, PTK’s mobile operator (VALA) and fixed network. This network structure is presented in Fig. 2.

Fig. 2. Network strucute VALA as a mobile operator has around 850.000 prepaid customers and around 30.000 as postpaid, providing voice, SMS, GPRS and MMS services. VALA has roaming agreements with around 240 worldwide operators. The coverage is achieved with more than 300 ALREADY implemented BTS. The actual coverage is presented in Fig. 3.

Fig. 3. Mobile coverage It is already completed expansion phase of VALA, based on NGN platform, with planned capacities to provide services for 1.200000 customers. On the fixed side, the complete network structure is digitalized, based on TDM and NGN concept. Both these structures are operational providing services for voice and data through ADSL fixed lines. PTK, as telecommunication service provider in Kosovo, has done a large step forward IP transformation by implementation of next generation network IP MPLS (MultiProtocol Label Switching) based, from the edge to the core level [4]. The future steps are oriented on end - to - end Ethernet transmission network transformation. In order to increase penetration for fixed services and create conditions for broadband access all over Kosovo, the Govenement of Kosovo has established national strategy for implementing fixed wireless access (FWA) in frequency band 3.4GHz – 3.6GHz inside territory of Kosovo. It is planned these services to be provided by four licensed operators [5]. Another factor with impact on spectrum management is recently licensed operator (IPKO) for fixed and mobile services. Becasuse of large international presence in Kosovo, different governmental or nongovermental organizations are using their own MW networks for civil or militarry purposes. Additonaly there are also ilegal operators providing wireless voice and data services IP

based. Since, Kosovo is a small country (10800 km2 ), and all present operators intend to as higher coverage level, the harmfull interfernce could be avoided only by sofisticated spectrum management system. These few factors make obvious neccesity of controlled and managed radiospectrum, what can be achieved by automation of spectrum management system. Up to early 2008, the frequncy allocation procces was under UN (United Nations) responsibility, and then by transfering compentencies to Kosovo’s government, this

WICT/III - 1569219751- 2509 © SoftCOM 2009

process in the future will be under responsibility of Telecommunication Regulatory Authority (TRA) of Kosovo [5]. Thus, further concepts for automation of spectrum management process are given. 3. SPECTRUM MANAGEMENT FUNCTIONS National frequency allocation policy goal is to establish legal basis for submissions and updating process for national frequency allocations data. This approach should provide a stable platform for national telecommunication infrastructure and major future investments. ITU, has developed an approach considering factors for spectrum management and monitoring, as it is presented in Fig. 4 [6].

Fig. 4. Spectrum management ITU approach Toward automation process, these factors can be categorized as following: • Strategic Spectrum Planning (Modify Table of Allocations, Adopt Channel Plans, Establish Fixed link Policy, Reserve Bands) • Spectrum Engineering (Establish Sharing and Interference Threshold Criteria) • Equipment Type Acceptance (Implement Band, Channel, Capacity Standards) • Frequency Assignment and Licensing (Identify Current Use, Implement Band, Service, Channel, Sharing Criteria) • Enforcement and Monitoring (Implement Band, Service, Assignment/License Standards)

Each of these spectrum management activities requires: Policies, Standards, Procedures, Automation and Staff Training, corellated with respective activity as presented in Table1 [3]. Table 1. Spectrum management functions

Above functions are general, and those could be operational only by being interlinked at respective data base and appropriate software. Thus, each country shoud establish its own single data base shared across all spectrum management activities. Then, spectrum management functions could be implemented by having access to respective national data base. The process chain is described through Fig. 2 [3].

Fig. 5. Spectrum management process

4. RADIO SPECTRUM MANAGEMENT

In order to build a computerized system for radio spectrum management for Kosovo, based on spectrum management processes as in Fig. 5, we have to consider three main elements:

• Database • Client Applications and • Network

NationalSpectrum

Management

Database

Monitoring and

Enfo

rcem

ent

Frequency

Assignm

ent

Type

Accep

tanc

e

Spectru

m

Engin

eerin

g

StrategicPlanning

POLICY

STRATEGIC

PLANNING

STANDARDS PROCEDURES AUTOMATION

SPECTRUM

ENGINEERING

TYPE

ACCEPTANCE

ASSIGNMENT/

LICENSING

ENFORCEMENT

TRAINING

WICT/III - 1569219751- 2509 © SoftCOM 2009

In contemporary systems, a computerized management system considers software and hardware resources to offer local and remote access to different databases and applications. The appropriate computerized radio spectrum management scheme is presented in Fig. 6.

Fig. 6. Computerized radio spectrum management scheme 4.1. Data Base Content Data base should content following groups of data which have to be considered through decision proccess for license issue. • Geographic and band occupancy with appropriate

usage statistics. • Tables of allocation, channel plans and sharing criteria. • Equipment characteristics and technical standarts. • Interference monitoring and reporting. • ITU and national regulatory criterias and rules. Geographic data base should be multilayered provided through GIS (Geographic Information System) as presented in Fig. 7 [3].

Fig. 7. Geographic information system

Kosovo’s Government has established a task group for creating GIS for Kosovo including above and under ground infrastructure. This is a major component toward spectrum management automation. Considering national data base, spectrum management functions, and then applying the process approach as in Fig. 5 the frequency assignments and licenses are defined and then issued. Further, spectrum management should be permanently monitored and these monitoring results have to be as inputs to national data base as in Fig. 4. Main goals of spectrum monitoring are: Surveying and idetifying changes in the RF environment and then locate the source of interfering or undesired signals. Results of these activities are recorded as histogram of RF environment for subsequent analysis. Usualy, the entire band from 100MHz to 40GHz and national wide is monitored. Specifically, for Kosovo case it is very important to be surveyed areas around borders, since Kosovo is a small country, so undesired signals can be present too deep within a country. Through this process the outputs should be in SIS (Spectrum Information System) as in Fig. 8 [3].

Fig. 8. Spectrum information system Spectrum Information System (SIS) should provide capabilities similar to GIS for frequency dependent information as are: Table of allocations, channel plans, predicted received signals, monitoring data and compliance measurements. All these data have to be organized into a central database managed and maintained by regulatorty body (RB) personnel. This regulatory body in Kosovo is known as TRA(Telecommunication Regulatory Authority) [5]. To create a proper database for including all required data, requires a big team work and respective time frame. Therefore for this paper we have created a sample database using Oracle10G Ex. The picture bellow shows the DB tables explored using Visual Studio.

WICT/III - 1569219751- 2509 © SoftCOM 2009

Fig. 9. Database tables

4.2. Client Applications Using visual studio, a sample client tool is developed to access locally and remotely the database server (see Fig. 6). The main tool’s window enables access to all data and applications within the management system, as in Fig. 10.

Fig. 10. Main application window Within frequency application module there is a FREQUENCY APPLICATION form for a new radio link request. This form contains all necessary information and parameters required for the equipment and configuration of two sites A and B. The frequency application module is presented in Fig. 11. All providers have to fill this form in order to obtain approval by RB (Regullatory Body) for every radio link they want to install. This request is treated by RB Frequency/Spectrum unit and the result afterwards is presented in application result module, and each applicant can see the result immediately after being processed by RB experts. This application for security reasons initially will be given only to licensed service operators. This will be the first step and in mean time online web-based application will replace this tool completely. Service operators will have certain number of authorized users to access the database and get information about spectrum allocation, GIS, SIS and read/write access to lectronic available applications for new radio links or any modification of existing radio links.

Fig.11. Radio link application module 4.3. Network

Operators should access the management system (Database) remotely, therefore network resources are necessary, to enable the functionality of the frequency management system. For access to the main RB server, providers will use internet broadband access. For management and internal processes within RB, the system will be reached by IP local area network (LAN). CONCLUSIONS The competition in the telecomunication market in Kosovo is alrerady reality. Incumbent and other alternative operators are using MW links because of lack on fiber deployement. Other wireless services for voice and data are growing rapidly. Thus the spectrum management have to be under control in order to be well managed. The orientation is building modern automated spectrum management system, based on experiences of EU countries and ITU recommendations. The spectrum management monitoring tools will be under the same investement program. Within this paper the concept of spectrum management and monitoring automation is given. REFERENCES [1] G.A. Codding, Jr and A.M Rutkowski, The

International Telecommunication Union in a Changing World, (Artech House, 1982).

[2] S. Hassler, “Engineering the megacity”, IEEE Spectrum, Vol. 44 (6), pg.14, June 2007.

[3] Alion (Science and Technology) , “Automation Aspects of National Spectrum Management”, presentation sl. 1-137, USTTI, Washington, DC, 2005.

[4] Sh. Cakaj, “From Analogue Infrastucture toward Converged Services in Kosovo’s Telecommunication Network”, 48th FITCE Congress, London, UK, Sept. 2008, pp. 196-200.

[5] http://www.art-ks.org/ [6] B. Balabanov, “Automation of the Spectrum

Management – Training Tools for Automating Utilization of the ITU-BR SM documents and Databases”, Centre of Excellence Workshop, 12-15 July, Kyiv, Ukraine, 2005.

WICT/III - 1569219751- 2509 © SoftCOM 2009

Selection of the base set of methods for computer forensic

Ivica Ružić1, Dalibor Marijanović2 and Mojca Ciglarič3

1University centre for professional studies, University of Split; 2 Business innovation centre of Croatia – BICRO; 3 Laboratory of Computer Communication, Faculty of Computer and Information Science, University of Ljubljana, Slovenia

Abstract: This paper explains and propose new basic set of methods for computer forensics while examines suspects computers or other storage media. Penetration of information and computer technology in all human activities extremely increased in last decades. Therefore, crime related to that technology has also increased. Appropriate measures are needed to meet new challenges raised by usages of new technologies.Insight in today most used methods for computer forensic, identify their weaknesses and deficiencies. Analyzing findings creates groundwork for development new set of methods.Result is new set of methods, including four steps and two objectives. Steps are: i) a detailed examination of the discovered situation, ii) computers and discs search, iii) analysis and iv) presentation. Objectives are: i) preservation of data and ii) authentication.Proposed set of methods ensures obtaining of authentic evidence that can be used in court. If we wish to use the evidence in real court proceeding, we must strictly adhere to all the proposed procedures, because this is the only way we can be sure that the court will not dismiss them as invalid. In case we do not intend to use the obtained evidence in court proceedings, it is not necessary to adhere strictly to all the procedures.

1. COMPUTER FORENSIC

Traditional forensic science developed from the forensic medicine practice already at the end of the 18th century, but computer forensics itself is a very young discipline and for the general public, including, unfortunately, many lawyers, it is still a novelty. Nevertheless, computer forensics has been used in court processes for decades and evidence obtained in such a manner has already been recognized in court. We can say that computer forensics is the discipline of computer science usage in legal processes.

Computer forensics is applicable in private as well as public sector. In the public sector, it deals with computer-related criminal activities and is implemented by court and law enforcement authorities. In the private sector, it deals with violation of individual organization rules or is used in civil actions. In these cases, the major part of the investigation can be carried out in the organization or insurance company itself without involving the police forces.

Apart from having roots in the forensic science, computer forensics can also be viewed as an area within information security field. An important feature coming from computer science is simple and well determined: digital evidence can be traced. Each data, file or document stored in the computer

has a unique fingerprint, maybe even the entire hard disc. By using computer forensics, it is even possible to examine previously deleted data as well as decrypt encrypted information.

J.M. Patzakis [3] defines computer forensics as the collection, preservation, analysis and court presentation of computer-related evidence, while I.Takahashi [6] defines it as the recovery of deleted computer-based information and the science of examining and putting back together the who, what, when, where, and how of computer-related conduct.

Computer-related crime is on the increase and the growth trend in recent years has reached 250% per year [3]. Often the computer-related evidence constitutes a major part of the evidence in court, by the size of the discs as well as number of computers and data storage media that need to be examined. On the other hand, the sluggishness of the legal system which adapts too slowly to the needs of computer forensics is an additional problem. An average investigation of the pedophile network may involve millions of people and consequently at least that many computers need to be examined. With regard to accessibility of data through broadband networks, we can expect the number of suspicious computers will increase even faster in the future.

Due to the nature of their usage, computers themselves can play an important role in various criminal activities and in many ways. Among other things, a computer could [1]:

• be the target of a crime, including information theft, financial fraud, denial of service, or other direct attack,

• be used to commit crimes against other computers,• be used to commit non-computer crimes, such as

creating false documents or counterfeiting currency,• be used to illegally copy or distribute copyrighted

materials such as music or movies, or to store illegal documents, such as child pornography or information stolen from government agencies or corporations, or

• contain information such as contact lists, copies of falsified documents, or email that documents a conspiracy, which investigators could use to prevent or solve crimes.

Experienced criminals avail themselves of the fact that digital data are much more permanent then an average user believes, so they hide illegal or controversial data among deleted files or in other places in the disc which are inaccessible for the majority of software equipment (available

1

WICT/III - 1569225051 - 2509 © SoftCOM 2009

to the average user for every-day use). Computer forensic investigators already encounter problems such as: increased technical capabilities of criminals, wide-spread usage of strong encryption, constant increase in the quantity of storable data, difficulties in examination of specific data formats, incompatible disk drives and operating systems, etc.

As a consequence of ever growing importance of computer forensics, numerous software tools are available, especially designed for use in computer forensics. Such specialized software is the most important tool for a computer forensics specialist. Actually, forensic analysis is impossible without it. In general, there are three primary reasons why specialized computer forensic software must be employed in order to conduct a proper computer investigation [3]: to preserve and properly acquire the evidence; to authenticate the data for presentation in court; and to recover all available data, including deleted files.

Internet Forensics is a separate area of forensics and is distinct form computer forensics. The computer forensics deals with law enforcement, while internet forensics deals with hacker community for network security [2]. The internet forensics specialist only has something to investigate if the packet filters, firewalls, and intrusion detection systems were preinstalled with the anticipation of the security breach.

1.1 Legal Background

Today, almost every company or organization is computer-dependent. E-mail has become the primary form of business communication. The technological advances in computer and network technologies have caused the creation of an enormous quantity of data. Almost all documents are stored in an electronic form and only a small portion thereof is printed out. Therefore, the need arose upon presentation of digital evidence in court. Digital documents have become as important evidence as hardcopies. Obtaining digital evidence has shown to be everything but simple and courts have realized that computer forensics could be the solution for obtaining and preservation of digital evidence applicable in court, by both the prosecutor and the defense. Today, computer forensics has become irreplaceable in court proceedings.

Mobility and network connectivity of modern computer systems opens numerous competency issues. Which laws were broken: local, national or international, or a combination of all those? Laws of which country are broken if a criminal activity is committed on the Internet? Are these the laws of the state in which that individual is physically located or the laws of the state in which the computer containing stolen or destroyed data is located, or are the laws of both states applicable? The legislation should follow the rapid development of technology and computer-related crime. Otherwise, legal complications can occur which the

criminals might use to evade law enforcement authorities and legal consequences.

Therefore, under the auspices of the Council of Europe i.e. the European Committee on Crime Problems (CDPC), a team of experts was formed in 1997 from the member states, namely, the Committee of Experts on Crime in Cyber-space (PC-CY), with the aim to prepare the so called Convention on Cybercrime which shall set international legal standards for combating computer-related crime. The Convention contains a preamble and four chapters. In the first part (Use of terms), it defines the used terms (computer system, computer data, service providers and traffic data); in the second (Measures to be taken at domestic level – substantive law and procedural law), measures (substantive criminal and procedural criminal) which need to be accepted on the national level; in the third (International co-operation), international cooperation (general principles of international cooperation, extradition of suspects and mutual assistance) and the fourth chapter contains the final clauses (on its validity, subsequent accession to the Convention).

The Republic of Croatia [9] has signed the Convention on November 23rd 2001 and ratified it on October 17th 2002. The Convention came into force on 1.7.2004. By doing so, Croatia has undertaken to modify and amend most of the Articles of the Criminal Code referring to computer-related crime. Therefore, among other things, the possession of child pornography has become a criminal offence and not only production and trade. Before that the law enforcement authorities could do nothing if someone was only watching child pornography over the Internet or if they possessed photos in the computer. After the laws were amended in compliance with the Convention, the aforementioned activities have become criminal offences.

2. COMPUTER FORENSICS AND ITS METHODS

A forensic specialist usually gathers a lot of information about the suspects on the basis of data stored in their computer, programs they have been using, e-mail correspondence and Internet usage. This information can vary, ranging from information about this person's daily activities, to his/her personal characteristics e.g: interests, financial situation, health condition etc. With the rapid scientific development, forensic specialists will soon face a range of new issues, including [1]:

• inevitable increases in criminals' technical skills,• the widespread use of strong encryption,• the increased volume of data that can be stored on disk

drives, and

2

WICT/III - 1569225051 - 2509 © SoftCOM 2009

• difficulties in examine data from legacy software applications or hardware that use proprietary formats, incompatible disk drives, or obsolete operating system.

New tools and techniques have increased the efficiency and speed of detective investigation, but new technologies continue presenting a challenge to forensic specialists. To fight the computer crime successfully, they need quality training and adequate certificates.

2.1 Models

What follows is a description of the three process models used in computer forensics: Each of them has certain shortcomings, and these are dealt with in our proposed set of basic methods in performance of (software) computer forensics.

Incident Response Process Model

Several organizations use this model, however since the incidents that take place in organizations are often not illegal, the model is not intended for police investigation. The examples of such incidents may include: computer use that extends beyond the rules of the organization, breaking into the organization data server with the aim of acquiring confidential information, server blocking. In the book Incident Response [4], an "incident response methodology" is given with the following phases:

• Pre-incident Preparation: Prepare for an incident with proper training and infrastructure.

