The Workshop on Semantic Web and New Technologies was held by second time at the Faculty of Computer Science of Benemérita Universidad Autónoma de Puebla, Mexico in March 2009. The Semantic Web provides a common framework that allows data to be shared and reused across application, enterprise, and community boundaries. Semantic Web technologies are beginning to play a significant role in many diverse areas, marking a turning point in the evolution of the Web. The goal of this workshop is to provide a forum for the Semantic Web community, in which participants can present and discuss approaches to add semantics on the Web, show innovative applications in this field and identify upcoming research issues related to Semantic Web. In order to fulfill these objectives, the more important workshop topics included Semantic Search, Semantic Advertising and Marketing, Linked Data, Collaboration and Social Network, Foundational Topics, Semantic Web and Web 3.0, Ontologies, Semantic Integration, Data Integration and Mashups, Unstructured Information, Semantic Query, Semantic Rules, Developing Semantic Applications and Semantic SOA.
Dr John Cardiff was the invited speaker in this Second Workshop on Semantic Web, who is a fulltime lecturer and lead researcher in the Social Media Research Group, based at the Institute of Technology Tallaght, Dublin, Ireland. He has previously held positions in the Department of Computer Science, Trinity College Dublin, and in the University of Queensland, Australia, where he obtained the degree of Ph.D. in 1990. He has extensive experience in semantic web technologies, heterogeneous database research and query processing and optimization. He collaborates closely with researchers of the National Language Engineering Laboratory at the Polytechnic University of Valencia, Spain, the Knowledge and Data Engineering Group of Trinity College Dublin, and the IBM Dublin Center for Advanced Studies. He is currently supervising four PhD students who are investigating semantic web based recommender systems, blogosphere analysis, and adaptive hypermedia systems. Dr Cardiff has a wide breadth of experience of research and management of large European Union funded projects under programmes such as RACE, Esprit, and AIM. He has over 20 refereed publications in international conferences and journals. Invited Paper The Evolution of the Semantic Web John Cardiff Exploiting Wikipedia as a Knowledge Base: Towards and Ontology of Movies Rodrigo Alarcón, Octavio Sánchez and Víctor Mijangos Translation of Verbal Expressions and Context of Use Extraction through a Corpus on Web Arturo Velasco, María J. Somodevilla, and Ivo. H. Pineda Dynamic Concept-Based Taxonomy used for image recovery based on their textual description Jaime Lara, María de la Concepción Pérez de Celis and David Pinto The Use of Document Fingerprinting in the Web People Search Task David Pinto, Mireya Tovar, Beatriz Beltrán, Darnes Vilariño and Héctor Furlog mQA: Question Answering in Mobile devices Fernando Zacarías F., Alberto Tellez V., Marco Antonio Balderas and Rosalba Cuapa C. 1 8 17 26 37 44 Semantic Routing for Structured Peer-to-Peer Networks Luis Enrique Colmenares Guillén, Omar Ariosto Niño Prieto and Leandro Navarro Moldes Some Considerations for the Semantic Web María Elena Franco Carcedo 67 76 83 97 The Evolution of the Semantic Web
Social Media Research Group, Institute of Technology Tallaght, Dublin, Ireland email John.Cardiff@ittdublin.ie Abstract — The Semantic Web offers an exciting promise of a world in which computers and humans can cooperate effectively with a common understanding of the meaning of data. However, in the decade since the term has come into widespread usage, Semantic Web applications have been slow to emerge from the research laboratories. In this paper, we present a brief overview of the Semantic Web vision and the underlying technologies. We describe the advances made in recent years and explain why we believe that Semantic Web technology will be the driving force behind the next generation of Web applications.
The World Wide Web (WWW) was invented by Tim Berners Lee in 1989, while he was working at the European Laboratory for Particle Physics (CERN) in Switzerland. It was conceived as a means to allow physicists working in different countries to communicate and to share documentation more efficiently. He wrote the first browser and Web server, allowing hypertext documents to be stored, retrieved and viewed.
The Web added two important services to the internet - it provided a very convenient means for us to retrieve and view information - we can then see the web as a vast document store in which we retrieve documents (web pages) by typing in their address into a web browser. Secondly, it provided a language called HTML, which describes to computers how to display documents written in this language. Documents, or web pages, are accessed by a unique identifier called a Uniform Resource Locator (URL) and are accessed using a Web browser. Within a short space of time, the WWW had become a popular infrastructure for sharing information, and as the volume of information increased its use became increasingly widespread.
Although the web provides the infrastructure for us to publish and retrieve documents, the HTML language defines only the visual characteristics, ie. how the documents are to be presented on a computer screen to the user. It is up to the user who requested the document to interpret the information it contains. This seems counterintuitive, as we normally think of computers as the tools to perform the more complex tasks, making life easier for humans. The problem is that within HTML there is no consideration of the meaning of the document, they are not represented in a way that allows interpretation of their information content by computers.
If computers could interpret the content of a web page, a lot of exciting possibilities would arise. Information could be exchanged between machines, automated processing and integration of data on different sites could occur. Fundamentally, they could improve the ways in which they can retrieve and utilise the information for us because they would have an understanding of what we are interested in. This is where the Semantic Web fits into the picture today's web (the "syntactic" web) is about documents whereas the semantic web is about "things" - concepts we are interested in (people, places, events etc.), and the relationships between these concepts.
The Semantic Web vision envisages advanced management
of the information on the internet, allowing us to pose
queries rather than browse documents, to infer new
knowledge from existing facts, and to identify
inconsistencies. Some of the advantages of achieving this
goal include [
meaning, eg. knowing when two statements are equivalent, or knowing that a reference to a person in different web pages are referring to the same individual.
Integrating information across different sources − by creating mappings across application and terminological boundaries we can identify identical or related concepts, Improving the way in which information is presented to a user, eg. aggregating information from different sources, removing duplicates, and summarising the data.
While the technologies to enable the development of the
Semantic Web were in place from the conception of the
web, a seminal article by Tim Berners-Lee, James Hendler
and Ora Lassila [
The paper is organized as follows. In Section II we describe the problems we face when trying to extract meaning from the web as it is today. Section III presents a brief overview of the technologies underlying the Semantic Web. In Section IV we give an overview of the gamut of typical Semantic Web applications and Section V introduces the Linking Open Data project. Finally, we present our conclusions and look to the future in Section VI.
II.
In Figure 1 we see a "typical" web page written in HTML which we will use to exemplify some of the drawbacks of the traditional web. This page lists the keynote speeches which took place at the 2009 World Wide Web conference1. To the reader, the content of the page can be interpreted intuitively. We can read the titles of the speeches, the names of the speakers and the time and dates at which they take place. Furthermore, by familiarity with browser interaction paradigms, we can realize that by following a hyperlink we can retrieve information about concepts related to the conference (authors, sponsors, attendees etc.). In this example, by following the hyperlink labelled "Sir Tim Berners-Lee" we will retrieve a document containing information about the person of this name. We intuitively assign a meaning - perhaps "has-homepage" - to the hyperlink, allowing us to assimilate the information presented to us.
A web browser cannot assign any to these links we see in this page − a hyperlink is simply a link from one document to another and the interpretation of the meaning of the link (and of the documents themselves!) is a task for the human reader. All that can be inferred automatically is that some undefined association between the two documents exists.
The problems are even more clear when we consider the nature of keyword-based browsing. While search engines such as Google and Yahoo! are clearly very good at what they do, we frequently are presented with a vast number of results, many (most?) of which will be irrelevant to our search. Semantically similar items will not be retrieved (for instance a search for "movie" will not retrieve results where the word "film" was used). And most significantly, the result set is a collection of individual web pages. Our tasks often require access to multiple sites (such as when we book a holiday), and so it is our responsibility to formulate a sequence of queries to retrieve the individual web pages, each one of which performs part of the task at hand. There are two potential ways to deal with this problem. One approach is to take the web as it is currently implemented, and to use Artificial Intelligence techniques to analyze the content of web pages in order to provide an interpretation of its meaning. This approach however would be prone to error and would require validation. Furthermore, the rate at which the web is growing would render it practically impossible to achieve.
The other approach is to represent the web pages in a form in which we can represent and interpret the data they contain. If there is a common representation to express the meaning of the data on the web, we can then develop languages, reasoners, and applications which can exploit this representation. This is the approach of the Semantic Web.