• Detection of the Incident: Identify a suspected incident.

• Initial Response: Verify that the incident has occurred and collect volatile evidence.

• Response Strategy Formulation: Determine a response based on the known facts.

• Duplication: Create a backup of the system.

• Investigation: Investigate the system to identify who, what, and how.

• Secure Measure Implementation: Isolate and contain the suspect system before it is rebuilt.

• Network Monitoring: Observe the network to monitor attacks and identify additional attacks.

• Recovery: Restore the system to its original state with additional security measures added.

• Reporting: Document the response steps and remedies taken.

• Follow-up: Review the response and adjust accordingly.

One of the main aims of this method is a complete system recovery following the incident, as soon as possible, to the state prior to the incident. It is perfectly understandable that this is the primary goal of an organization, but we have set greater goals to also be able to solve different court cases. This model can successfully be applied for monitoring of the employees in a company, as well as monitoring of computer usage by children (parental controls)

Law Enforcement Process Model

The U.S. Department of Justice (DOJ) has published a process model in the Electronic Crime Scene Investigation Guide [8] that is supposed to help with gathering and storing all kinds of digital evidence. The following phases are given:

• Prepare equipment and tools to perform needed tasks during an investigation.

• Collection: Search for and collect electronic evidence.o Secure and Evaluate the Scene: Secure the scene

to ensure the safety of people and the integrity of evidence. Potential evidence should be identified in this phase.

o Document the Scene: Document the physical attributes of the scene - including computer photos.

o Evidence Collection: Collect the physical system or make a copy of the data on the system.

• Examination: A technical review of the system for evidence.

• Analysis: The Investigation team reviews the examination results for their value in the case.

• Reporting: Examination notes are created after each case.

Unlike the incident response model, much greater attention has been paid to analysis in this model, but we still find it unsatisfying. This model derives from a classic model for collecting physical evidence, but it has not been successfully adapted to collecting the digital evidence. In other words, individual processes are not clearly defined in this model, and the additional inconsistence is created by the fact that, according to this model, digital data is inspected only after all of the evidence has been collected.

An Abstract Process Model

The investigators at the U.S. Air Force have combined several different models into a new abstract model with the following phases [5]:

• Identification: Detect the incident or crime.

• Preparation: Prepare the tools, techniques, and obtain approval.

3

WICT/III - 1569225051 - 2509 © SoftCOM 2009

• Approach Strategy: Develop a strategy to maximize the collection of evidence and minimize the impact on the victim.

• Preservation: Isolate and secure the physical and digital evidence.

• Collection: Record the physical crime scene and duplicate digital evidence.

• Examination: Search for evidence relating to the suspected crime.

• Analysis: Determine significance and draw conclusions based on the evidence found. Repeat examination until a theory has been supported.

• Presentation: Summarize and provide an explanation of the final conclusions and theory.

• Return Evidence: Return the evidence that was removed from the scene back to the owner.

Like the previous model, this one identifies and collects digital evidence within two phases, Examination and Analysis. The names themselves could be confusing because their meanings are similar, even the same for many investigators. It is very likely that two investigators would name an identical procedure differently, e.g.: one of them could call it »analyzing a system« and another one would say »examining a system«. Although we agree these phases should be (and are) distinguished in this model, we have to point out that the naming similarity could lead to confusion.

2.2 Proposition for the basic set of methods

In this section, we will suggest a basic set of methods that every forensic specialist should follow when conducting computer forensics. The proposed set of methods contains all of the good characteristics that other models possess and it is not complicated. Unlike other models, the proposed procedures are clear, precisely defined and ensure the obtaining of authentic evidence that can be used in court. In case one would not follow the proposed procedures, it just might occur that the evidence will be eliminated in court.

A complete forensic investigation includes the following services [6]: testimonies of experts, preparation of reports, examination and reconstruction of critical events, forensic analysis where necessary, and legal and technical advice. The data must be kept in such a way, that it can be obtained and analyzed in a manner without damaging them. Just as well, we must keep their authenticity.

Without the appropriate procedures, the forensic data could become completely useless in a court proceeding. In relation to this, every laboratory that conducts computer forensics should have a document which would set the procedures in advance with the following content [7]:

1. how to manage the incident,

2. how to preserve potential evidence,3. how to analyze the data and information obtained, and4. presentation in court.

After the forensic specialists decide what computers to check, their work must be entirely systematic. From the moment they enter the space with the computer, until completion of the report for the party (legal party or natural person), the forensic specialist must document every procedure, even when it seems insignificant at the first glance. Since the computer is an integral part of the space where the criminal conduct took place, or it is the crime scene itself, it is necessary to act with great awareness in order to protect potential electronic or other evidence. Destroyed, damaged or improperly handled evidence cannot be used in court.

In order to achieve the abovementioned goal, we propose the following four steps and two objectives which should be adhered to:

1. A detailed examination of the discovered situation2. Computers and discs search3. Analysis4. Presentation

• Preservation of data and• Authentication.

We shall now describe in more detail each proposed step and desired objectives.

Figure 1- Steps and objectives of a forensic investigation

A detailed examination of the discovered situation

If you suspect that the computer of the accused contains important evidence, first we must document the physical position of the computer. It is best to photograph it from various angles to record all the components of the hardware and the way the equipment is interconnected. It is desirable to mark the cables, so that we can later reconstruct the position of all devices into the initial state (the moment of dispossession). In case the computer is switched on, we must

4

WICT/III - 1569225051 - 2509 © SoftCOM 2009

look what is on the screen. We check the written notes beside the table and on the labels near the computer. We must check whether there is a note on passwords, which, surprisingly, often happens.

We then take the computer to a location of a highly restricted access. It is recommended to start a diary in order to document chronologically every person that accessed the evidence material, as well as actions performed on the evidence. This diary is called the Chain of Custody and it can be used, if necessary, to prove that during the investigation we did not damage the evidence.

It is important that we do not start any program, or turn on the computer. These actions can change numerous significant parameters. A lot of potential evidence is situated in the unused space of the file slack, and that is the space from the end of the database to the end of the cluster in which it is saved. These actions could even destroy the evidence that is situated in the virtual memory (swap file), i.e. in the database that the operative system creates when starting a computer.

Taking into consideration various parameters, the operative computer system being the most significant one, we decide in what manner we shall turn off the computer. In case it is a Unix operative system, the computer can be turned off normally, but if the operative system is Windows, then we must decide between two options. The first option is to turn off the computer by force, i.e. to interrupt the electric circuit. We must take into consideration the fact that some computers can have internal uninterruptible power supply (UPS), and in that case, we must go around that device. This way we wish to preserve the data that are located in the system databases, in the virtual memory, as well as data on the processes that have been taking place in the background. We must be aware that a forceful turning off of a computer leaves the possibility of damaging the database system or perhaps the hard disc, and that in this way we can lose important information. On the other hand, by turning off the computer in a normal manner, we can lose information from the system logs, virtual memory or the background processes. Regardless of the manner in which we turn off the computer, it is very important that the entire process of turning off the computer is clearly and systematically documented.

Computers and discs search

It is strongly suggested not to search through the original information, because we might alter some data that could serve as key evidence, making them useless in court. It is therefore necessary to create an image before we begin any kind of search – a copy of the suspicious media. When doing this it is completely irrelevant whether it is a flash drive, hard drive, MP3 Player, iPod, PDA, cell phone or some other device for storage of data. With special technologies every sector can be copied exactly as it was in the moment of dispossession of the computer. The section Preservation of data describes in detail, how to reach this goal in detail in.

After we make a copy of the data, we must be sure that we have obtained the full data. We must also make sure that any search of data will not change some other data. When we know that the data are authentic we can start the search. That step also includes the returning of deleted databases and/or description of the encrypted ones, when necessary, of course. Each step of the search must be conducted exclusively on the image of the media to exclude the possibility of change or erasure of data from the original.

We must be aware that a large number of databases which might be evidence in crime actions will have to be checked. In addition, deleted databases (if we succeed in obtaining them) will also have to be checked, as well as databases that are hidden, encrypted, protected with a password, temporary or databases in the virtual memory, i.e. all databases in which there might be potential evidence, regardless whether they use applications or an operative system. Many programming tools can help us in such searches, especially if the sought evidences are in some type of textual databases (e-mails, MS Word document, etc.). In such examples, the program tools can efficiently search the media by key words. Unfortunately, in example of picture searches, which are most often the case in criminal actions relating to child pornography, the tools cannot be of much assistance. In that case, it is necessary to secure a sufficient number of persons who will manually search through the suspicious databases.

We must give additional attention to databases whose extension has been altered. The only sensible reason why an average computer user would alter an extension of a database is to hide a database. Modern programming tools have an in-built function to search for such databases.

Analyses

By analyzing the obtained database, the investigator can see which documents «live» on the drive, as well as the content of certain databases. The investigator can see the exact time when the databases were accessed (both textual and pictorial), which databases are available and how the databases system saves information. It is very important that the forensic specialist describes and documents all the actions and procedures that have been done, writes down all the suspicious databases and computer applications, as well as all the relevant information that has been obtained by the described procedures.

It is necessary to write down a list of key words that have been used in the search, as well as a list of objects that are likely to contain evidence. Depending on the course of the investigation, it is possible that in the same court preceding a photograph is needed at one point, and a textual document in another. In such cases, these lists significantly facilitate a repeated search.

We must pay special attention when analyzing information situated in the disc areas that have not been located. These areas are the ones that are currently not being used by the

5

WICT/III - 1569225051 - 2509 © SoftCOM 2009

operative system, but can contain data which might be used as relevant evidence. Since these parts of the disc are not seen through the usual computer usage, i.e. the operative system, they are very suitable for hiding data. As has already been mentioned, the potential evidence can be situated in the database file slack, as well as in the virtual memory (swap file), therefore it is necessary to check these areas with great attention. A very significant source of information can be the so-called hidden information like the meta data of certain documents that are being created and added at every opening of the document.

There are many programming tools that can assist us in the search and the analysis of the obtained data. Currently there are two that integrate all the mentioned procedures. These are the Forensic Toolkit (AccessData company) and the EnCase Forensic (Guidance Software company).

Presentation

After the completed analysis, the forensic specialists make their reports which are then presented to the party; natural person or the court. This is the last thing they have to do in order to conclude a given cycle of computer forensics. The investigators must recognize the significance of evidence that is applicable in court. Also, they must present only authentic evidence, which means, it must exist also on the original media, regardless of the manner in which it was obtained. Due to this fact, the report must clearly state the procedure of obtaining a certain piece of evidence.

The evidence must be relevant for a certain case, which means that is must assist the court in making a court decision. After the court accepts the evidence as relevant, the defense of the accused tries to prove that it is invalid in a manner that it tries to disqualify its authenticity, that is to say, it tries to prove that there was an illegal obtaining of evidence during the procedure. This is why the mentioned documenting is so important in all procedures.

The documentation must contain:

• the status of the suspicious computer at the moment of its dispossession,

• technical data on the computer,

• information on databases (number of databases, size, type, ...),

• methods and the programming tools used in the making of an image of the suspicious media and

• procedures, methods and the programming equipment used in the search and the analysis of evidence.

The report made for the court procedure must contain all the data. It must be pointed out that it is not necessary to make a detailed report for every case, but it is necessary that the report contains a clear explanation of evidence that is significant for a certain party. If an entrepreneur is checking

the computer usage of an employee, the proofs of authenticity are not necessary, because this is not a court proceeding and there is no lawyer who will try to eliminate evidence. A detailed report is also not necessary when parents are checking how their children are using the computer.

In any case, the report must clearly describe everything that is necessary for a certain case for which the analysis was conducted. If there is a possibility that in the procedure someone might try to eliminate evidence, then it is definitely necessary to prove the authenticity of such evidence.

Preservation of data

One of the goals that we must achieve is to preserve data. It must NOT happen that any piece of information from the original media be altered or ruined, that is to say, destroyed. It is necessary to make such a copy of the suspicious media that is identical to the original. We call such a copy an image. After we make this copy, we must be sure that we have obtained all the data, just as we must be sure that by implementing some kind of search we shall not make alteration of other data.

There are several ways to make an image of the media. There is specialized mechanical and programming equipment for their creation. If we wish to make an image that will be identical to the original to the last bit, at the very beginning of its creation we must render impossible any kind of writing on the original media. This is performed with the so-called write blockers. Today there are numerous programming tools that already have an built-in write-blockers function, but it is much faster and simpler to make an image by mechanical equipment. Sometimes it is impossible to come to the suspicious media and sometimes it might result in very high expenditures (e.g. hard disc in a server that must not be disconnected), and in such cases we are left only with the option of making the image by mechanical equipment.

After the write blockers function has been enabled on the original media, it is necessary to decide whether we wish to make an image of the entire disc or only of a certain part. When it is a crime investigation, it is recommended to create an image of the entire disc, because during the investigation a need for additional searches might arise and we would most probably lose evidence that is situated in the part of the disc that we have not copied. We can make an image of only a certain part of the media when we are sure that we know exactly where the databases that are evidence material are located or if the investigation of the media serves solely to discover other media which hold key evidence.

The last decision we must make when creating images is to decide whether we can perform a compression of data (images) or not. In many cases due to an enormous amount and size of images, we decide to compress data. When doing this, we must be extremely careful because we might lose the deleted databases. A compressed image of the existing database will still be identical to the original, but will not be

6

WICT/III - 1569225051 - 2509 © SoftCOM 2009

copied every time in exactly the same shape as it was on the computer at the moment of dispossession. This is also the reason why deleted databases that could be evidence material in some further computer searches are lost.

Today in the group of programming tools the most widely spread and the safest is the Forensic Toolkit (AccessData company) and the EnCase Forensic (Guidance Software company). Logicube is the leading company in mechanical equipment.

Authenticity

The second goal that we must ensure is authenticity that tells how one obtains and analyses information in order to prove a conclusion. The image created on a media that previously did not use the function of write block, evidence that is not identically copied from the original, searches that alter other pieces of evidence, all of these are events that must not occur. If something of the above does happen, then evidence will not be used in court because it is not authentic.

In order to avoid such problems, that is to say, to preserve the authenticity of evidence material, cryptographic hash functions are used to make the so-called fingerprints of certain databases or the entire disc. At the entrance the hash function has a symbolic array of variable sizes, which as a result, give a symbolic array of a fixed size and this is the fingerprint of the entry array (clusters, databases, disc or any other array). In that way, computer forensic specialists can create a unique relation between the data on the original media and the data on the image of the media, and in that manner, confirm the authenticity of the searched database or evidence material.

3. CONCLUSION

The proposed set of methods that every forensic specialist should follow when conducting computer forensics, guarantees the authenticity and validity of obtained evidence. If we wish to use the evidence in some court proceeding, we must strictly adhere to all the proposed procedures, because this is the only way we can be sure that the court will not dismiss them as invalid.

In case we do not intend to use the obtained evidence in court proceedings, it is not necessary to adhere strictly to all the procedures. For example, the review of the discovered state can be left out, e.g. in cases of control of the computer use (when parents check how their children use the Internet, or in case of monitoring of the work of the employees in a company). In such cases it is not necessary to take care of the preservation of data or their authenticity or presentation. Finally, let us point out once again that for obtaining data for court proceedings, it is necessary to adhere to all the described steps and objectives.

REFERENCES

[1] W.H.Allen: "Computer Forensics", IEEE Security & Privacy, IEEE Computer Society July/August 2005, str. 59-62.[2] H. Berghel: "The discipline of Internet forensics" Communications of the ACM, Association for Computing Machinery, št.46, zv.8 2003, str. 15-20.[3] J. M. Patzakis: "Computer forensics-from cottage industry to standard practice", Information System Control Journal, Inf.Syst.Audit. & Control Assoc, št.2, 2001.[4] Chris Prosise and Kevin Mandia, Incident Response: Investigating Computer Crime,McGrawHill Osborne Media, 2001.[5] Mark Reith, Clint Carr, and Gregg Gunsch, An Examination of Digital Forensics Models,International Journal of Digital Evidence, Fall 2002.[6] I. Takahashi: "Legal system and computer forensics business", 2004 International Symposium on Aplications and the Internet Workshops, IEEE Computer Society, jan. 2004.[7] Y.Wang, J.Cannady, J.Rosenbluth: "Fundations of computer forensics: A technology for the fight against computer crime", Computer Law & Security Report, Elsevier Ltd. 2005, št.21, str. 119-127.

Other:

[8] Electronic Crime Scene Investigation: A Guide for First Responders. Available at:http://www.ncjrs.org, July 2001.[9] http://conventions.coe.int/Treaty/Commun/ChercheSig.asp?NT=185&CM=8&DF=3/8/2009&CL=ENG

7

WICT/III - 1569225051 - 2509 © SoftCOM 2009

Recognition of damaged characters – Monte Carlo method Željko Deljac,dipl.ing, HT-Croatian telecommunications, Zagreb Robert Moštak,dipl.ing, HT-Croatian telecommunications, Zagreb

Predrag Brođanac, V. High School, Zagreb Contacts: 098/336-526; e-mail: zeljko.deljac@t.ht.hr, robert.mostak@t.ht.hr, predrag.brodjanac@zg.t-com.hr

Abstract: The aim of this work is to research the possibilities of increasing the percentage of correctly identified characters. Matter for analyzing are those characters which, because of too much damage, could not be recognized in the process of recognition by the OCR software. The basic idea of this method is to insert accidentally positioned segments into the image of damaged character and restart process of recognition on updated picture. This process is repeated a given number of times and program notes the result of any recognition. For the purposes of research it was developed special software that automates the entire process and measure the effectiveness of methods, and provides statistics on the percentage of identified characters.

1. MONTE CARLO METHOD - USING IN THIS WORK

Monte Carlo method uses computer algorithms to generate

random values with the aim of solving a variety of experimental and practical problems from different research areas. This method is use in many scientific disciplines such as mathematical integration of multidimensional integrals, physics, chemistry, biology, economics, geography ... In this paper, this method is used to generate two-dimensional field of random variables. Sample image that we are processing is located in the two-dimensional array. Monte Carlo method helped us in this case to investigate two things. First, we investigated the impact of noise on the accuracy of recognition for given character. The noise generator was a source of pseudo randomly positioned pixels in the image of character, and we investigated the sensitivity of recognition depending on the amount of noise. Second, in the second part of the work Monte Carlo method was used for determining the position of accidentally inserted segment with the aim of increasing the percentage of correctly identified characters. Probabilistic field can have many different distributions depending on the applied function of density distributions, and in this paper we have used two of them, uniform distribution and Gaussian distribution. The pictures below show an example of uniform and Gaussian distributions, applied to the X axes, Y axis or both axis. Figure 1a) shows the Normal-Normal distribution, 1B) Normal-Gaussian, 1.c) Gaussian-Normal, 1.d) Gaussian-Gaussian.

Figure 1a Normal-Normal distribution, 1b) Normal-Gaussian, 1.c) Gaussian-Normal, 1.d) Gaussian-Gaussian

2. ABOUT THE CLASSIFICATION BY USING NEURAL NETWORKS

Neural networks have many fields of use, and one of these

is the classification. We could say, the process of recognition of characters is a certain kind of classification, where input samples, in this case the images of letters, we classify into the pre-defined class, in this case joining the ASCII codes as output data recognition. Reliability of classification is measured by size of the qualification degree. Another term that is often used is the "Sensitivity analysis", in which we want to calculate the impact of individual variations of input parameters to change the output value. Equations that we use here are based on the first derivative of each input variable and the gradient refers to the output impact. Usually this is a more variable entry, and analysis found application in various areas such as calculating the impact on air pollution, water pollution, effects of various shares income funds, etc. In this paper we are focused on the application of sensitivity measurement, combined with the application of Monte Carlo method. Two-dimensional probabilistic field are used as the input variable, whose values are generated using Monte Carlo method. We examine the level of accuracy depending on the quality of input images, which represents the character of specified letters. So we vary the input character in a way that we add the random noise in the form of black dots, and it is constantly subjected to the process of recognition. Process of adding dots goes by the principle of Monte Carlo, where the probability of placing dots has a uniform distribution by both image axis. Figure 2 shows what happens with the image of a character in which is added a) 15 additional pixels, b) 50 additional pixels c) 70 additional pixels.

WICT/III- 1569225307- 2509 © SoftCOM 2009

Figure 2a) 15 additional pixels, 2b) 50 additional pixels 2c)

70 additional pixels

The principle is that we add one mark on clean characters image, then try to identify, after that add another, then again try to identify and so 25 times step by step. Of course in a certain moment the images will be so changed that character will no longer be recognized and the probability of recognition after it falls to zero, in the example below it is exactly 7 additional dots. If we display it in the graph where in X axis put the number of additional pixels and in Y axis the probability of classification arises diagram which is shown on figure 3.

0

0,2

0,4

0,6

0,8

1

0 5 10 15 20 25

number of added pixels

reco

gniti

on p

roba

bilit

y (p

)

Figure 3

The procedure can be repeated m times so that every time

we have a different schedule and different order of adding a dot, and on the base of the results we calculate the distribution, median, and get the graph shown on the figure 4.

0

0,2

0,4

0,6

0,8

1

0 5 10 15 20 25

number of added pixels

reco

gniti

on p

roba

bilit

y (p

)

Figure 4

From the obtained data we can calculate the mean of the Gaussian distributions and the maximum deviation for each x. This shows the next figure.

0

0,2

0,4

0,6

0,8

1

1,2

0 5 10 15 20 25

number of added pixels

reco

gniti

on p

roba

bilit

y (p

)

Figure 5

Now we can identify certain parts of the graph and we can

identify the area of high probability of recognition, the area of large risk of recognition and the area of low probability of recognition, which is shown on figure 6.

Figure 6

Next step is to create a similar procedure in the opposite

direction, with removing of black pixels from the image. This is shown as the x values go to the left (negative area of the X axis). The higher negative value the higher the number of deleted black pixels. In other words black pixels are transformed into white color. The following pictures show samples with deleting of pixels, figure 7.a) 20 deleted pixels, images 7.b) 60 deleted pixels.

Figure 7.a) 20 deleted pixels, 7.b) 60 deleted pixels

WICT/III- 1569225307- 2509 © SoftCOM 2009

Again (similar as in Figure 4) we continued with recognition after each deleted pixels, the whole procedure consists of repeated m-cycle, calculated mean values for all negative x values and figure 8 is showing drawn graph.

0

0,2

0,4

0,6

0,8

1

-60 -55 -50 -45 -40 -35 -30 -25 -20 -15 -10 -5 0

number of removed pixels

reco

gniti

on p

roba

bilit

y (p

)

Figure 8

We connected figure 4 and figure 8 together and we got

Figure 9, which actually represents a sensitivity analysis of recognition letters "k", with one input variable (number of deleted or added pixels). Negative values represent deleting pixels, positive values represent adding pixels.

0

0,2

0,4

0,6

0,8

1

-90 -85 -80 -75 -70 -65 -60 -55 -50 -45 -40 -35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35

number of added/removed pixels

reco

gniti

on p

roba

bilit

y (p

)

Figure 9

The same procedure can be repeated for some other images

of characters, eg the character "V". Figure 10 shows analysis of character "V".

0

0,2

0,4

0,6

0,8

1

-90 -85 -80 -75 -70 -65 -60 -55 -50 -45 -40 -35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35

number of added/removed pixels

reco

gniti

on p

roba

bilit

y (p

)

Figure 10

It is evident that the area of high values of recognition is

different for these characters. We can conclude that for some characters, in this case the letter "k", has greater area of recognition of some other letters, in this case the letter "V". This means that the letter "k" is more resistant to all kinds of noise and interference in the process of recognition. On the other hand, characters with low recognition area are less resistant to noise and this often leads to wrong character

recognition. Respecting this, we could define the coefficient of reliability of recognition cr which would be the ratio of the length of recognition area (mediaa-medianr) and the total number of pixels.

p

rar n

median-medianc =

We can see the calculated results of the reliability factors of recognition and for the rest of the letters, shown in Table 1.

Table 1 - Reliability factors

3. USING MONTE CARLO METHOD IN INCREASING THE PERCENTAGE OF RECOGNIZED

CHARACTERS

As already mentioned at the beginning the goal of this

work is increasing the percentage of characters recognition. Further work in this paper does not consider the characters which are in the area of high recognition over 95%, but the signs whose recognition are fewer than 95%. In fact further work deals only with signs that were not properly recognized. These are the signs which due to the large degree of damage can not be recognized by the OCR. Further considerations are focused on the attempt to find methods that will ensure that the samples lost in the process of recognition somehow get out of the dead. Idea is the implementation of the Monte Carlo method, for inserting

WICT/III- 1569225307- 2509 © SoftCOM 2009

random segments, and the presumption is that the updated image has a chance to be recognized.

4. SET OF DAMAGED CHARACTERS

As previously mentioned, the objective of this work was to increase the recognition rate of damaged characters. For the purpose of testing it is needed to generate a set of damaged characters (Step No. 1). The task was made by particular supporting application. Input data for the application were pictures of original characters. The process of damage included the following: adding a noise, adding randomly positioned vertical line and adding randomly positioned spots in the picture. The goal of this process was to simulate damages that occur in the real process of scanning images. Next pictures show a few so obtained test characters. Comparison provides images of original characters and images of damaged characters. Images have a x and y resolution of 96 pixels / inch and their dimensions are about 45 x 55 pixels. It is evident that in the process of damaging characters got smaller or larger damage, which was expected because the process of damaging is a random process.

Figure 11. Examples of damaged characters

The next step (step 2.) was an attempt to recognize test characters. The application that performs the recognition has built-in tools for recognition, it is the free optical character recognition engine Tesseract 2.0. All test characters were subjected to the process of recognition, the result was that some characters recognized correctly, others were not recognized (or recognized as a completely different characters), while the third were not even recognized as characters. It was made filtering and for further processing took only characters from the third group. The process continue with only those characters which in the process of recognition does not recognize as legal characters, ie those whose ASCII code are in the range between the listed below:

(ASC(C) <> " " Or (Asc(C) > 49 And Asc(C) < 116)) Any given input samples are inserted in the application

and the application attempted to recognize characters without the usage of Monte Carlo method. Application stores the results of recognition and adds visual indicator. The indicator changes colour depending on whether the character is recognized or not. The red colour indicates that the sign is not properly recognized, while the green colour indicates the sign is recognized

Figure 12. Without the usage of MC method

Next step (step 3) is the application of Monte Carlo

method. Input data are signs that are not been recognized at all as legal characters. Method works as follow, segment 3x3 pixels (black colour square) was inserted in the picture. Segment position is selected randomly in the area of the image where is the character.

y_sl = CInt(Int(((bm.Height - 4) * Rnd()) + 1)) x_sl = CInt(Int(((bm.Width - 4) * Rnd()) + 1)) That randomly inserted segment complements the image.

The image had already turned into a black&white. This image goes back into the process of recognition, and output data is the ASCII character code. This process is repeated 20 times for each character. Why we decided to use only 20 iterations? On the one hand, the number 20 was chosen with a purpose to ensure a representative sample, on the other side it shouldn’t have been too large that they would not increase the duration of the process. Since this is a random input, and output data depends on a case by case basis, the final data for recognition is the output which appears the most times. All test characters are subjected to the procedure of recognition again, but this time with using of the Monte Carlo method. Now we can see the effectiveness of methods for the cycle of recognition. Characters that have been identified are painted in green:

WICT/III- 1569225307- 2509 © SoftCOM 2009

Figure 13. Result of the usage of MC method

Repeating the method with large number of cycles, we

reach the average efficiency of the amount of 57.66% for the selected set of test characters (mean = 0.5766, standard deviation = 0.0476).

It is also evident that some characters with significant damage were not recognized, while others which were less damaged were identified in 100% of cases. Of course all of this depends on the selected recognition tool.

0,00%10,00%20,00%30,00%40,00%50,00%60,00%70,00%80,00%90,00%

100,00%

Q h b 9 D 8 P 6 k 4 3 2 S 1 C 5 Y W V M E Z T N I R L G F 7

Figure 14. Result of recognition by signs

The final result is shown in the picture below (mean = 0.5766, standard deviation = 0.0476).

05

10152025303540

40,00

%

43,33

%

46,67

%

50,00

%

53,33

%

56,67

%

60,00

%

63,33

%

66,67

%

70,00

%

73,33

%

Figure 15. Gauss distribution of recognition cycles

Therefore, all damages on the samples, which have been

appeared in the process of sampling, lead to decreasing of percentage of successfully characters recognition. However, as can be seen from the previously mentioned, the use of Monte Carlo methods for sample supplementation may increase the percentage of recognition. Increasing percentage of recognition can be summarized in the diagram as an

expansion of the curve in recognizing high-risk area of recognition, as shown in the following diagram, Figure 16.

Figure 16. Extended recognition area

5. CONCLUSION

The subject of research in this work were damaged characters, which could not be identified by OCR program, 100% of test samples were not recognized in ordinary way. Introducing method of Monte Carlo percentage of recognition accuracy of these characters were 57.66%, which represents significant progress. Which gain does usage of Monte Carlo method bring? Naturally, increases the speed of recognition. Increasing of the speed comes from the fact that we took fewer points into consideration. We used only 20 positions, instead of 45 * 55 = 2475 which means that results arrived about 120 times more quickly. This method is not applicable in applications which require 100% recognition accuracy. It can be used in those applications in which we want to raise the accuracy of reading. These are the cases where the damage occurred in the process of reading character. The application can also be found in shape recognition of a very large number of samples. In those cases, this process can increase sigma accuracy and reduce the standard deviation and the number of defects.

BIBLIOGRAPHY

[1] Steve McConnell 1995: Kod iznutra. [2] Christopher M. Bishop 2005: Neural Networks for Pattern

Recognition [3] S.N. Srihari and S.W. Hull 1995: Character Recognition [4] John Clark Craig and Jeff Webb 1997: Visual Basic [5] Alexander J. Faaborg: 2002: Using Neural Networks to

Create an Adaptive Character Recognition System [6] Computerworld : http://www.computerworld.com [7] Vision Research and Image Science Laboratory Technion -

Israel Institute of Technology, Haifa : http://visl.technion.ac.il [8] Thearon Willis and Bryan Newsome: Visual Basic 2008

WICT/III- 1569225307- 2509 © SoftCOM 2009

New approach to designing a Zero-overshoot Automatic Level Control for high-power amplifiers

Bertrand Gerfault Thales communications R&D

Cholet, France

Balwant Godara ISEP

Paris, France

Frank Chahbazian GERAC R&D

Le Barp, France

Abstract—This article describes a new approach in designing a zero overshoot ALC (Automatic Level Control) for power amplifier using a combination of forward and feedback loops. In previously designed amplifier system the ALC is always feedback loop, and the input signal is narrow bandwidth and well known. This article presents a ALC algorithm that can cope with wide bandwidth unknown signal and a new approach for testing the ALC behavior. This approach is based on a customized simulator. The simulation is done in the time and frequency domain as well as in the analogue and digital domain. This simulator allows to test all the behaviors of the ALC with a simple user interface. This Wide bandwidth feedback and feed-forward ALC will provide a new capability for the next generation of amplifier system able to cope with up-coming generations of communication protocols and modulations such as cognitive radio, wireless ad-hoc and ultra wide band communications…

Index Terms — ALC, Automatic Level Control, Power amplifiers, Zero Overshoot.

I. INTRODUCTION

The principle of the ALC (Figure 1) is to control the output level and ensure that the output will remain constant in any input conditions. This is done for several reasons. The most important one is to protect the system that is placed after the amplifier, for example a Transmit/Receive switch (T/R switch) which is very sensitive to overstress.

The aim of the study is to design an automatic level control loop that can avoid any overshoot: no overshoot means no over stress for the T/R switch.

The major contributions of the article are to propose solutions to three problems that have not been solved in the previously designed ALC: the output should remain constant in any condition of an unknown input signal, by opposition of ALC for GSM application only[1], the simulation of the overall system will be simulated, by apposition to simulate partially. The detector and the ALC will allow to cope with wide bandwidth signal.

The output level must remain constant for any condition of the input signal. It has to be considered that the input signal

can be as wide as the input bandwidth of the amplifier and can change level or frequency as fast as the highest frequency that this amplifier can accept. This is a brand new approach compared to the systems that have been previously designed. In previous systems, even if the bandwidth of the amplifier could be wide (20 MHz to 3000MHz), the instantaneous input signal remained narrow, only a few MHz.

Figure 1: System block diagram

When designing an ALC, the first decision concerns the architecture. The most common ALC is a feedback loop. The feedback loop principle is simple and can work perfectly in some conditions. A feedback loop measures the output signal and compares it to a reference, and then controls the input signal according to the error. If the error is positive, the ALC decreases the input, and if the error is negative the ALC increases the input. This ALC architecture has an important drawback because the error is calculated using the output signal. The responding will always be delayed to apply the correction, so this principle will never be well protected from overshoot.

It is thus necessary to design a system avoiding any overshoot. There is only one method to achieve this goal which is to detect the input. If the input signal changes, the input will be corrected immediately. This loop called feed

WICT/III- 1569227687- 2509 © SoftCOM 2009

forward loop has also a drawback, the accuracy of the output is very poor.

To avoid overshoot and improve performance, it has been decided to combine both loops. To ensure that both loops will not try to negate each other and make the system unstable, the overall system was simulated.

The top-level design of RF systems involves the simulation of the interaction between different blocks (attenuators, amplifiers, delay lines, couplers, antennas, etc.). Each of these blocks can be approached by a transfer function in the frequency domain or in the time domain, or by a transfer function in both domains simultaneously.

Radio transceivers are built of digital and analog blocks. Digital sub-systems can be used in a closed loop in order to design more complex control functions.

To evaluate the system, a simulator was designed using wavelet transform [2]. The aim of this article is to propose a new method for ALC and use a simulator to perform exhaustive tests of the ALC behavior.

In the section II, a conventional RF ALC system will be introduced as a state of the art. Section III presents the new approach. Section IV presents the simulation built by using wavelet transform simulator.

II. STATE OF THE ART

A. FeedBack ALC

Figure 2: Feedback ALC

The feedback ALC can be represented by Figure 2. Its transfer function is give by:

)(.)().().(1

)().().()( sref

sAsEsLsAsEsL

sS+

= (1)

The most common way to do a loop control is to use the PID controller, which means Proportional, Integral and Derivative.

sserrKds

serrKiserrKpsc ).(.

)(.)(.)( ++= (2)

But the error is always calculated from the output signal so if the input signal changes fast, the output will change before a new error can be calculated. Whatever the speed of the loop, the output will always change first, causing an overshoot.

A Mathlab Simulink simulation was done to emphasize the result of the PID controller. On the Figure 3 the result of the feedback loop can be observed. When the input (blue curve) changes suddenly the output (red curve) overshoots before stabilizing to the desired output value.