The Semantic Web describes a web of data rather than
documents. And just as we need common formats and
standards to be able to retrieve documents from computers
all over the world, we need common formats for the
representation and integration of data. We also need
languages that allow us to describe how this data relates to
real world objects and to reason about the data. The famous
"Layer Cake" [
In this section, we summarize and discuss the key aspects
shown in the Layer Cake diagram. Firstly we describe the
core technologies: the languages RDF and RDFS. Next we
describe the higher level concepts, focusing in particular on
the concept of the ontology which is at the heart of the
Semantic Web infrastructure. Finally we examine the trends
and directions of the technology. For further information on
the concepts presented in this section, the reader is referred
to a more detailed work (eg. [
By way of example, the statement “The Adventures of Tom Sawyer” was written
by Mark Twain could be expressed in RDF by a statement such as <rdf:Description rdf:about=www.famouswriters.org/twain/mark> <s:hasName>Mark Twain</s:hasName> <s:hasWritten rdf:resource=
www.books.org/ISBN0001047> </rdf:Description> At first glance it may appear that this information could be equally well represented using XML. However XML makes no commitment on which words should be used to describe a given set of concepts. In the above example we have a property entitled "hasWritten", but this could equally have been "IsAuthorOf" or another such variant. So, XML is suitable for closed and stable domains, rather than for sharable web resources.
The statements we make in RDF are unambiguous and have a uniform structure. Concepts are each identified by a Universal Resource Identifer (URI) which allows us to make statements about the same concept in different applications. This provides the basis for semantic interoperability, allowing us to distinguish between ambiguous terms (for instance an address could be a geographical location, or a speech) and to define a place on the web at which we can find the definition of the concept. To describe and make general statements collectively about groups of objects (or classes), and to assign properties to members of these groups we use RDF Schema, or RDFS3. RDFS provides a basic object model, and enables us to describe resources in terms of classes, properties, and values. Whereas in RDF we spoke about specific objects such as '“The Adventures of Tom Sawyer” and "Mark Twain", in RDFS we can make general statements such as "A book was written by an author"
<rdf:Property rdf:ID=“HasWritten” <rdfs:domain rdf:resource=“#author”\> <rdfs:range rdf:resource=“#book”\> <\rdf:Property> An expansion of these examples, and the relationship between the graphical representations of RDF and RDFS is shown in Figure 3. 2 www.w3.org/RDF/
RDF and RDFS allow us to describe aspects of a domain,
but the modelling primitives are too restrictive to be of
general use. We need to be able to describe the taxonomic
structure of the domain, to be able to model restrictions or
constraints of the domain, and to be able to state and reason
over a set of inference rules associated with the domain. We
need to be able to describe an ontology of our domain.
The term ontology originated in the sphere of philosophy,
where it signified the nature and the organisation of reality,
ie. concerning the kinds of things that exist, and how to
describe them. Our definition within Computer Science is
more specific, and the most commonly cited definition has
been provided to us by Tom Gruber in [
In relation to the Semantic Web, for us to share, reuse and reason about data we must provide a precise definition of the ontology, and represent it in a form that makes it amenable to machine processing. An ontology language should ideally extend existing standards such as XML and RDF/S, be of "adequate" expressive power, and provide efficient automated reasoning support. The most widely used ontology language is the "Web Ontology Language", which curiously has the acronym "OWL"4. Along with RDF/S, OWL is a W3C standard and augments RDFS with additional constraints such as localised domain and range constraints, cardinality and existence constraints, and transitive, inverse, and symmetric properties.
Adding a reasoning capability to an ontology language is tricky since there will be a trade-off between efficiency and expressiveness. Ultimately it depends on the nature and requirements of the end application, and it is for this reason that OWL offers three sublanguages, 4 www.w3.org/2004/OWL
OWL Lite supports only a limited subset of OWL constructs and is computationally efficient, OWL DL is based on a first order logic called Description Logic, OWL Full offers the full compatibility with RDFS but at the price of computational tractability.
Examples of applications which could require very different levels of reasoning capabilities are described in the following section.
The top layers of the layer cake have received surprising little attention considering that they are crucial to successful deployment of Semantic Web applications. The proof layer involves the actual deductive process, representation of proofs, and proof validation. It allows applications to inquire why a particular conclusion has been reached, ie. they can give proof of their conclusions. The trust layer provides authentication of identity and evidence of the trustworthiness of data and services. It is supported through the use of digital signatures, recommendations by trusted agents, ratings by certification agencies etc.