Figure 3: ALC Feedback simulation result

B. Simulation

Time-domain or frequency-domain. Most of the simulators in the market (Simulink, Spice,

ADS, System view etc) are dedicated or optimized to simulate in the time domain or in the frequency domain. They are also dedicated to the analogue domain or Digital domain. And all these simulators will not be appropriate or optimized for to mixed all these domains.

Because it is very important to simulate the complete

behavior of the system before starting to implement the final system, it has been decided to design a simulator dedicated to this application able to simulate in time, frequency, digital and analogue domains.

III. NEW ALC WITH FEEDBACK & FEEDFORWARD

The input signal can change very fast, so the only way to

prevent this change is to detect it at the input, and then apply the correction before the signal is transmitted to the amplifier. The type of loop is called a feed-forward loop.

A simple feed-forward loop has an important drawback; the accuracy of the amplifier system depends on the accuracy of the gain of the amplifier because the gain accuracy for a wide band amplifier can vary a lot across the frequency range.

Another possibility is to use both loops (Figure 4), one before the amplifier using the feed forward loop and one

WICT/III- 1569227687- 2509 © SoftCOM 2009

controlling the amplifier. This seems to be the solution but if both loops are independent they will compensate each other and the system will become unstable.

The solution is to have a system with both loops, but working together. To ensure the stability, the feed forward loop must be as fast as possible and the feedback loop slower than the feed forward loop response time.

Figure 4: combined feed forward and feedback loop

In the main part of the system a delay [D(t)] is introduced to ensure that the feed forward loop will be faster than the feedback loop.

Then the PID is also used to control the loop; but the

main difference is the way it is used. The feedback loop uses the Integrate of the output for the accuracy and the feed forward loop used the derivative of the input for the speed.

ssrKds

serrKiserrKpsc ).(.

)(.)(.)( ++= (3)

Using this principle, if the signal rises very fast, the Kd

parameter of the equation (3) will bring a very fast compensation in the loop for a short period of time, giving the time for the feedback loop to compensate (Figure 5).

Figure 5: Combined loop result

IV. SIMULATION OF THE NEW ALC

A. Simulator description

This simulation method has to be able to answer two challenges. The first one is the simulation of a system made up of blocks with transfer functions in different space (time, frequency and time-frequency). Whereas systems with only transfer function in time space or in frequency space can be theoretically solved, complex heterogeneous systems have to be solved numerically. The second challenge is the simulation of system made up of Digital and Analogue sub-systems. With numerical simulation, this can be reduced as the simulation of a system made up of two sub-systems with different sampling frequency.

This approach to these two problems is to describe the

input signal in a time-frequency space and to solve the overall system in an incremental way.

B. Simulation of Transfer function

The simulation of the system shown of the Figure 6 has

been performed. Some blocks were characterized in the time domain (Delay, VVA and detector) some in the frequency domain (Amplifier, directional combiner and Antenna), some in both domains (Input signal) some in the digital Domain (ALC) with a slow sampling frequency and the other in the analogue domain with a very high sampling frequency.

Figure 6: ALC simulation

C. General Specifications

The purpose of this ALC is to control an amplifier system

with the following main specifications. These values are typical specifications that a modern communication amplifier will have to stand for.

• Wide bandwidth frequency range for 20MHz up to 3

GHz.

• Nominal power of the input signal 0dBm, with crest factor up to 10dB.

• Instantaneous bandwidth of the input signal up to 100MHz

• High Gain of the high-power Power Amplifier.

• Ripple in the PA’s response as large as +/-3dB

WICT/III- 1569227687- 2509 © SoftCOM 2009

• Phase delay of the high-power Power Amplifier as large as 200ns.

• VSWR of the antenna as high as 3:1

Using the simulator, the Input Signal can be programmed to generate a variety of signals in the time and frequency domain, to reproduce all types of signals that can be used with the system. The more capability is given to the generator the more test of the behavior of the ALC can be performed.

Then the signal is split in two parts; one will be called the main part and one called the ALC part.

The main part is composed of the main elements of the amplifier: the VVA (Voltage Variable Attenuator), the amplifier, the delay, the directional coupler and the antenna. The delay is used to emulate the amplifier delay. For the simulation purpose, the delay has been separated from the amplifier. The directional coupler is used to measure the forward and reflect the signal from the antenna and calculate the VSWR. The antenna VSWR can be reproduced to perform a more realistic simulation.

The ALC patch is composed of the input and output detector, ADC (Analog to Digital Converter), ALC block and DAC (Digital to Analog Converter)

The ALC block (digital system on Figure 7) received the signal from the input and both signals from the output. It also gets some control signal. Then, this block performs the ALC algorithm and generates the error signal to be sent to the VVA.

The ALC block uses a forward and feedback algorithm to provide a zero overshoot signal at the amplifier output. All parameters of the algorithm have been optimized during the simulation and most of the amplifier system has been simulated to verify the ALC behavior. The main behaviors to be tested are:

• Verification of the accuracy of the output • Verification of the reaction time • Verification of the overshoot

The following variations are added to the signal to determine the system’s response to them:

• Gain flatness versus frequency • Amplifier delay • Detector linearity • Antenna VSWR • Directional combiner gain flatness

And with all the input conditions such as:

• Single tone and multiple tones • Instantaneous bandwidth • Level variation

• Crest factor All these behaviors can be easily tested with this simulator and all ALC parameters can be adjusted quickly. This simulation improves the amplifier design lead-time and ensures that the amplifier meets the entire system requirement. .

Figure 7: ALC Simulator main window

The Figure 7 is the main window of the simulator. It shows all the blocks. For each one, all the parameters can be varied.

D. ALC simulation result To verify that the ALC works correctly, a lot of simulations were made. Below are the most significant results for the most difficult problem that ALC had to cope with:

Figure 8: Input level window for programming any signal waveform

Example 1: Two tones separated in the time domain

The signal is composed of two continuous wave (CW)

signals separated in the time domain: one starts at the beginning of the simulation at 200MHz with a level of 0dBm, then the second signal starts immediately after the first one has stopped. The second one has a frequency of 300MHz

WICT/III- 1569227687- 2509 © SoftCOM 2009

with 3dBm of level. It can be seen on Figure 9 that the signal is perfectly controlled and no overshoot is generated.

The simulation result is composed of three figures: • One in the frequency domain at a time delay of 101us • One in the time domain at the frequency of 300MHz • And one in both time and frequency domains, (this is

done using the wavelet transform of the signal).

Figure 9: simulation result of 2 signals at 200MHz and 300MHz

Example 2: Two tones un-separated in time The second example is a two-tone signal that appears at the middle of the simulation. From 100us to 200us, two signals are present at the input.

Figure 10: 2 tones simulations

In this example is can also be observed on Figure 10 that

the signal is perfectly controlled, no overshoot is generated. Also because the regulation is based on a peak detection and

regulation, even with the two tones, the peak is controlled not to overshoot the 20dBm.

To be able to have a signal regulated on the peak value

for such a bandwidth, a very high bandwidth peak detector must be used [3] and also an ALC that will calculate the peak of the signal and regulate on the peak. This ALC must have a detection bandwidth compatible with the instantaneous bandwidth of the input signal.

V. CONCLUSIONS

The ALC algorithm proposed in this article ensures that the output signal will never exceed the targeted reference value. A state of the art feedback loop with log detector will generate an output overshoot for two reasons, the signal can’t be detected fast enough and the differences in frequency of the two tone signals will be larger than the video bandwidth. Only this ALC with both Forward and Feed Back loop with peak detector will achieve this level of performance. It is also very important to choose the right detector which can provide enough bandwidth. The overall system is able to provide a regulation on the peak of the signal without overshoot for the entire instantaneous bandwidth of the input signal.

Using this customized simulator by opposition to use an off the shelf simulator, it has been possible to make an in-depth test of the ALC within the entire input configuration. During the entire simulation, the parameters have been adapted to have the best speed without any compromise on the overshoot. This ALC algorithm has been implemented in the latest state of the art amplifier system and all measurements that have been made confirm this simulation.

In the future design, the knowledge of the ALC behavior and the simulator will provide a serious advantage for designing the next generation of amplifier system, compatible with the new generation of wide bandwidth communications signals.

ACKNOWLEDGMENT

The work is supported by Thales Communications France, GERAC, and the ISEP Institute Supérieur d’Electronique de Paris.

REFERENCES

[1] D. Ripley, “Power detection and Control for Mobile Handset

applications” RF designline June 09. [2] M. Nau, B. Godera, and B. Gerfault, “Applying Wavelet

Transformation to RF System Modeling,” unpublished. [3] B. Gerfault, B. Godara, “Novel Methodology for Choosing detectors

for the Automatic Level Control of High power amplifiers,” Proceedings of the 3rd international Conference on Anti-counterfeiting, Security, and Identification In communication, Aug 2009 p378-381.

WICT/III- 1569227687- 2509 © SoftCOM 2009

A Design and Implementation of Smart Guide Services with the Semantic Service Discovery

HyunKyung Yoo*, JeongHwan Kim**, SangKi Kim*

Service Convergence Research Team*

Convergence Service Platform Research Department** Electronics and Telecommunications Research Institute (ETRI)

Daejeon, Republic of KOREA E-mail: {hkyoo, ditto, kimsang}@etri.re.kr

Tel: +82-42-860-6768, Fax: +82-42-861-1342

Abstract: In these days, there is a flood of numerous services with various capabilities. So how to find and provide a suitable service among various services is an important issue. To solve this, we present that the service discovery and composition framework which can provide the service to satisfy a user’s request or context. For the service discovery and composition, the services become the components which can be reusable and have a generic interface and the components have the service metadata. Service providers can publish, discover and manage the services using this metadata. In this paper, we propose the service discovery and composition framework and the use case scenarios, that is, exposition smart guide service. Service subscribers can use each service according to their preference in same location and discover whether a wanted service is being or not. Moreover our framework provides the service composition function which combines services dynamically when there is not a suitable service.

1. INTRODUCTION

In these days, there is a flood of numerous services with various capabilities. Context-awareness has become a key technology for the integration of the next generation network (NGN) and pervasive computing[1]. In this circumstance, it is useful to have the service discovery and composition framework which can provide services to satisfy a user’s explicit request or context[2]. For this framework, the services become the components which can be reusable and have a generic interface and the components have the service metadata. And then service providers can publish, discover and manage the services using this metadata.

In this paper, we proposed the service discovery and composition framework and the use case scenarios about the exposition smart guide service. Service subscribers can use each service according to their preference in same location and discover whether a wanted service is being or not. Moreover our framework provided the service composition function which combines services dynamically when there is not a suitable service.

This paper is organized as follows. Section II defines the necessity and architecture of a proposed service. In section III, we propose exposition smart guide service scenario and flow. Finally, in section IV, we bring to conclusions.

2. ARCHITECTURE

Tourism is one of the major sources of income for many countries. Therefore, providing efficient, real-time service for tourists is a crucial competitive asset which needs to be enhanced using major technological advances[3]. And to take advantage of telecom services with various capabilities, it is not reasonable for users to be aware of the published services and subscribe to the wanted services.

By making that the subscriber’s profile and preference are stored in the platform and managed, the platform enables to provide the available services based on user’s context and preference.

Our platform is a de-centralized system that supports description, publication, discovery and brokering components. Components, the basic building blocks for services, are an executable unit independently and can be reused with SW unit.

Our platform consists of the service framework and the recommendation framework. The former is for provisioning services dynamically and the latter is for recommending services suitably.

Figure 1 shows service framework that consists of service annotation function, service publication and discovery function, service brokering function and service metadata registry. Services annotation function generates the metadata of services registered to registry. This function is performed by service developers or service managers.

WICT/III- 1569230467- 2509 © SoftCOM 2009

Figure 1 –The architecture of service framework

Service annotations contain three main aspects: functional metadata, non-functional metadata and semantic annotation[4]. They are completed to the service description documents. In our framework, services are defined with owl(Web Ontology Language) and annotated with owl-s, that is, Semantic Markup for Web Service. Figure 2 show the example of service annotation that includes a service name, contact information, service category, service classification, service product and text description.

Figure 2 –The example of service annotation Service publication and discovery function provides the interface to publish and discover service description documents. In our framework, we used OWL-S UDDI MatchMaker as discovery mechanism. After published owl-s services to the service registry, users can discover wanted services with query. In case of composite services, service brokering function provides the reference point for the users to retrieve components. The service metadata, that is, the means to specify components are stored in service metadata registry. They enable service discovery, selection and composition.

3. SERVICE SCENARIO AND FLOW

The proposed use case scenario is an exposition smart

guide service. Fig.3 shows the network architecture to provide this service. It consists of application server, platform, location server, message server and terminals.

Platform is composed of service registry, SMS/MMS function, location function, recommendation function and service broker. The recommendation function stores the subscriber’s profile and preference and recommends the suitable service to user’s context or request. The service broker performs the service composition to a request and sends the composed service to application server.

A pplication Server

Location server M essage Server

term inal

Profile

location

Preference

SM S

M M S

Service BrokerService R eg istry

M ap

R ecom m ender

Figure 3 –Architecture In this use case, subscribers, Tom and John, work in a

same company and go on a business trip to an unfamiliar place, Ocean Expo. Tom requests the service discovery and the service composition. John receives the composed service which Tom requests.

A service scenario of user’s point of view is as follows.

- Tom and John arrive at the Ocean Expo. - Platform recommends them to the available services

based on users’ context and preference. Tom is provided an exhibition service and John is provided an event service of smart guide services in Ocean Expo. They use their own services.

- While a yacht race, Tom wants to see a final yacht race with his colleague, John. He wonders whether his wanted service is being or not.

service annotation function

service publication & discovery function

service brokering function

Service R eg istry

WICT/III- 1569230467- 2509 © SoftCOM 2009

- Tom requests a service search with a keyword that is to send MMS with the direction map from a current position to the yacht race field.

- Platform tries to discover a requested service considering semantic relations of services. As a result, there is no matched service.

- Platform suggests a service composition. Tom consents the service composition with the service keyword, message contents and phone number of receiver.

- Platform does compose the components, that is, John’s location, direction map and MMS.

- As a service composition result, John receives a MMS with the direction map from his current position to the yacht race field from Tom.

The semantic relations of service used in service discovery,

mean that services have service metadata. So when the SPs or users request a service, platform can discover the suitable service by service keyword, category and service information. And the discovered services have the degree of match for service recommendation or user’s best choice.

Figure 4 shows the service flow that is composed of entities, that is, application server, platform, context generator, Tom and John’s UE.

A S platform

Context G enerator

Tom ’s U E John’s U E

Receive user’s location

Search profile and preference

PushBestService(Expo sm art guid e service)

SearchService(keyw ord: send M M S w ith the direction m ap from a current position to destination)

Search service

Service Com position

Send M M S(direction m ap to yacht race field)

SelectService(Expo sm art guide service)

SelectService(Expo sm art guide service)

PushBestService(exhibition service)

PushBestService (event service)

N oM atchedService()

RequestCom position()

getM ap()

Figure 4 –Service flow A context generator plays a role of the location server that

recognizes a user’s location transition. Platform searches profile and preference in a recommendation framework when receives user’s location. And then, platform pushes best services to users. Tom and John are recommended the suitable services and can set ups the services they select. Tom requests a service search and receives the result that is no matched service. Platform performs the composition of a

location and a MMS in platform and a map in application server.

Service subscribers can use each service according to their preference in same location and discover whether a wanted service is being or not. Moreover service framework provided the service composition function which combines services dynamically when there is not a suitable service.

Figure 5 –Tom’s terminal UI

Figure 5 show a Tom’s terminal UI. The left is the

composition inputs that are service keyword, message and receiver. The right is composition flow that shows the input and output parameters of basic services. Finally a location service, a direction map service and a MMS service are composed and send to John.

4. CONCLUSION

In this paper, we presented the service discovery and composition framework and the use case scenarios about exposition smart guide service. Service subscribers can use each service according to their preference in same location and discover whether a wanted service is being or not. Moreover our framework provided the service composition function which combines services dynamically when there is not a suitable service. For this framework, the services become the components which can be reusable and have a generic interface and the components have the service metadata. And then service providers can publish, discover and manage the services using this metadata.

Semantic service discovery can provide and rank not only keyword matched services but also related services in concept automatically.

With a user’s point of view, it doesn’t need for users to be aware of the published services and subscribe to the wanted services. By making that the subscriber’s profile and preference are stored in the platform and managed, the

WICT/III- 1569230467- 2509 © SoftCOM 2009

platform enables to provide the available services based on user’s context and preference.

In service composition, platform is positioning the user’s location which is not fixed and provides direction map from that location to the destination. This service can be used to context aware tourism services.

ACKNOWLEDGEMENT This research is supported by the IT R&D program of MKE/IITA of South Korea. [2009-F-048-01, Development of Customer Oriented Convergent Service Common Platform Technology based on Network]. REFERENCES [1] Run, Y., Yun, Y., Xudong, M. and Zhengkun, M., “The study on the context-aware application architecture in NGN”, IET Conference Publications, 2006, pp.424. [2] Peng, R., Mi, Z., Wang, L., An OWL-S based adaptive service discovery algorithm for mobile users, 2008 International Conference on Wireless Communications, Networking and Mobile Computing, WiCOM 2008, art. no. 4677964 [3] García-Crespo, A., Chamizo, J., Rivera, I., Mencke, M., Colomo-Palacios, R., Gómez-Berbís, J.M., SPETA: Social pervasive e-Tourism advisor, Telematics and Informatics [4] SPICE Deliverable D1.5: Revised Requirements and Scenarios, 2007.