C. Recent Trends and Technological Developments
As with any maturing technology, the architecture will not
remain static. In 2006 Tim Berners Lee suggested an update
to the layer cake diagram [
Rules and Inferencing Systems. Alternative approaches to rule specification and inferencing are being developed. RIF (Rules Interchange Format) is a language for representing rules on the Web and for linking different rule-based systems together. The various formalisms are being extended in order to capture causal, probabilistic and temporal knowledge.
Database Support for RDF. As the volume of RDF data increases, it is necessary to provide the means to store, query and reason efficiently over the data. Database support for RDF and OWL is now available from Oracle (although at present the focus is on storage, rather than inferencing capabilities). Other open source products include 3Store5 and Jena6. The specification of a query language for RDF, SPARQL, was adopted by the W3C in 2008.
RDF Extraction. The language GRDDL: ("Gleaning Resource Descriptions from Dialects of Languages") identifies when an XML document contains data compatible with RDF and provides transformations which can extract the data. Considering the volume of XML data available on the web, a means of converting this to RDF is clearly highly desirable.
Ontology Language Developments. The OWL language was adapted as a standard in 2004. In 2007, work began on the definition of a new version, OWL 2 which includes easier query capabilities and efficient reasoning algorithms scaled to large datasets. Even though Semantic Web technology is in its infancy, there are a wide range of applications in existence. In this section we give a brief overview of some typical application areas.
E-Science Applications. Typically e-science describes
scenarios involving large data collections requiring
computationally intensive processing, and where the
participants are distributed across the world. An
infrastructure whereby scientists from different disciplines
are able to share their insights and results is seen as critical,
particularly when we consider the availability of large
volumes of data becoming available online. The Gene
Ontology7 is a project aimed at standardizing the
representation of genes across databases and species.
Perhaps the most famous e-science project is the Human
Genome Project8 which identified the genes in human DNA
and which includes over 500 datasets and tools. The
International Virtual Observatory Alliance9 makes available
astronomical data from a number of digital archives.
Interoperation of Digital Libraries. Institutions such as
libraries, universities, and museums have vast inventories of
materials which are increasingly becoming available online.
These systems are implemented using a range of different
technologies, and although their aims are similar it is a huge
challenge to enable the different institutions to access each
7 http://www.geneontology.org/index.shtml
8 http://www.ornl.gov/sci/techresources/Human_Genome/home.shtml
9 www.ivoa.net
other's catalogues. Ontologies are useful for providing
shared descriptions of the objects, and ontology mapping
techniques are being applied to achieve semantic
interoperability [
Travel Information Systems. The goal of building an application which would allow a user to seamlessly book and plan the various elements of a trip (flights, hotel, car hire etc.) is highly desirable. Ontologies again could be used to arrive at a common understanding of terminology. The Open Travel Alliance is building XML based specifications which allow for the interchange of messages between companies. While this is a first step, an agreed ontology would be needed in order to achieve any meaningful interoperation.
Although many potential applications can be identified, there are less deployed at this time than we might expect. One possible reason is the lack of a common understanding of what the Semantic Web can offer, and more particularly what the role of ontology. At one end of the spectrum we find applications which take the "traditional", or AI view of inferencing, in which accuracy is paramount. Such applications arise in combinatorial chemistry for example, in which vast quantities of information on chemicals and their properties are analysed in order to identify useful new drugs. By coding the required drug's properties as assertions will reduce the number of samples which need to be constructed and manually analyzed by orders of magnitude. In cases such as these, the time taken to perform the inferencing is less important, since the trade-off will be a large reduction in the samples to be analyzed.
At the other end of the spectrum, we have "data centric" web applications which require a swift response to the user. Examples of this type of application include social network recommender systems such as Twine10 which make use of ontologies to recommend their users to other individuals who may be of interest to them. While it is clear that a response must be generated for the user within a few seconds, we can observe too that there can be no logical proof of correctness and soundness of the answers generated in this type of case! Accordingly, the level of inferencing required in this type of application is minimal.
V.
While we have stated that the Semantic Web focuses on data in contrast to the document centric view of the traditional web, this is not the complete picture. In order to realize value from putting data on the web, links need to be made in order to create a "web of data". Instead of having a web with pages that link to each other, we can have (with the same infrastructure) a data model with information on each entity distributed over the web.
The Linking Open Data [
In a recent talk at the TDC 2009 conference11, Tim Berners Lee gave a powerful motivation example for the project: scientists investigating the drug discovery for Alzheimer's disease needed to know which proteins were involved in signal transduction and were related to pyramidal neurons. Searching on Google returned 223,000 hits, but no document provided the answer as nobody had asked the question before. Posing the same question to the linked data produces 32 hits, each of which is a protein meeting the specified properties.
At the conception of the project in early 2007, there were a reported 200,000 RDF triples published. By May 2009 this had grown to 4.7 billion [dh]. Core datasets include
containing over 274 million pieces of information.
The knowledge base is constructed by analyzing the different types of structured information, such as the "infoboxes", tables, pictures etc.
The DBLP Bibliography, which contains
bibliographic information of academic papers,
Geonames, which contains RDF descriptions of 6.5
million geographical features.
So where is the Semantic Web? In a 2006 article [
The last five years have seen Semantic Web applications
move from the research labs to the marketplace. While the
use of ontologies has been flourishing in niche areas such as
e-science for a number of years a recent survey by Hendler
[
As we have discussed, the "chicken and egg" dilemma is resolving itself with tens of billions of RDF triples now available on the web, and this number is continuing to increase exponentially.
Also, it is becoming easier for companies to enter the market of Semantic Web applications. There are now a wide range of open source applications such as Protégé13 and Kowari14 which provide building blocks for application development, making it more cost effective to develop Semantic Web products.
Some observers argue that the Semantic Web has failed to deliver its promise, arguing instead that the Web 2.0 genre of applications signifies the way forward. The Web 2.0 approach has made an enormous impact in recent years, but these applications could be developed and deployed more rapidly as their designers did not have the inconvenience of standards to adhere to. In this article we have demonstrated the steady infiltration from the research lab to the marketplace being made by the Semantic Web over the last decade. As the standards mature and the web of data expands, we are confident that the Semantic Web vision is set to become a reality.
Gruber, T. 1993. Toward principles for the design of ontologies used for knowledge sharing. In Guarino N, Poli R (eds). International Workshop on Formal Ontology, Padova, Italy, Hendler, J., 2008. Linked Data: The Dark Side of the Semantic Web, (tutorial), 7th International Semantic Web Conference (ISWC08), Karlsruhe, Germany.
Linking Open Data Wiki, available at
http://esw.w3.org/topic/SweoIG/TaskForces/CommunityProjects/LinkingOpenData
Manning, C., Schütze, H., 1999. Foundations of statistical natural
language processing. MIT Press.
[
http://www.w3.org/2000/Talks/1206-xml2k-tbl/slide10-0.html.
W3C. Exploiting Wikipedia as a Knowledge Base: Towards and Ontology of Movies
Abstract. Wikipedia is a huge knowledge base growing every day due to the contribution of people all around the world. Some part of the information of each article is kept in a special, consistently and formatted table called infobox. In this article, we analyze the Wikipedia infoboxes of movies articles; we describe some of the problems that can make extracting information from these tables a difficult task. We also present a methodology to automatically extract information that could be useful towards the building of an ontology of movies from Wikipedia in Spanish. 1 Introduction
Wikipedia is a free encyclopedia of open content that has become an important resource towards the construction of the Semantic Web. Since it beginnings, in the year 2001, the English version has achieve more than 2 million of articles, while the Spanish version has around 480 thousand of articles. All of the content has been written and edited by volunteers from different countries in many different languages, and it is covered by GFDL (GNU Free Document License), which makes possible to freely use them.
One important thing about the structure of Wikipedia is the social control executed by the community, which is able to avoid the spam, the nonsense and other kind of vandalism that is recurrent on some media sites. Besides, this same control makes possible to constantly increase the quality and precision of the articles.