WICT/III- 1569230467- 2509 © SoftCOM 2009

Providing Data Grid Capability Using Society Network on Mobile Phones

Meisam Hejazinia 1, Mohammad Reza Razzazi 2 1 Amirkabir University of Technology, Dept. of Computer and IT Engineering, Tehran, Iran, hejazinia@aut.ac.ir 2

Amirkabir University of Technology, Dept. of Computer and IT Engineering, Tehran, Iran, razzazi@aut.ac.ir

Abstract: Today mobile phones are the most widespread devices in the world. The numbers of mobile phones today are, four times more than the number of PCs. This provides high potentials for providing some capabilities like data grid on mobile phones. Unfortunately, there is no suitable software system to provide this capability on mobile phones; however, we have enough hardware capabilities. All previous proposals called mobile grid only focused on hybrid mechanism, which at most mobile phones could conduct process, but all scheduling activities were still done on server. The hardware enhancement of mobile phones has become steady, and Moor’s law is not applied to mobile phones anymore due to the limitation in battery power. In this paper, we focus on data grid capabilities, and try to provide the same functionality for mobile phones. Constraints of mobile phones make them unsuitable to port data grid on them, and we should change many mechanism in order to provide this functionality. We consider the non-functional requirements of mobile phones in our proposed system. We define a new concept society network, which is necessary to overcome other constraints of mobile phones. So far, there were not such a system developed on mobile phone, and our idea is quite new to the field.

1. INTRODUCTION Capabilities of mobile phones are the same as personal computers in early 90s. This guides us to the fact that much work could be done to enhance processing over mobile phones. This fact also applies to data delivery. Of course deep research has been conducted on grid computing [1]. No one has yet leveraged that research to enhance mobile phone platforms to provide data and service grid over them. Computational grids represent a new and rapidly evolving research area, which has gained a lot of attention in the past several years. A computational grid can be viewed as a transparent aggregation of many computing devices on a network that enables sharing of distributed resources. Typically, computational grids are considered in the context of sharing processor power of many computers interconnected by a wired network [4]. Mobile equipments are integrated either as mechanisms for interfacing the grid or as active resources for the grid itself. In the case of access mechanisms, every incoming mobile device is integrated in a completely transparent way; that is, the mobile user can access the grid and move within the environment without any manual configuration operation. In particular, the mobile user gets automatic access to services as soon as he enters the environment. In addition, whenever he moves within the environment, his device is automatically located and his context follows him [5]. Although concept of mobile grid has been investigated before [4-12], but this concepts, look at mobile devices as terminal. In mobile grids, mobile devices computational, and memory power is also used, but still all grid activities were conducted still on servers, which are backbone of the grid. In this paper we are looking at mobile phones, as only infrastructure of grid environment, and redesigned mechanisms according to specific requirement of this new environment.

At the first glance, we may think that it would not be possible to have such a system on mobile phones, while most of grid mechanisms, need a high computational power specially for discovery, publishing [8][9][11]. Even replication mechanism [12] of data grid seems impossible to be ported on mobile phones. These all are rooted in two assumptions that we should change when we look at mobile phones. First, we should look at mobile phones as servers, not as a simple lightweight terminal. Second, we should change the goal of data grid for mobile phones. Data grid is provided for scientific data [4]. But on mobile phones, we should think about other data like tacit knowledge, the experiment that everybody has to help others, for example, knowing the best restaurant in London. These are people’s experiences and usually put on forums on the internet but if people don’t have time, what should they do? The solution is using their mobile phones, and it is the purpose of our paper, to provide the needed functionality on mobile phones.

There are many similarities between grid and mobile phones, which we could leverage in order to provide required functionality of data grid on mobile phones. First, grid resources are dispersed and we don’t know where they are [1]. Thus, we need to have discovery services [11]. In grid, data are geographically distributed, and we have the same thing on mobile phones. We have heterogeneity in grid environments, and we have the same thing in terms of operating system, hardware, of course J2ME, solved this problem somehow. But the challenge that is not present in grid is the network heterogeneity. We have it on mobile phones as well, that means that mobile phones are not always connected to the internet like grids, and we should leverage different network infrastructures, like Bluetooth, WLAN, GPRS, and GPS in order to communicate [7]. This characteristic of mobile phone makes this platform more challenging. Although many works have been conducted on mobile grid, but mobile devices in mobile grid, use seamless network, like GPRS to connect to each other, but focus of this paper is on mobile Ad-hoc network (MANET). Collaboration services enable users utilizing mobile devices to share information and communicate with each other. Research has been conducted on various models that support collaboration [4][13]. However most of these models are not scalable and depend on specific technologies, most of them were centralized, and have single point of failure, they are based on GPRS, and this imposed high cost on users, which makes these models unusable. Also they are not configured according to unique specifications of mobile phones, which we will discuss. Additionally these models are not extensible, meaning that they only support designated services. The models depend on an existing wireless infrastructure and operating within the range of a wireless access points. Ad-hoc or infrastructure wireless environments, allow users to make use of these services and not rely on a wireless access point. Reference [13] discusses the design and development of an extensible model to support collaboration services between mobile devices in a wireless

WICT/III-1569231353- 2509 © SoftCOM 2009

environment. However their proposal is only focused on providing communication layer. In this paper we go further and defined detail mechanism of data grids for this environment.

We have the concept of autonomy on grids, that each site could have its own policy [1]. This is the same on mobile phones, while they are personalized; people want to have their own policies according to their mobile’s capabilities, and their own preferences. It means that the same concept of Virtual Organization should be provided on mobile phones. The next point is resource sharing like grids that we wanted to use perishable computing services [1]. We have the same meaning on mobile phones that mobile phones together could provide a huge amount of storage for providing data. The main challenge is that mobile phones should have a mechanism to handle distribution, which is much higher than grids. This distribution in grids is on different servers which have enough capacity around the world but we cannot have such thing on mobile phones, and distribution is dominant, than each capacity. The other characteristic of grids is that they are dynamic; nodes could join, or leave anytime [10]. This characteristic is more severe on mobile phones while they use battery power and wireless network, and this makes this platform highly dynamic. Any node could leave for the purpose of failure, or wireless network out of range, or battery drainage.

Apart from the similarities and differences we have previously mentioned between mobile phones and grids, we have more sever challenges in terms of processing power, and memory capacity. This makes this platform more challenging. Like grids [1] we could not have a single organization centralized control on mobile phones, they are not static. We should have the same thing like QoS concern on mobile phones, and we should let different mobile phones set their own policies.

In data grids we have a middleware which supports authentication, and authorization. We could do something much better on mobile phones by using SIM card’s security. In grids we could check the status of data. Data management service includes metadata, indexing caching, reliable transfer. For information discovery and monitoring we are aware of available resource lists, we have the ability to put constraints on this directory, for relevancy, and the best suit resource, we have a publish status, and a list of available resources, and we have hierarchical information discovery service. We must provide the same functionality on mobile phones in order to have data grid capabilities on them, and this is our main contribution, that we will describe in next sections.

The system we provide on mobile phones should satisfy some non-functional requirements, which are: scalability, like grids mobile phone for the same characteristics should provide it; adaptability, which means the system should be adaptable for different user preferences, and device capabilities. Thirdly, it should be lightweight that means it should be optimized, while we have constraint in terms of mobile phones’ processor and memory capacity.

Remaining part of this paper is organized as follows: Section 2 will discuss related works, on mobile grid. Section 3 discusses scenarios and requirements of data grid capabilities on mobile phones. Section 4 provides a definition of Society network, and explains why it is suitable for mobile platforms. Section 5 discusses our solution for data grid functionality on mobile phones. Moreover, it provides the details how differently the challenges should be overcome. Section 6 discusses the evaluation issue related to our proposed solution, and section 7 concludes the paper with highlighting our contribution.

2. RELATED WORK Mobile grid means that movable wireless devices are integrated into traditional wired grid through wireless channel to share grid resources (CPU power, storage capacity, instrument, devices, data, software, etc), meanwhile, mobile devices can provide service or resource to grid users, such as PDAs, cellular phones, handsets or wearable computers, laptops with GPS service, mobile service, etc [4]. The main drawback of mobile grid is they still use current grid infrastructure of grid, and they should connect to this infrastructure. This means people should pay high cost, due to high cost of internet data exchange on mobile phones. We tried to solve this problem to redefine data grid only on mobile phone, and without basing our infrastructure on current infrastructure, and consequently, having lower cost of communication. Reference [4], present a prototype of a grid-based problem-solving environment for wireless mobile devices with limited processing power Its primary purpose is to allow mobile devices with limited resources to solve problems that they would not be able to solve individually. This goal is achieved by redistributing the computational load among many computing devices. Reference [5], presents a middleware infrastructure able to integrate mobile devices in the grid. As a matter of fact, classic grids do not provide mobile users with support to access resources and services at all (i.e. classic grids consists of wired, pre-configured, powerful stations). Moreover, whenever a user wants to execute her own application, she typically has to: (i) ask the grid for resources; (ii) allocate tasks; (iii) launch and control executions; (iv) get results; and (v) release resources. Reference [6], proposes a MABS platform on the computational Grid for mobile distributed computing. The findings of the experiment show that a Grid based MABS platform can provide a scalable simulation environment for mobile distributed applications. Mobile grid combines mobile computing and grid computing and develops rapidly. Reference [7], provides design of a general architecture of mobile grid, illustration of mobile grid architecture design principle, presentation of a novel architecture, analyzing of the mobile grid constitutes and logical structure. It also discusses mobile grid resource and service management and allocation mechanism from description, discovery, security, QoS and selection and assignment factors, etc.

3. SCENARIO FOR DATA GRID ON MOBILE PHONES

The first question is why we really need data grid on mobile

phones. The aim of data grid was providing transparency, and share resources across the world, with the core concept of virtual organization. That platform includes huge data related to genes, huge financial data related to financial markets, or huge astronomical data. Scientists use these dispersed data through grid.

We do not have such a thing on mobile phones, and they are not even suitable for this kind of job. Mobile phones are portable. Thus, they are very useful, when it is difficult for you to use your laptop, or terminal to your server, when you do not want to bother yourself to go and find a PC to handle your job.

Mobile phones are ideal tools for tacit knowledge sharing. Although people usually have knowledge about something, they do not bother themselves to share it and it is difficult for them to open their laptop to write tacit knowledge they have. People have so much free time during the day, for example when they sit in a bus, or when they are at a beach with their family, or on a journey. This free time could be used to insert their tacit knowledge on their mobile phones.

WICT/III-1569231353- 2509 © SoftCOM 2009

Let us explain ourselves with a scenario: You have a cardiac problem, while your doctor is not aware of that. You know the pain is definitely there. Most of the time, you ask your family members if they know a good doctor. It is the tacit knowledge between people. They have experienced many doctors in their lives but they have never written who they think is better for a specific problem. The idea is that, your mobile phone uses data items related to this area by interacting many other mobile phones automatically while you are doing you ordinary job during the day.

The other scenario could be when you need to use some heuristics to enjoy buying well suit stocks in a stock market. If people have good heuristics on this domain of time, why do not you use them?

The other scenario is when you go to London, in a specific area in the town. You have a meeting and you intend to go to a good restaurant in neighborhood. If other people provide this tacit knowledge, and you consume these data services, the world becomes a heaven. You may want to go mountain climbing in a specific mountain then you can use the tacit knowledge of the people you have never met. In the middleware have we provided, you will be able use it?

You may have asked these questions previously in forums on the internet. However, people usually do not have enough time to check forums. You may be on a trip, and it is difficult for you to open your laptop, and you may see many people, in you society network, you can just leverage this new channel. Additionally, this new ability will provide usability for publishing tacit knowledge.

4. SOCIETY NETWORK

There are four options available for providing a middleware over

mobile phones: Bluetooth, SMS, WLAN, and GPRS. GPRS is based on GSM. In the middle of these two technologies is the improved version of GSM, which cellular operators enhance their systems, to improve to GPRS, and SMS is built over GSM, it is simple, fast highly flexible, scalable, wide spread and user friendly. The specification of each of these infrastructures could be seen in figure 1.

To make decisions over these infrastructures, Bluetooth shows

up, for it is free but has some shortcomings upon its limited range. Bluetooth and WLAN support the concepts of same time and same place. At first sight, we may think that this would not be a suitable infrastructure to conduct data grid services. But when we look thoroughly, we find out that there is another thing, which is society network. Society network means people we meet every day, in buses or bank queues, waiting for our flights. We are not staying in a place day and night, and we are moving, and this is a great opportunity for providing exchange with many people. Although we meet and talk with limited people, our device is not limited like us. It could automatically interact with many people we may not see, but they

are in our proximity, and it could match profiles, use their processing power, or memory. Our device could even trigger and guide us to the right person. With this mentality, Bluetooth is not limited anymore, while we are not Robinson Crusoe who don’t have any people around us in an isolated island, even when we are in an island there are many people around us, and this means a huge opportunity for computing and commerce. The main advantage of Bluetooth is that, we are not controlled by central agents anymore, there is no operator, and we could conduct commerce without the middleman (operators).

After that we have SMS, which is accessible from anywhere, but has a shortcoming in terms of cost and middleman controls and regulations. The third suitable choice is WLAN which is also free and suitable, with the ability to connect to the internet, and the fourth one is GPRS, due to having a high cost of transfer; it makes it difficult for a normal application to communicate.

To overcome the shortcomings of each of these options there is a solution, which is a software layer over them for roaming [6] transparently, so that the application over them would not be notified that the underlying layer has changed the network medium. Thus, the first underlying layer searches for the mobile phones in the proximity to check whether they have the special service. If they do not, the software layer under it does the roaming and goes to WLAN or SMS according to the user preferences between cost and time. In addition, if it is not found, then automatically it connects via GPRS to another mobile phone, or sends an SMS to a default mobile phone that he knows it contains the required services.

5. PROPOSED SOLUTION FOR DATA GRID ON MOBILE

PHONES

The main requirement to provide data grid on mobile phones is publishing service. The main issue is how to publish data on mobile phones, especially on society networks. In Grid we have a distributed system consisting of replica metadata, replica catalog, replica search, replica selection, replica location, replica location index. Grids provide a hierarchical system, and while we have a seamless network, on grid, this is possible, but our proposed society network is not seamless, so such solution does not work. The broader challenge is that we do not even know the topology of the society network.

What we propose is an adaptive request/response method. Each person could request a specific data item. In our proposed system, if it does not receive a response from the approximate mobile phone, the request could become persistent on their mobile phones. If they change their places, and new people locate in their proximity, which means they become members of their society networks. Then, they ask this request, as an agent of the previous mobile phone. Finally, if they get the requested data, they could send that to the person next time he locates in their proximity, or they could save the address of that person and send him, the response through an SMS, or send them the address of that person who owns the requested data.

Our adaptive request/response method keeps track of data items, which is done by the provider of data. In this case, if the provider of data finds out that this mobile phone asks for specific data more than a specific threshold, the replica could be created on that mobile phone according to local policy of the requester. Thus, after a number of times, the optimal place of the replica could be found, and this could be used by the requester. Our approach consists of two main parts, firstly the queue, which preserves the request until it is responded, or the deadline expires. The second part is the self aware data, which put addresses of the requester in their cache, and

Bluetooth WLAN GPRS

Bandwidth 1 Mbps 11Mbps 115-117 kbps

Power 1-10mw 50-70mw 200-800mw

Range 10-100m 100-200m 1KM Cost None Low High

Frequency 2.4GHz 2.4GHz 900/1800MHz Fig. 1: Network Medium Comparison

WICT/III-1569231353- 2509 © SoftCOM 2009

if the requester asks for the same data more than a threshold, it will provide the replica on that node.

This queuing also helps us to have routing information on mobile phones; it means that, each intermediate node that keeps the information requested for some time could keep track of who answers this request. Instead of broadcasting the request in next stages, it will send the request to a specific address. This address could be either intermediate or final.

You may think that this approach only solves the problem of not having a seamless network, what about the lightweight requirement of an application. What we propose is using a mechanism like cache; we could implement Least Recently Used mechanism. Each service could only have a limited cache that could be configured according to a specific mobile phone’s requirement. We will explain this issue, later in this section.

The solution provided previously, was only a post solution, which means it is adaptive, but what we could do for the first time, when we do not have any routing directives, and no replica is the main challenge. What we propose is to first having a hierarchical structure, which could be adapted then. We could have a domain and some sub domains. At first, this network is configured with people’s expertise. For example, when you are a computer engineer, data requests related to the computer domain, are routed to you. People themselves could configure this routing table according to their knowledge. For example, Jack knows his friend Jenny, who is computer engineer, so it configures his routing table accordingly, and so when a request for computer came to his mobile phone, this request will be routed to Jenny’s mobile phone, automatically according to configuration. Consequently, the primary table could be constructed, and then it would be adapted according to our approach. This is the same thing we do in our society. For instance, when our friend has a specific illness, if someone with the same illness wants to know who the best doctor in this field is, we guide them to our friend. Our contribution is that we make this process automatic.

In our adaptive approach, we perform omissions, which mean if more than some time we send someone a request and they do not answer, that next time the probability that we send our request to this person will decrease. Our approach uses probability that is a number between 0 and 1. Moreover, our adaptation is done by simply calculating the number of time we did calculation over the number of time we requested for each mobile phone.