studies1, where it is possible to see the growth of academic interest in this
encyclopedia. This interest is related to the use of Wikipedia on different academic
studies and as a knowledge base for developing specific tools. On one hand, to
mention a few, some works have focused on the social theme that represents
Wikipedia [
In this paper we present a work in process for the elaboration of an ontology of movies from Wikipedia on Spanish language. First we will briefly present an overview of some studies related to the use of Wikipedia for semantic relation extraction and ontology construction (2). Then we will explain the first step towards the elaboration of an ontology of movies (3). This step includes: a) the description of the so-called infobox, which is part of each movie of Wikipedia and contains specific data about the film (3.1); b) the specific relations to automatically extract (3.2); c) and our proposed XML schema to represent these relations (3.3). Finally, we will discuss our preliminary results and present the future work (4). 2
Wikipedia as a Semantic Knowledge Base
different purposes. As we have mentioned before, one of this interest is the extraction of semantic information that could be helpful on the process of giving more meaning to the Web. In Wikipedia, the meaning could be seen as the knowledge about things represented in different ways: definitions, descriptions, images, numeric data, etc. Furthermore, the meaning of each concept explained on the encyclopedia is related to the meaning of other concepts, which becomes a helpful semantic network to understand concepts on the field where they belong.
semantic information between concepts. A general overview of how Wikipedia could
be used to extract concepts, relations, facts and descriptions can be found in [
In [
The research presented in [
semantic relation extraction or ontology construction from Wikipedia. Our main purpose is to state both the interest that has woken up in the area of extraction and organization of semantic information, and some of the automatic analyzes and procedures that are possible to develop taking into account Wikipedia’s structure. Nevertheless, as we will see in this paper, this structure is often not well organized and makes it difficult to implement automatic processes. 3
Towards an Ontology of Movies In order to develop an ontology of movies we have stated three main steps that can lead us to our purpose. The first one is to collect our input corpus from Wikipedia movies articles and the analysis of the infobox structure on them. After that, the second step is the delimitation and automatic extraction of specific semantic information. Finally, as a third step we consider the implementation of the extracted information into a XML schema that will conform the basis for another later annotation schema. 3.1
Movies infobox structure The first step of our methodology was to conform a corpus from the articles of the films by year category. We use the categories tree option to find a list of the movies titles from the year 1892 to 20082. After that, we use the export pages option to retrieve all the articles of this list. We found a total of 5,561 articles, where the opening and closing infoboxes tags ({{Fields…}}) was on 5,092 cases. This late number represents the total of articles from our corpus.
Wikipedia to summarize and group the information about specific data on some articles. In general words, its purpose is to make the information on a more available format and it can be use as a resource to other applications.
are consider as required: film title and original title. The infobox will be framed in {{Fields…}}, and each field inside will be preceded by a vertical bar “|” and followed by an equal sign “=” and the specific information. Fields without descriptions will remain empty after the equal sign. That means it will have the following structure: | Field = description of the field
| genre = Science fiction
Table 1. Infobox template in Spanish. Fields título original título índice imagen nombre imagen dirección dirección2 dirección3 dirección4 dirección5 dirección6 dirección7 dirección8 dirección9 ayudantedirección dirección artistica producción diseño de producción guión música sonido edición fotografía montaje vestuario efectos reparto país país2 país3 país4 estreno estreno1 género duración clasificación idioma idioma2 idioma3 idioma4 productora distribución presupuesto recaudación precedida_por sucedida_por imdb filmaffinity sincat
can introduce. We see information about dirección (direction), estreno (premiere), idioma (language, language2, language3, etc.), as well as país (country country2, country3, etc.), IMDb (Internet Movie Data Base) or Filmaffinity links (external Web sites with movies information).
Furthermore, one of the problems presented in the infoboxes is the lack of standardization. Some of the elements established by Wikipedia are written in an indistinctive way by the authors of the articles, while others have typographical errors. For example, the field dirección (direction) appears also as director (director); the field título original (original title) can be found as título en España (title in Spain), título principal (main title), título traducido (translated title), among others. More complicated is the case of estreno (premiere), which presents variations like año (year), fecha (date), fecha de estreno (premiere date), or primera emisión (first emission).