To complete our approach, for lightweight services, which means the content data is only a small number of an SMS volume, the requester, himself, could have a list of popular services. As a result, he could ask for the service through an SMS, which is something like having a phone book for data services.

We have two kinds of identifier in our system. The first one is a UID for each user, and the second one is a DID, which is for data, and both are unique over our platform. They are simply calculated by mapping user/data metadata to one number. We will discuss the shortcoming and solution related to searching data in this section.

The request for specific data roams over a society network. We should have a mechanism to trust the data that we acquire. For this purpose, we propose a rating mechanism that different people rate one person, in a specific domain of knowledge. This could be done first according to the previous knowledge or afterwards when people use data services. This rating mechanism could also have negative effects on the rate of each data provider. As far as according to economical theories, large people could not coalite, this rating could provide a solution for the challenge, since we have a high distribution.

We also define the shortest path optimizer over a society network. Thus, we put the depth of request in each request that is the number of people this request is roamed over, and if it exceeds a certain threshold, then the request will be discarded. We could also put the availability of a specific person, and the time it takes to response the data request, and optimize routing table accordingly.

Additionally, we have a credit mechanism in our system. Each person has a specific level of credit by default. Later on, according to data services this credit system could exchange the credits of people. This system prevents people from free riding, this is somehow an incentive mechanism, and if people want to use tacit knowledge of other people according to this system, they should also provide tacit knowledge to others. The primary credit is provided in order to provide liquidation to the platform, but its price must be investigated.

The search service is not a big deal for grids, but when it comes to mobile phone platforms, it is challenging, because mobile phones do not have enough processing power, and they do not have access to huge metadata related to data items. Even if we want to have a simple search over mobile phones, the input capability is very limited, and makes it difficult to search over them. What is the solution for mobile phones? If we cannot search, then there will not be any data grid over mobile phones.

We think we should use the PC platform to provide search, but on mobile phones, each data has a specific DID as we explained before. We leverage a PC platform just to find DID. Since the data we are searching for could be repeated over time, a simple mapping of domain, especially ontology to data items, could solve our problem. This mapping is between metadata and DID by using hash functions. But let us review our scenario. If a person does not have time to reach a PC, what should they do? We could have two solutions. At first, a centralized service exists, which you can contribute by SMS. There could be an automatic machine in that center, which according to your history of preferences helps you. Secondly, there could be a person the behind SMS system or you can use a call center to find out DID.

A question may be raised, that doesn’t our system, contradict its aim? If we could find the index from the computer, why should we use a mobile phone? The point is that, we have tacits on mobile phones and we do not search for tacits on PCs, we just search for DID domains, and we could have many versions, created by many people about this DID. Each mobile phone also has a cache of favorite domains, thus, it should not refer to the search system for DIDs.

We have different data items, in our system, with many replicas, but how could we handle consistency? First, only a person who provides data could change it. Data items have versioning systems, they have a field called validity of data, which we explained before, and searching for a special DID could have the characteristics of validity, which means that if the validity of data is lower than the threshold of the query, it will not be sent.

Each data item has a creation time that will be returned with that data, this could help the requester to decide whether it is valid for his request or not.

Each mobile phone has its own replica policy. According to that, it will let the data producer whether replicate the data over it or not. Replication could also be done by a pull model, which means the person who requests the data after using it, could publish this data as a replica; it will also advertise itself to the approximate node by a push model. The point is that each mobile phone has its own policy to accept the pushed or advertised data.

WICT/III-1569231353- 2509 © SoftCOM 2009

Since we have no updates, everyone could provide data items through changing previous data items. But the point is that, they would be creators, and should consider rating, which will have effect on their credits. Each mobile phone could have a data invalidator, which means if the validity of a replica becomes lower than a certain threshold it will be omitted.

Each data item could have an expiration time attribute. This is very good for news, and data that need to be overwritten periodically, it is used by data invalidator for replicas; update time attribute could be used also by replicas, to be updated or omitted.

Finally, all data in our system are broken into SMS size data, and the production is done based on it. Large data are broken into multiple SMSs, so apart from DIDs, they have a package number, which is used to find different packages, from multiple sources, and then reassemble them in the last node.

Figures 2 shows our system modules, detail of their interaction as

well as mechanisms described in the text.

6. EVALUATION

To evaluate our system, we should first have other systems to compare with. Unfortunately, there is no such a system to compare with, and also previous works on grid were not suitable for this platform due to the unique characteristics of this platform. But we compare the proposed system with non-functional requirement of this specific platform.

Being lightweight is the main non-functional requirement, which is supported by our layered architecture. Modularity of proposed system provides flexibility, and adaptability. This modular system could provide even lighteweighness, which means someone who has a mobile phone with a higher memory and processing capacity can install more modules, enjoy the optimized application, and less battery power usage. Our proposed system also provides adaptability, which means the threshold of caches and other parameters could be set according to special devices’ requirement.

Our mechanisms are optimized and less computation, and more distribution, makes it more suitable for mobile phones platform. Additionally, our platform is scalable, since our mechanisms are configured for distribution, and do not depend on central components.

7. CONCLUSION

Today mobile phones, are widespread, however there is no software system to provide data grid capabilities on mobile phones. Most previous works were focused on looking at mobile phone as terminal devices. We changed our view, and look at mobile phones as servers, we defined the concept of society network, which could be leveraged to enhance mobile phone data grid capabilities. Our contribution was redefining mechanisms of data grid modules on

mobile phones. Due to unique characteristics of mobile phones, previous solutions are not suitable, and we define the mechanism ourselves. Our publish service was not done globally like grid, but we use locally published self aware data items’ concepts. Our approach is adaptable, and makes the performance better overtime. We used caches, to improve performance. Our architecture provides flexibility, adaptability, and performance for mobile phones.

Our work could be enhanced by investigating thresholds for different mobile phones. Moreover, incentive system, like grid economy should be developed in order to launch the system, and prevent free riders.

7. ACKNOWLEDGEMENT

This research was created through a financial grant offered by Iran Telecommunication Research Canter. REFERENCES [1] F.Magoules, J.Pan, K.A.Tan, A.Kumar, “Introduction to grid computing”, Taylor & Francis Group, 2009. [2] Y.Deng, F.Wang, “A Heterogeneous Storage Grid Service”, ACM. 2007. [3] P. Kunszt, “European DataGrid project: status and plans”, Proc. Elsevier, 2003. [4] S.Kurkovsky, Bhagyavati, A.Ray. “Modeling a Grid-Based Problem Solving Environment for Mobile Devices”, Proceedings of the International Conference on Information Technology: Coding and Computing 2004. [5] A.Coronato, Gi. D.Pietro: MiPeG: A middleware infrastructure for pervasive grids. Future Generation Comp. Syst. 24(1): 17-29 (2008). [6] D.Mengistu, P.Davidsson, L.Lundberg, “A Grid Based Simulation Environment for Mobile Distributed Applications”, Proceeding International Conference on Multimedia and Ubiquitous Engineering, 2007. [7] W.Y.Zeng Y.L. Zhao, J.W. Zeng , W.Song , “Mobile Grid Architecture Design and Application”, Proceeding 4th International Conference on Wireless Communications, Networking and Mobile Computing, 2008. [8] K. Katsaros and G.C. Polyzos, "Evaluation of scheduling policies in a Mobile Grid architecture," Proceeding International Symposium on Performance Evaluation of Computer and Telecommunication Systems, 2008. [9] Chunlin Li, Layuan Li, “Energy constrained resource allocation optimization for mobile grids”, Energy constrained resource allocation optimization for mobile grids. In Press, Corrected Proof, Available online 24 June 2009. [10] M. Parashar and J-M. Pierson, "When the Grid becomes Pervasive: A Vision on Pervasive Grids,", Proceedings of the 8th Hellenic European Research on Computer Mathematics & its Applications Conference, 2007. [11] R. Moreno, “A Hybrid Mechanism for Resource/Service Discovery in Ad-Hoc Grids”, Journal Future Generation Computer Systems , 2008. [12] A. Litke, D.Skoutas, K.Tserpes, T.A. Varvarigou: Efficient task replication and management for adaptive fault tolerance in Mobile Grid environments. Future Generation Comp. Syst. 23(2): 163-178, 2007. [13] L.Son, M.,Calitz, A.P. A Model for Collaboration Services between Mobile Devices. Proceeding the South African Telecommunications Networks and Applications Conference ,2006.

DID finder module

Self aware data manager

Favorite service cache

Queuing service

Favorite DID cache

Credit manager

Local policy manager

Data item creator

Data Invalidator

Short path optimizer

Routing manager

Figure 2 Proposed System Modules

WICT/III-1569231353- 2509 © SoftCOM 2009

Usporedba metoda za određivanje praga u svrhu segmentiranja igrača

Vladimir Pleština1, Nikola Rožić2, Vladan Papić1

Prirodoslovno matematički fakultet1 Fakultet elektrotehnike strojarstva i brodogradnje2

Sveučilište u Splitu Split, Croatia

e-mail: vladimir.plestina@pmfst.hr, nikola.rozic@fesb.hr, vladan.papic@pmfst.hr Sažetak: Jedan od najjednostavnijih načina za izdvajanje objekata sa sive slike je primjena praga. Za automatsko određivanje praga postoji više metoda. U ovom radu su prikazane dvije metode koje se temelje na ukrštenoj entropiji, te Otsu metoda, temeljena na histogramu. Eksperimentom je na kraju prikazano koja je metoda najbolja za odvajanje igrača sa slike.

1. UVOD Problemi praćenja objekata je jedan od čestih problema u računalnom vidu. Za samo praćenje korištene su razne metode. Statističke metode su pokazale prilično dobre rezultate i uglavnom su dio svake nove metode. Kod praćenja samo jednog objekta na slici, stvari se uvelike pojednostavnjuju, dok u slučajevima više objekata (više igrača na terenu, molekule plina, skupina ljudi ispred nadzorne kamere) potrebno je naći primjerenu metodu pomoću koje ćemo dobiti najbolje rezultate. Algoritmi za pronalaženje objekata na slikama se uglavnom temelje na pretpostavkama o onome što bi se trebalo pratiti. Neki autori su u pred obradi slike uzimali u obzir već ranije poznatu pozadinu koju bi oduzimali i na taj način dobivali željene objekte. Sullivan u svom radu [4] u kojem vrši praćenje nogometaša uzima u obzir poznatu pozadinu i boju dresova. Koristeći Gaussianovu distribuciju, pronalazi objekte, odnosno igrače, na slici. Ovakvo označavanje igrača kao objekata daje mogućnost praćenja istih, ali ne definira konkretnog igrača. Iwase u svom radu [5] vrši jednostavno oduzimanje slike od poznate pozadine. Sliku zatim pretvara u binarnu, filtrira šumove i detektira pojedine regije koje je dobio. Problem koji se pojavljuje su sjene objekata i preklapanja jer je potrebno imati dodatni algoritam koji razdvaja preklopljene objekte, odnosno ljude koji se gibaju. Na isti način Figueroa u svom radu [6] započinje algoritam odvajanja, a dobivene segmente naziva „blob“ koje poslije prati. Nadalje u algoritmu, Figureoa koristi model nogometaša gdje pomoću vertikalne distribucije intenziteta i modela određuje kojoj momčadi igrač pripada. Gedikli [7] segmentira nogometaše pomoću poznatog modela boja, zatim pomoću poznatih predložaka i distribucije boja računa najsličniji model. Nakon toga koristi predviđanje i uz ranije izračunate podatke razdvaja igrače i označava ih.

Xu us svojim radovima [8] i [9] se oslanja na odvajanje poznate pozadine uz korištenje Gaussiana. U slučajevima kada postoji gibanje u određenim sekvencama, najjednostavniji način za pronalaženje objekta je oduzimanje trenutnog framea od poznate pozadine. Pretvaranje takve slike u binarnu sa određenim pragom daje nam objekte koji se inače ne nalaze na slici. Metoda ukrštene entropije (Cross-Entropy Method, CEM) je metoda koja potječe iz područja simulacije rijetkih događaja gdje je potrebno što točnije odrediti vjerojatnost pojave događaja male vjerojatnosti. Simulacija rijetkih događaja i uzorkovanje po važnosti osnove su ove metode. CEM ima dvije faze: • generiranje slučajnih uzoraka rješenja prema određenom mehanizmu • obnavljanje parametara mehanizma u svrhu dobivanja boljih uzoraka U računalnom vidu obrada i segmentacija slike je jako bitan faktor. Segmentacija se uglavnom vrši na slikama u svom tonu kod kojih je potrebno odrediti prag gdje se odlučuje što pripada objektu, a što pozadini. Prag je moguće subjektivno odrediti, ali se češće koriste metode za automatsko određivanje praga na slici. Jedna od najčešće korištenih metoda za određivanje praga je Otsu metoda i u ovom radu će biti uspoređena ova metoda sa dvije metode koje se temelje na ukrštenoj entropiji.

2. IZDVAJANJE INTERVALA SIVIH TONOVA Identificiranje dijelova slike koji predstavljaju neki objekt je jedan od najvažnijih problema kojima se bave sustavi vida. Segmentacija slike je dijeljenje slike u regije. Izdvajanje intervala sivih tonova je metoda kojom se želi postići konverzija slike s više intenziteta sivih tonova u binarnu sliku s ciljem odvajanja objekta od interesa i pozadine. Za uspješno provođenje metode potrebno je da kontrast između objekta i pozadine bude dovoljno velik te da je poznat nivo intenziteta i objekta i pozadine [3]. Prema tome, vrijednosti intenziteta određuju vrijednost praga s koji je potrebno postaviti. Neka je binarna slika B(i,j) ista kao i slika koja je dobivena izdvajanjem intervala sivih tonova fT(x,y) koja je dobivena korištenjem praga s za originalnu sliku f(x,y). Stoga se može pisati:

WICT/III B - 1569231401- 2509 © SoftCOM 2009

),(),( yxfjiB s= gdje za tamniji objekt na svjetlijoj pozadini

⎩⎨⎧

>≤

=syxfsyxf

yxfT ),(,0),(,1

),(

Zbog jasnijeg označavanja, umjesto B(x,y) pišemo B(i,j). Također, na ovaj način se ističe i vrsta varijable, tj. točke slike imaju cjelobrojne koordinate. Ako je poznato da su vrijednosti intenziteta objekta iz intervala [ ]21 , ss , tada se može koristiti slijedeća relacija:

⎩⎨⎧ ≤≤

=inace

syxfsyxf s ,0

),(,1),( 21

Opći zapis, kada nivoi intenziteta nekog objekta mogu biti iz više disjunktnih intervala jest

⎩⎨⎧ ∈

=inace

Zyxfyxf s ,0

),(,1),(

gdje je Z skup vrijednosti intenziteta komponenti objekta.

a) b)

Slika 2.1. Binarizacija slike korištenjem jedne vrijednosti praga (a) i dvije vrijednosti praga (b)

Opisani postupak je način isticanja željene vrijednosti binarizacijom slike. Pikseli kojima je nivo sivih tonova unutar vrijednosti koje se žele istaknuti poprimaju vrijednost 1, a ostali poprimaju vrijednost 0 (Slika 2.1.). Ukoliko se ne radi samo s binarnim slikama, moguće je pikselima čija je vrijednost unutar željenog opsega pridružiti neku višu vrijednost, a ostale ne mijenjati (Slika 2.2.).

Slika 2.2.

Pikselima čija je vrijednost unutar intervala [ ]21 , ss pridružuje se neka viša vrijednost, a ostale ne mijenjamo. U ovom slučaju izlazna slika nije binarna. Za ispravno izdvajanje objekta od pozadine tj. postavljanje prave vrijednosti praga potrebno je znanje o sceni i problemu kojeg imamo na originalnoj slici [10]. Jedna vrijednost praga može biti dobra u jednom slučaju, a ista vrijednost za drugi slučaj (drugu sliku) može biti potpuno neupotrebljiva. Za postavljanje ispravne vrijednosti praga (pragova) sustav može prvo analizirati scenu. Automatsko binariziranje postavljanjem vrijednosti praga često je prvi korak u analizi slika sustava računalnog vida. Postoji mnogo tehnika razvijenih za korištenje distribucije intenziteta slike i znanja o objektima koje promatramo za odabir prave vrijednosti praga.

3. METODE ZA ODREĐIVANJE PRAGA 3.1. Metoda minimalne ukrštene entropije (Li i Lee metoda) Ukrštena entropija je informacija o teoretskoj udaljenosti D između dvije distribucije vjerojatnosti. Neka su P i Q dvije razdiobe vjerojatnosti , , … , i

, , … , ukrštena entropija D(P,Q) je definirana kao:

, ∑ log (3.1) Kako ∑ ∑ 1, Gibbsova nejdenakost ( ∑ log ∑ log ) osigurava da D(P,Q) ≥ 0 i da jednakost jedino vrijedi 1,2, … , . Jasno je da D(P,Q) nije simetrično, odnosno da D(P,Q) ≠ D(Q,P). Simetrija se može dobiti na način:

, , (3.2)

Neka je , slika s L sivih tonova. ,1,2, … , je skup sivih tonova. Neka je hi frekvencija sivih

tonova i, a pi=hi/(M x N) vjerojatnost pojavljivanja sivih tonova i. Neka je s pretpostavljeni prag, tada s dijeli sliku na dva područja, objekt i pozadinu. Možemo pretpostaviti da sivo područje [0 - s] određuje područje objekta, a područje [(s+1) - L] određuje pozadinu. Li i Lee [1] su definirali ukrštenu entropiju segmentirane slike sa pragom s kao:

∑ log ∑ log (3.3.) gdje su

/

/

WICT/III B - 1569231401- 2509 © SoftCOM 2009

3.2. Palova metoda određivanja praga ukrštenom entropijom Pal [2] se u svom radu nadovezuje na Li i Lee algoritam. Smatra da je jednadžba 3.3. razuman kriterij za segmentaciju. Metoda koju Pal predlaže koristi simetričnu verziju razdiobe:

, ∑ log ∑ log (3.4.)