Table 2. Infobox template in English Fields name image image_size caption director producer writer narrator starring music cinematography editing studio distributor released runtime country language budget gross preceded_by followed_by
template. It is important to notice the fact that most of another languages follow a similar structure like the one described for English. There is a similar template to the English movies infobox in French Wikipedia, with some added elements like format, awards, and IMDb. In Italian, the infobox determines general fields for different genres of films: generic, animation or film a episodi (films conformed from several short films), with specific fields for each genre; while in German, the fields specifies a more generic data, i.e., title, original title, producer or cameraman.
premiere. There are also coincidences in other fields, for example music and photography. Between English and Spanish there is a coincidence in preceded_by and followed_by. Furthermore, in Spanish, as well as in French, there is the field of IMDb, while Italian or English do not include. However, in English links to IMDb or Allmovie can appear within the article as external links and not inside the template of the infobox. These external links are also a valuable information to extend the semantic data for an ontology, as they can add more information about the films that does not appears in Wikipedia, or be used to complete the empty fields of the infoboxes. Nevertheless, there is also no consistence between the occurrences of the tags with external links. In our corpus, the IMDb tag occurs approximately in the 80% of the articles, while Filmaffinity occurs around in the 5%. 3.2 Extracting specific relations data
exploited with relative easiness. We decide to automatically extract the title, original title, director, premiere year and genre, in order to generate a database with all of this information. Although, not all of this information is present on all the movies articles founded in the films by year category.
director field, we found it with complete information in the 5,092 occurrences of the articles with infoboxes, however the field genre occurs only in 4,499 of these cases. Taking into account that the inconsistencies of the metadata make more difficult the process of automatic relation extraction from the films information, we achieve to obtain the data through the process we hereby describe.
although the field name from many of them was not the same and a review had to be made in order to compile a list of ad hoc synonyms for searching this specific field. The synset was formed by dirección (direction), director (director) and dirigida (directed). Also, after the equal sign that should follow the name of the field, the kind of following blanks was not always the same. Sometimes there were tabs; some other, more than one simple space; and even other, without spaces. Many of the director’s names are also entries of Wikipedia, so many users decided to establish links to their names, using the symbol “[[” followed by the name of the director and closing with “]]”. This has the purpose of specifying to the wikiengine that there is a link: [[link to the article]]. But not all of them had those brackets, and it caused troubles while parsing the data with the aim of recovering the director’s name of the film associated with the title of the entry.
movie. Despite the fact that this field does appear in all the infoboxes, not all of them appear with information, which means that there are articles with the original title field empty. It does not contain information in 195 articles occurrences in the corpus.
most of the films had different words to express the premiere year, for example año (year), fecha de estreno (premiere date) or *añoacceso (acces year). In this case we decide to mine only the año (year) and estreno (premiere) variants, because of the wide range of structural possibilities. We found that 23 films infoboxes do not contain a premiere year, sometimes it was in the title and sometimes were completely absent.
Table 3. Numerical data found over the analysis of infoboxes
Director Título Título ID Título orignal Año Género Director Título
last one in more tan 500 cases. It is important to mention that the title of the movies was not obtained from the infobox but directly from the XML given by the Wikipedia, mainly because it is well demarcated by the labels <title> </title>; in the same way, we obtained the id used by the Wikipedia to identify each article.
3.3 Proposed XML schema With the data from the infoboxes that were exploited, we decided to generate a first XML scheme, which should give basic information about the film. This scheme can be expanded as we extend our extraction processes of the information contained in the
first tag will consist of the director’s name. Taking into account that directors can have more than one film, we decided to introduce a filmography tag to include them. This last tag will include each film with title, original title, year and genre tags. On the opening film tag we added an attribute with Wikipedia’s title id number. An example of the schema can be seen below.
Proposed XML schema for the organization of movies data on Spanish Wikipedia.
As we can see in this example, the root tag is <director></director>. It is followed by the director’s name tag <name></ name >. At the same level there is the tag <filmography></ filmography>. This tag nests the film tag <film wiki_id=“”></film>, which contains the relevant information of each film: <title></title>, <original_title></original_title>, <year></year> and <genre> </genre>.
information that we have considered at this first stage of the ontology construction. As we have said, this is not the full final scheme because as more data is extracted, the more can be added. This scheme is currently based on Wikipedia films articles in Spanish language, however it can be extended to fit another kind of relevant information, for example the country, external links (IMDb) or the id of directors or genre from Wikipedia. Furthermore, it will be possible to use this scheme in order to exploit Wikipedia in other languages, which could make possible to fill the empty fields in one language by relating them with the information on another languages, as well as to make multilingual queries. 4 Conclusions and future work Nowadays, Wikipedia can be explored with the aim of obtaining information on different ways. The information added in a manual way by the users is generally well organized and semi-structured. Also, many entries from Wikipedia have infoboxes with summarized specific information about the theme treated in the article. We have mentioned that the structure of Wikipedia has made possible to exploit the information in order to extract semantic data. The extraction of semantic relations is one of the growing interests aiming to the construction of the Semantic Web.