S je prag razdiobe. Promatrana distribucija vjerojatnosti sive skale područja objekta je definirana kao:

, , … , , a područja pozadine: , , … , gdje su

, 1,2, … , ;

, 1, 2, … , ;

vrijedi da je ∑ ∑ 1 Neka je QO i QB distribucija vjerojatnosti područja objekta i pozadine temeljen na nekom modelu. Drugim riječima , , … , i

, , … , su vjerojatnosti temeljene na nekom modelu. Ukrštena entropija područja objekta prema Palu je DO(s)=DS(PO,QO), a pozadine je DB(s)=DS(PB,QB). Ukrštena entropija segmentirane se tada može pisati:

log

log log

log

Da bi segmentirali sliku, minimiziramo D(s) s obzirom na s. Da bi mogli upotrijebiti ovu jednadžbu, potrebno je prvo odrediti QO i QB . Histogram sive slike je uglavnom modeliran kao mješavina normalnih distribucija. Noviji radovi histogram sive slike modeliraju sa mješavinom poissonovih distribucija. Prema Palu [2] modeliranje sive slike histograma mješavinom poissonovih distribucija se temelji na teoriji oblika slike. Koristeći poissonov model za dobivanje QO i QB pretpostavlja se da vrijednosti objekta imaju poissonovu distribuciju sa parametrom λO, a vrijednosti pozadine imaju poissonovu distribuciju sa parametrom λB.

Prema tome:

! , 1,2, … . , .

! , 1, 2, … . , . gdje su λO, i λB.

/

/

3.3. Otsu metoda za određivanje praga Otsu metoda [13] kreće od pretpostavke da su na slici dvije regije, tj. da su na histogramu intenziteta slike dva vrha koja nastaju kao superpozicija dvije razdiobe s različitim srednjim vrijednostima. I u ovom slučaju se primjenjuje iterativni postupak određivanja vrijednosti praga koji razdvaja dvije razdiobe. Za neku početnu vrijednost praga s, izračuna se srednja vrijednost 1μ za grupu R1, te 2μ za R2. U narednom koraku izračunava se težinsko razdvajanje srednjih vrijednosti:

( ))()(

)()()()()(21

2

21221

sNsNNsNssNssD

⋅⋅⋅−⋅

=μμ

μ

i

⎥⎦

⎤⎢⎣

⎡⋅⋅

⋅−⋅−=

)()()()()()(4)(

21

1222

212

2sNsNN

sNssNssD μμσ

σ

gdje je 2σ varijanca svih piksela slike, N ukupan broj piksela, N1 – broj piksela regije R1, a N2 – broj piksela regije R2. Prag s se odabire tako da se minimizira izraz

)()()(

2 sDsDsD

μσ

μ

+

U pojednostavnjenom obliku, Otsu metoda minimizira μD .

4. EKSPERIMENT SA SLIKOM NOGOMETNE UTAKMICE

U eksperimentu su korištene tri slike sa nogometne utakmice. Slike su u boji i sadrže teren, te igrače u plavom i bijelom dresu. Prvi korak je da se slike prebace u sive slike. Nakon toga je moguće odrediti prag i segmentirati pojedine dijelove slike.

WICT/III B - 1569231401- 2509 © SoftCOM 2009

4.1. Slika 1. Učitavamo prvu sliku i prebacujemo je u sivu skalu. Nakon primjenjujemo ranije opisane metode.

Metoda Lee Pal Otsu Prag 136 212 145

Brzina (sec) 0,6777 0,0775 0,0313

Slika 4.1. Originalna i segmentirana slika Lee metodom

Slika 4.2. Originalna i segmentirana slika Pal metodom

Slika 4.3. Originalna i segmentirana slika Otsu metodom

4.2. Slika 2.

Metoda Lee Pal Otsu Prag 98 212 148

Brzina (sec) 0,6096 0,1522 0,6097

Slika 4.4. Originalna i segmentirana slika Lee metodom

Slika 4.5. Originalna i segmentirana slika Pal metodom

Slika 4.6. Originalna i segmentirana slika Otsu metodom

4.3. Slika 3.

Metoda Lee Pal Otsu Prag 98 210 147

Brzina (sec) 0,6752 0,1858 0,0297

Slika 4.7. Originalna i segmentirana slika Lee metodom

Slika 4.8. Originalna i segmentirana slika Pal metodom

Slika 4.9. Originalna i segmentirana slika Otsu metodom

WICT/III B - 1569231401- 2509 © SoftCOM 2009

4.4. Analiza rezultata i poboljšanje Napravljeni eksperiment prikazuje usporedbu algoritama za određivanje praga za segmentiranje sive slike. Određivanje da li je rezultat dobar odnosno prihvatljiv može se zaključiti uspoređivanjem segmentirane slike i dobivenih vrijednosti praga. Analizirajući rezultate, vidi se da Leeova metoda daje najlošije rezultate, a prilično slične slike imaju znatno različit prag. Palova i Otsu metoda ističu svijetle igrače koji se u daljnjoj obradi slike mogu izdvojiti.

Metoda Lee Pal Otsu Slika1 136 212 145 Slika2 98 212 148 Slika3 98 210 147

Tablica 4.1. - prag

Metoda Lee Pal Otsu Slika1 0,6777 0,0775 0,0313 Slika2 0,6096 0,1522 0,6097 Slika3 0,6752 0,1858 0,0297

Tablica 4.2. - brzina obrade

Palove metoda je poboljšana Leeova metoda i očekivano je da pri segmentaciji daje bolje rezultate. Problem se javlja kod Lee metode jer primjenom entropijske metode i uzimanjem slučajnih uzoraka, prag se postavlja nisko. U tablicama 4.1. i 4.2. se mogu vidjeti rezultati brzine obrade za pojedinu metodu. Može se primijetiti da je Otsu metoda daleko brža od metoda ukrštene entropije i to je razlog zašto se ona primjenjuje u raznim aplikacijama za automatsko određivanje praga. Ovisno o onom što nam je potrebno, ove metode je moguće primijeniti na drugim vrstama slika. Primjenom ovih metoda samo na H komponenti u HSI modelu boja očekujem da bi dobili bolje rezultate i segmentirali bi samo značajne vrijednosti (dominantne boje) na slici. Isto tako u slučaju primjene algoritama za određivanje više pragova, entropijske metode bi se mogle koristiti za segmentiranje točno željenih objekata na slici.

4. ZAKLJUČAK

U ovom radu su teorijski opisane metode koje se koriste za određivanje praga i segmentiranje slika. Prve dvije metode se temelje na ukrštenoj entropiji i uspoređene su sa trenutno najraširenijom i najprimjenjenijom metodom. Ideja je bila usporediti metode temeljene na ukrštenoj entropiji za automatsko određivanje praga sa Otsu metodom. Danas se statističke metode i metode temeljene na entropiji često koriste u računalnom vidu. Ako u provedenom eksperimentu usporedimo samo metode ukrštene entropije, tada možemo vidjeti da je subjektivno

(gledanjem dobivenih slika) i po brzini obrade Palova metoda puno bolja od Lee-ove metode. Iako je u ovom radu prikazano korištenje ukrštene entropije na jednostavnim primjerima i na slikama koje imaju samo vrijednosti sivih tonova (od 0 do 255), smatram da se prikazani algoritmi mogu koristiti i na složenijim slikama (slike u boji) ili u drugom području (HSI). Analiza je izvršena na Dell OptiPlex 320 računalu sa procesorom Pentium IV 3GHz i sa 1 GB rama pomoću Mathoworks Matlab-a 2007a. ZAHVALA Ovaj rad je podržan od Ministarstva Znanosti i Tehnologije Republike Hrvatske pod projektima: Računalni vid u identifikaciji kinematike sportskih aktivnosti (177-0232006-1662) i ICT sustavi i usluge temeljeni na integraciji informacija (023-0231924-1661) LITERATURA [1] Li E. H., Lee C. K., "Minimum cross-entropy thresholding", Pattern Recognition, 617-625, 1992. [2] Pal N. R., "On minimum cross-entropy thresholding", Pattern Recognition, vol. 29, pp. 575-580, 1996. [3] Ramesh Jain, Rangachar Kasturi, Brian G. Schunck "Machine vision", MIT press, 1995 [4] Josephine Sullivan and Stefan Carlsson: "Tracking and Labelling of Interacting Multiple Targets", In Proc. 9th European Conf. on Computer Vision (ECCV 2006) [5] Sachiko Iwase and Hideo Saito: "Tracking Soccer Players Based on Homography among Multiple Views" , Visual communications and image processing. Conference, Lugano , ITALIE (08/07/2003) 2003, vol. 5150 (3), pp. 283-292 [6] Pascual J. Figueroa, Neucimar J. Leite, Ricardo M.L. Barros: "Tracking soccer players aiming their kinematical motion analysis", Computer Vision and Image Understanding, Volume 101 , Issue 2 (February 2006) Pages: 122 - 135 , Publication: 2006 [7] Suat Gedikli, Jan Bandouch, Nico v. Hoyningen-Huene, Bernhard Kirchlechner, and Michael Beetz: "An Adaptive Vision System for Tracking Soccer Players from Variable Camera Settings", In Proceedings of the 5th International Conference on Computer Vision Systems (ICVS), 2007 [8] Ming Xu James Orwell Graeme Jones: "Tracking football players with multiple cameras" , ICIP 2004: 2909-2912 [9] Ming Xu Liam Lowey James Orwell: "Architecture and algorithms for tracking football players with multiple cameras", IEE proceedings. Vision, image and signal processing, 2005, vol. 152, pp. 232-241 [10] Seul, O'Gorman, Sammon. "Practical Algorithms for Image Analysis", Cambridge university press, 2000

WICT/III B - 1569231401- 2509 © SoftCOM 2009

Comparison of threshold methods for players segmentation

Vladimir Pleština1, Nikola Rožić2, Vladan Papić1

Faculty of Science1 Faculty of Electrical Engineering, Mechanical Engineering and Architecture2

University of Split Split, Croatia

e-mail: vladimir.plestina@pmfst.hr, nikola.rozic@fesb.hr, vladan.papic@pmfst.hr Abstract: The simplest way to extract objects from the gray scale image is threshold appliance. To automatically set the threshold there are number of methods. In this work we present two methods based on cross entropy and Otsu method, based on histogram. With experiment we compare methods and determine best method for separating players from the image.

WICT/III B - 1569231401- 2509 © SoftCOM 2009

Abstract: In an environment where employees and their experience are of central value to the company, human resources management (HRM) represents a significant aspect of business efficiency. In this paper we present a model of an HRM information system for universities. Special attention is paid to aspects of the system that support processes specific to science and higher education. UML was used for modelling the global and detailed system architecture and database model. FURPS+ methodology was used for classification of requirements and the MoSCoW method for analysis of requirement priority.

1. INTRODUCTION A human resources management (HRM) system is an

important part of every business organization. Science and higher education institutions are no different in this respect. In addition to regular human resource (HR) data, as in any other organization, such institutions need to include specific data in accord with national and institutional laws and regulations. As a result it may be necessary for such institutions to develop a specialised HRM system.

In this paper we describe the development of the HRM system in use at the University of Dubrovnik .

The basis for the development of a model for the HRM system is described in the second section of the paper. The third section describes the methodology used for the requirement analysis. The fourth section explains in detail the academic professional development structure in Croatian science and education institutions, which determines the principal part of a specific HR system for science and education institutions.

The structure of the information system is described in the fifth section. Special attention is paid to processes specific to

science and higher education. The sixth section describes the architecture of the final HRM system implemented at the University of Dubrovnik.

2. BASIS FOR MODEL CREATION Interviews with the head of Human Resources at the

University of Dubrovnik as well as a study of their business data served as the main sources of information used for the development of the model. Research on the Law on Higher Education and Science provided additional information for setting up a model.

For the first stage we implemented a version of the HRM system as a character application, applicable to a range of industrial and business situations with a view to establishing the feasibility of database re-engineering and migration to another platform. Users entered real data in order to verify that the application suited their needs. After a series of test iterations we began the development of a new version of the HRM application.

At the next stage we re-wrote a complete legacy application on Windows GUI platform. In order to ensure parallel functioning of the old and newer versions of the HRM system we were forced to keep the legacy data structure but we conducted a database re-engineering wherever possible. For that reason there is some redundancy in the database.

The final development stage involved customisation by incorporating some new functionalities and features specific to higher educational institutions according to customer requirements.

Ivona Zakarija Department of Electrical Engineering and Computing

University of Dubrovnik Ćira Carića 4, Dubrovnik, Croatia E-mail: ivona.zakarija@unidu.hr

Zoran Skočir

Faculty of Electrical Engineering and Computing University of Zagreb

Unska 3, Zagreb, Croatia E-mail: zoran.skocir@fer.hr

Krunoslav Žubrinić Department of Electrical Engineering and Computing

University of Dubrovnik Ćira Carića 4, Dubrovnik, Croatia

E-mail: krunoslav.zubrinic@unidu.hr

Human resources management system for Higher Education institutions

WICT/III - 1569232113- 2509 © SoftCOM 2009

3. REQUIREMENT ANALYSIS The requirements engineering approach was used to

capture system requirements. Requirements engineering [1] is a term used to describe the activities involved in eliciting, prioritizing, documenting, and maintaining a set of requirements for a software system. It is a process of assessment of what stakeholders need the system do for them.

Since there are often conflicting requirements that must be balanced, a degree of compromise and negotiation is generally required. According to FURPS + classification requirements are placed into these categories:

functional requirements describing the functionality that the system is to execute; what behaviour the system should offer (capabilities);

non-functional requirements are ones that act to constrain the solution.

FURPS+ methodology has been adopted by many organizations and integrated into an international standard ISO/ IEC 9126. [2] It is an acronym for a model for classifying software quality attributes:

Functionality Usability Reliability Performance Supportability

“+” (ancillary) : Implementation Interfaces Operations Packaging Licensing

We used the MoSCoW analysis to achieve a clear prioritization of requirements. MoSCoW stands for:

Must Have Should Have Could Have Won’t Have This Time Around [3]

Requirements are documented in the standard form where functional requirements are separated and the MoSCoW list is integrated. Each requirement has a Priority attribute that can take one of the values M, S, C, or W. Modelling customer requirements in this way make a significant contribution in determining the project scope.

4. ACADEMIC PROFESSIONAL DEVELOPMENT University personnel comprise two major groups: administrative staff ; academic staff.

Besides existing standard personnel data the HRM system for Higher Education institutions must take account of academic professional development. Under the law [4] in higher institutions scientific research activities can be performed by:

scientists and researchers; teachers and associates .

of appropriate level. Employees occupy academic positions according to [5] with official document issued by higher institutions.

4.1 Scientists and Researchers The work of scientific research is carried out by qualified

scientists and researchers, who, under existing rules and procedures [6][4][7], can be classified into the following grades [8]: Scientists:

research associates, senior research associates and research advisors,

Researchers: expert assistant, younger assistent, assistent, senior

assistent

4.2 Register of Researchers

The Ministry of Science, Education and Sports keeps records of scientists and researchers in the Register of Researchers. The following categories appear in the Register:

research associates, senior research associates and research advisors;

assistant professors, associate professors and full professors;

external associates – assistant professors and senior assistant professors,

persons with a doctoral degree. Universities initiate procedures of entry into the Register

based on the submission of an application to the Ministry with all relevant documentation. Each person has a unique ID number assigned – identity number of scientists. In the HRM system applications are supported by additional employee data.

WICT/III - 1569232113- 2509 © SoftCOM 2009

Fig. 1 UML activitiy diagram shows conducting of the grade appointment procedure

4.3 Teachers and Associates The work of education is carried out by either full time

teachers or associate staff; most university teachers are also active as scientists and researchers at their university or other research institute. Under the rules and procedures in force teachers can be of the following grades: • scientific-research grades:

assistant professor, associate professor, full professor, professor emeritus.

• teaching grades: lecturer, senior lecturer, professor of high school,

lector, senior lector, repetiteur, senior repetiteur. Associate grades:

expert assistant, younger assistant, assistant, high school assistant, senior assistant.

Academic grades with their corresponding required skills in teaching and research are defined by the University Law. Promotion of academic staff is designed to recognise and reward sustained excellence. Assessments are made by committees of peers through a process designed to enable fair

and consistent application of standards. To be promoted, academic staff members must, on objective evidence, attain an appropriate standard.

5. DESCRIPTION OF INFORMATION SYSTEM MODEL

Based on the results of research and interviews a information system model has been set up. HRM system consists of following segments as presented in Figure 2. We will not in detail elaborate each particular segment, though special attention is paid to aspects of the system that support processes specific for science and higher education.

5.1 Conducting the Grade Appointment Procedure Universities teachers are appointed to scientific-research

and teaching grades for a period of five years. Appointments are made on the basis of open application.

Based on the regulations and procedures in force [9] the University announces vacancies for scientific-research and teaching post appointments. The University aims to ensure

WICT/III - 1569232113- 2509 © SoftCOM 2009

equity, transparency and fairness in all aspect of the appointment process.

The University must inform all applicants of the criteria and procedures for appointment. Faculty Council make a decision to initiate the procedure and select the board members of the Expert Committee.

After the official announcement the Committee accept applications and documentation and when the application is closed examines those materials. Based on their reports the competent Registrar Board make a decision and list applicants for promotion of grade and academic position. After confirmation from Area Council/Senate the appointments are recognised. Figure 1 shows the UML activity diagram of conducting the grade appointment procedure.

Fig. 2 HRM system segments

The HRM system supports appointment/re-appointment processes through the module for managing the academic grade records as shown on Figure 4. It is a part of Higher Education additional data administration segment. Detailed data on all recognized grade appointments are recorded, and the system gives a review of warnings and notifications of grade appointment expiry.

This is important because the procedure for a grade application must be initiated 3 months before the grade expires and the institution must prepare applications within that period of time.

Evidence for assessment for promotion to a higher academic grade may include: articles in journals, including electronic journals; articles in the proceedings of, and presentations to, national and international conferences. In the HRM system for employees records of such scientific papers are in the form of title, type of work, publishing date, and URL to the CROSBI where it is stored in electronic form. A module of HRM system is called Published scientific papers records. Universities can take an up to date

bibliography for employees without the need to enter a large amount of data. The Croatian Scientific Bibliography (CROSBI) is a digital archive which stores all scientific publications in Croatia [10].

Complete documentation related to promotion procedure applications, committee structure, decision of grade appointment and decisions of Registrar Board can be stored electronically in a document repository. In the HRM system it is called Attached document records – and involves simply entering the path where a document is stored in the repository. ActiveX control built into the HRM system ensures access to documents in the repository regardless of format (.pdf ,. doc,. xls,. jpg,. html)

Fig. 3 System architecture

6. SYSTEM ARCHITECTURE To achieve easier implementation and maintenance of the

application two-tier architecture is employed in our system. The architecture is composed of two layers: the database layer and the business logic. The database layer and the business logic are located on separate physical machines.

There are two kinds of clients: • “thin clients” - connect via RDP protocol to the

application terminal server where the complete application logic is performed. Terminal server handles the database and the business logic. The main role of “thin client” is to present application data in user interface.

“fat clients” - where the complete application logic is performed; such clients connect directly to the database server.

Figure 3 shows system architecture.

Terminal server

constraints{Web browser}

«execution environment»Terminal server :Panther run-time

Database server

constraints{UNIX server (AIX, LINUX,...)}

«SQL data...

Thin client

constraints{RDP client}

tagsOS = Windows XP/2000/2003RAM = 128 M B and more

PC client (fat)

constraints{Web browser}

«execution environment»PC client :Panther run-time

«Panther ...:BASE

«Panther ...:HRM

«Panther ...:WINM

«Panther ...:BASE

«Panther ...:HRM

«Panther ...:WINM

0..*«TCP/IP»1

1. .*«TCP/IP»1

0..*

«TCP/IP»

1

WICT/III - 1569232113- 2509 © SoftCOM 2009

Fig. 4 UML activity diagram showing a module for managing the academic grade records

7. CONCLUSION In the paper an HRM system for higher education

institutions is presented. Implementation of this system at the University of Dubrovnik has confirmed that it provides effective support for certain specific processes in the appropriate context. The use of the system for grade appointment procedures is shown in detail.

REFERENCES [1] J.Arlow, I.Neustadt, UML 2 and the Unified Process:

Practical Object-Oriented Analysis and Design, Addison-Wesley Object Technology Series,2005

[2] R. S. Kennet, E. R. Baker, Software Process Quality, Marcel Dekker Inc., New York – Basel, 1999.

[3] S. Ash, MoSCoW Prioritisation Briefing Paper, DSDM Consortium, http://www.dsdm.org/knowledgebase/details/165/moscow-prioritisation-briefing-paper.html (20.11.2008.)

[4] Law on Scientific and Research Activity (in Croatian), Official Gazette of Republic of Croatia “Narodne Novine”, No. 29/96.

[5] Regulation of the Job Titles and Complexity Coefficients in Science and Higher Education (in Croatian), Official Gazette of Republic of Croatia “Narodne Novine”, No. 38/01, 112/01, 62/02, 156/02, 162/03, 39/05, 82/05, 133/05, 30/06)

[6] Law on Higher Education (in Croatian), Official Gazette of Republic of Croatia “Narodne Novine”, No. 54/96.

[7] Law on Science and Higher Education (in Croatian), Official Gazette of Republic of Croatia “Narodne Novine”, No. 123/03, 105/04 i 174/04

[8] R.Zelenika, Methodology and technology of scientific and professional publication (in Croatian), Faculty of Economics Rijeka, Rijeka, 2000

[9] Regulation on Conditions for the Selection of Scientific Positions (in Croatian), Official Gazette of Republic of Croatia “Narodne Novine”, No. 84/05

[10] Croatian Scientific Bibliography (CROSBI) , Ruđer Bošković Institute, Ministry of Science, Education and Sports , http://bib.irb.hr/ (10.09.2008.)

WICT/III - 1569232113- 2509 © SoftCOM 2009

Abstract— In this paper we investigate influence of unsharp

masking technique on OCR performance. Three different filters are used for unsharp masking: Laplacian, adaptive Lap lacian and Teager filter. In order to reduce noise sensitivity we propose adaptive version of Laplacian unsharp masking: Our approach corrects output of Laplacian filter according to local variance of pixel values. Character images are taken from ICDAR 2003 dataset and enhanced with these 3 variations of unsharp masking filtering. Second step is binarization using Otsu’s criterion and processing with OCR software. Results show that approach with Teager filter shows best recognition rate. Unsharp masking based on Laplacian and adaptive Laplacian filter has weaker performance, but they still improve recognition rate in comparison with original images.

Index Terms— unsharp masking, Teager filter

I. INTRODUCTION Unsharp masking technique is often employed for

improvement of visual image quality. In this technique high pass filtered, scaled image is added to original image in order to enhance edges and detail information. Classic linear unsharp masking filtering uses linear high-pass filter (e.g. Laplacian), but this approach has two main drawbacks: noise sensitivity and overshooting artifacts.

Various methods are employed to reduce noise sensitivity. In [1] quadratic Teager filter, that can be approximately characterized as a local-mean-weighted adaptive high-pass filter, is used instead of linear one. Approach presented in [2] describes adaptive unsharp masking algorithm that enhance contrast in high detail areas only. Locally adaptive first-order recursive (YENI) filter is used in [3] to prevent over/under shooting artifacts.

In [4] it is shown that quadratic non-linear filters eficiently enhance character edges. Mancas-Thillou and Mirmehdi [5] propose super resolution technique for low resolution text images. They use unsharp masking with Teager filter in order to increase recognition rate of OCR software.

In this paper we investigate how image enhancement with unsharp masking techniques influence on OCR recognition rate. For this purpose three versions of unsharp masking are

used: Laplacian based, adaptive Laplacian based and Teager based. We propose adaptive Laplacian unsharp masking that corrects output of Laplacian filter in order to avoid noise amplification in smooth areas.

Character images enhanced with these 3 versions of unsharp masking are then binarized using Otsu’s criterion and processed with OCR software. Images are taken from ICDAR 2003 dataset containing a wide variety of single character images in natural scenes.

The paper is organized as follows. Section 2 describes preprocessing techniques for character recognition. Section 3 discusses unsharp masking with linear and adaptive linear filter. Section 4 describes Teager filter and its usage in unsharp masking. Results are presented in section 5 and conclusions are made in section 6.

II. IMAGE PREPROCESSING FOR CHARACTER RECOGNITION

According to [8] conventional preprocessing steps before character recognition in document images include noise removal/smoothing, skew detection/correction, connected component analysis, normalization, slant detection/correction, thinning and contour analysis.

Image smoothing or blurring removes small details from image and reduce noise. Linear filtering is usually used for this purpose. Simplest method is replacing the value of pixel with an unweighted average over a fixed region. This is the same as convolution with block of ones multiplied with constant. Filter kernel that gives better results is symmetric Gaussian kernel:

, exp ) (1)

This kernel forms a weighted average that weights pixels at its center stronger than pixels at boundaries.

In low contrast, blurred images smoothing is not appropriate approach. In that case unsharp masking technique can be used for sharpness enhancement. In [4] it is showed that Teager unsharp masking improve readability by a human user and recognition by OCR software.

Skew visually appears as a slope of text lines with respect

Influence of unsharp masking on OCR performance

Matko Šarić, Dinko Begušić and Hrvoje Dujmić

Faculty of electrical engineering, University of Split R. Boskovica bb, Split, Croatia

msaric@fesb.hr, Dinko.Begusic@fesb.hr, hdujmic@fesb.hr

WICT/III-1569237815- 2509 © SoftCOM 2009

to the x-axis and it can significantly affect recognition accuracy. Skew is first detected using average angles between connected component centroids or using projection analysis. After the skew angle has been detected, a coordinate rotation transformation is used to correct it.

Slant is character inclination normally found in cursive writing. Slant correction decrease variation of the script and enhance quality of characters. After slant angle estimation, a horizontal shear transform is used to shift characters to the left or to the right.

Character normalization is very important preprocessing operation that maps character image onto a standard plane (typically 32x32 or 64x64) to give representation of fixed dimensionality for classification. There are two different approaches to character normalization: linear and nonlinear normalization.

Contour tracing extract important information about character general shape. Main benefit from contour analysis arises from fact that contour is small subset of the total number of pixels representing a pattern. Therefore the amount of computation is greatly reduced. Square-tracing algorithm is often used for contour tracing because of its simplicity.

Thinning algorithms extract character skeleton that shows general shape of pattern and some important features can also be extracted from it.

III. UNSHARP MASKING WITH ADAPTIVE LAPLACIAN FILTER In unsharp masking algorithm (figure 1) the enhanced

image y(n,m) is obtained from input image x(n,m) as y(n,m)=x(n,m)+λz(n,m) (2) where z(n,m) is output of a high pass filter and λ is scaling

factor that determines level of contrast enhancement. The usual choice for high pass filter is Laplacian:

z(n,m)=4x(n,m)-x(n-1,m)-x(n+1,m)-x(n,m-1)-x(n,m-1) (3)

Usage of Laplacian filter suffers from two problems.

Firstly, whole system is very sensitive to noise because method sharpens smooth areas. Second problem is excessive enhancement of high-contrast areas in comparison with areas

Figure 1. Unsharp masking for image enhancement that do not have high image dynamics (overshooting

artifacts). In order to reduce noise amplification we exploit

pixel activity level ([2]). It is measured as local variance computed over 3x3 pixel block:

, ∑ ∑ , , (4) where , is pixel value level and , is average

pixel value level. Let t be positive threshold. If , < t then pixel belongs to smooth region where is desirable to avoid contrast enhancement. According to this scaling factor λ from (1) is defined as

0, ,

, , (5)

Equation (4) shows that output from Laplacian filter is set

to zero in case of low local dynamics. With this adaptive approach noise amplification in smooth areas is avoided.

Figure 2 shows original image, image enhanced with Laplacian unsharp masking and image enhanced with adaptive Laplacian unsharp masking. It is obvious that adaptive Laplacian filter enhances edges without noise amplification.

a) b) c)

Figure 2 a) original image b) unsharp masking with Laplacian filter c) unsharp masking with adaptive Laplacian filter

IV. UNSHARP MASKING WITH TEAGER FILTER 2-D Teager filter is class of quadratic Volterra filters[1]. 1-

D quadratic digital Volterra filter is given by the 2-D convolution of the 1-D sample products · with a 2-D kernel , :

∑ ∑ , · , (6)

Teager’s algorithm [6], which estimates measure of signal

energy, is a 1-D quadratic Volterra filter defined by

1 · 1 (7) The output from Teager’s algorithm can be approximated as

1 1 (8) where 1 1 /3. From (8) it

is clear that Teager filter can be considered as mean-weighted high pass filter. The 1-D Teager filter can be extended to

WICT/III-1569237815- 2509 © SoftCOM 2009

create 2-D Teager filter:

, 3 ,12 1, 1 1, 1

12

1, 1 1, 11, 1,

, 1 , 1 (9) This filter has stronger response in regions of high average

intensity than in regions of low average intensity. Thus unsharp masking with Teager filter has weaker enhancement in dark regions and stronger in bright areas satisfying Weber’s law [7]. Figure 3 illustrates effect of Teager unsharp masking.

a) b)

Figure 3. a) original image b) unsharp masking with Teager filter

V. EXPERIMENTS AND RESULTS We have tested three variations of unsharp masking :

Laplacian based, adaptive Laplacian based and Teager based. Character images are firstly enhanced using these 3 techniques in order to increase contrast between character and background. ICDAR 2003 dataset used in our experiments contains character images affected with different kinds of degradations like low contrast, blur, variation of illumination, noise, complex background etc.

Enhanced character images have to be binarized before OCR processing. Usually, thresholding of color images is done by first converting RGB image to grayscale and then applying different algorithms that can be divided in two categories: global and local or adaptive. In our experiments global Otsu method is chosen because of its simplicity and processing speed.

Our test set contains 400 single character images. Main criterion for comparison of different unsharp masking techniques is recognition rate, but we also use two measures that show in how many cases enhancement or degradation is achieved in comparison with original image:

(8)

(9)

Character is classified as enhanced when original image

isn’t recognized and filtered image is recognized. Similarly, character is degraded when original image is recognized and filtered image is not recognized. These 2 measures make possible deeper analysis of tested enhancement methods.

Original Laplacian UM

Adaptive Laplacian

UM

Teager UM

132 (47%) 137 (49%) 134 (49%) 139 (50%)

Table I. OCR recognition rates. Total number of characters

is 280

Method Enhancement rate Degradation rate

Laplacian UM 20 (7.1%) 15 (5.3%)

Adaptive Laplacian UM 18(6.5%) 16(5.7%)

Teager UM 17 (6%) 10(3.6%)

Table II. Enhancement and degradation rates. Total number

of characters is 280 Recognition rates are shown in table I. All three variations

of unsharp masking improve performance in comparison with original image. Unsharp masking based on Laplacian and adaptive Laplacian have same recognition rate while Teager unsharp masking shows best performance.

Although adaptive approach in case of Laplacian filter give perceptually more pleasant image with sharp edges without noise amplification (figure 4 and 5), recognition performance is same as in case of classical Laplacian unsharp masking. Adaptive version has even slightly weaker enhancement and degradation rate (table II). From this observation it is obvious that amount of noise added with Laplacian unsharp masking doesn’t have serious influence on recognition performance. Maybe the more sophisticated binarization method would show benefit from noise reduction.

Teager filter has best recognition rate. Enhancement rate is slightly weaker than in case of Laplacian filtering, but degradation rate is lowest. Presence of degraded characters can be explained by fact that unsharp masking enhance not only character edges, but also small details that can seriously affect recognition accuracy.

All three variations of unsharp masking technique mainly enhance perceptual quality, but obviously this kind of enhancement doesn’t necessary improve recognition accuracy. In order to test general impact of unsharp masking data set

WICT/III-1569237815- 2509 © SoftCOM 2009

used in this work contains characters with various types of degradations and complex backgrounds. As it is expected, recognition rate is increased mainly in character images with low contrast and blur.

VI. CONCLUSION An impact of unsharp masking to character recognition

accuracy is investigated in this paper. We tested three versions of unsharp masking technique: Laplacian based, adaptive Laplacian based and Teager based. Results demonstrate that all three methods enhance recognition results in comparison with original images. Teager unsharp masking gives best result. Although approach with adaptive Laplacian filter reduce noise sensitivity and results with perceptually better image, recognition rate is same as with standard Laplacian filter. All three methods enhance sharpness of character images, but recognition accuracy is improved mostly for low contrast and blurred images.

In future work we will try to combine unsharp masking with other filtering techniques like median filtering in order to remove unnecessary details and preserve character edges.

a) b)

c) d) Figure 4 .a) original image b) unsharp masking with

Laplacian filter c) unsharp masking with adaptive Laplacian filter d) unsharp masking with Teager filter

a) b)

c) d)

Figure 5 .a) original image b) unsharp masking with

Laplacian filter c) unsharp masking with adaptive Laplacian filter d) unsharp masking with Teager filter

REFERENCES [1] S.Mitra, G.Sicuranza. "Nonlinear Image Processing," Academic Press, 2000. [2] A. Polesel, G. Ramponi, and V. J. Mathews, "Image enhancement via adaptive unsharp masking," IEEE Trans. Image Processing, vol. 9, pp. 505–510, Mar. 2000. [3] T. Arici and Y. Altunbasak, "Image local contrast enhancement using adaptive non-linear filters," IEEE International Conference on Image Processing (ICIP-06), Oct. 2006, pp. 2881–2884. [4] G.Ramponi, P.Fontanot, "Enhancing document images with a quadratic filter," Signal Processing, 33:23-34, 1993. [5] C. Mancas-Thillou and M. Mirmehdi, "Super-resolution text using the teager filter," First International Workshop on Camera-Based Document Analysis and Recognition, pages 10–16, 2005. [6] J. F. Kaiser, "On a simple algorithm to calculate the ‘energy’ of a signal," Proc. IEEE ICASSP, Albuquerque, NM, vol. 1, pp. 381-384, 1990. [7] A.K.Jain. "Fundamentals of Digital Image Processing," Prentice Hall, 1989. [8] Mohamed Cheriet, Nawwaf Kharma, Cheng-Lin Liu, "Character recognition systems: a guide for students and practioners," Wiley-Interscience, 2007

WICT/III-1569237815- 2509 © SoftCOM 2009