automatically exploiting it. To summarize a few, there are: a) the fact that the field’s names are not respected; b) typos by human errors; c) lack of information; and d) differences on the infobox structure between languages. The latter should not be seen as a problem, however it would be advantageous to have standard fields on different languages.
of writing or editing an article use a check-bot to confirm the information of the infoboxes templates. Thus, the fields not belonging to the template would be alerted, as well as typos on the field names. Furthermore, the same check-bot could be used to seek the existing fields looking for inconsistencies in the infoboxes or the whole articles.
The work that we have presented here is a first approach towards the elaboration of an ontology of movies from the Wikipedia in Spanish. We have showed the kind of semantic relations that are possible to mine, as well as a first scheme to represent them. We are conscious that this scheme may well be improved for achieving a complete ontology of movies. The future work will include: a) to define a scheme to represent subject, relation and predicates between the extracted information, for example a RDF scheme; b) to implement this new scheme for making the information available and sharing it with systems dedicated to the construction of the Semantic Web; c) to develop a movie-ontology query system capable of retrieve the information on specific ways related to directors, titles, genres and years fields. Acknowledgments This research was made possible by the financial support of CONACYT (82050) and DGAPA-PAPIIT (IN403108). The authors wish to thank Sarahi Abrego Romero for the proofreading of this paper. 3.1 Algortihm K-means Advantajes.
MacQeen J. [
The process, which is called “k-means”, appears to give partitions which are reasonably efficient in the sense of within-class variance, corroborated to some extend by mathematical analysis and practical experience. Also, the k-means procedure is easily programmed and is computationally economical, so that it is feasible to process very large samples on a digital computer.
Likewise [
Taking as a framework and as an extension and update, the k-means shortcomings
that are identified in [
According to [
Authors
[
Nayeripour.
[
Title and Commentary “K-means Clustering Algorithm with improved initial Center.” It avoids the initial random assignment of centers. Use strategy called "sub-merger" “A New Evolutionary Algorithm for Cluster Analysis”.
It not depend on the initial centers. Algorithm PSO-SA-K
combines the algorithms "Particle Swarm Optimization (PSO),"
Simulated Annealing "(SA) and K-means.
“Local vs global interactions in clustering algorithms: Advances
over K-means.” It focuses on the algorithm's sensitivity to initial
conditions. Incorporate information on the role of overall
performance. Define three new algorithms: Weighted k-means
(WK), Inverse Weighted K-means (IWK) and Inverse Exponential
k-means (IEK).
[
Rouseeuw.
[
Hybrid techniques called K-NM-PSO-based K-means, NelderMead Simplex search and optimization of exchange of particles. “Enhancing K-Means Algorithm with Initial Cluster Centers Derived from Data Partitioning along the Data Axis with the Highest Variance.” Title explicit.
A method for initialising the K-means clustering algorithm using
kd-trees” A kd-tree used to calculate an estimate of the density of
data and to select the number of clusters.
“Analysis of Global k-means, an Incremental Heuristic for
Minimum Sum of Squares Clustering”. Commentary on work
[
L. Kaufman and P. Rouseeuw. Finding Groups in Data: An Introduction to Cluster analysis: Text. Def. K-means. “A clustering technique for summarizing multivariate data”, (ISODATA). Perform dynamic estimation of K. 3.2.2 The convergence of algorithm to a local optimum rather than a global optimum.
According to [
Authors
[
Rosy Basave T. Héctor
Fraire H.
[
And Tung A.K.H.
Títle and Commentary “Local vs global interactions in clustering algorithms: Advances over K-means.” Addresses the algorithm's sensitivity to initial conditions. Incorporating global information on the performance function. Define three new algorithms: Weighted k-means (WK), Inverse Weighted K-means (IWK) and Inverse Exponential kmeans (IEK). “Improvement the Efficiency and Efficacy of the K-means Clustering Algorithm through a New Convergence Condition”.
Improvement to the k-means algorithm by new convergence
conditions. Experimentally analyze the local convergence of
kmeans.
“On the Lower Bound of Local Optimums in K-means
Algorithm.” Estimate lower limit for local optimum.
[
“Genetic K-means algorithm”. Hybrid scheme based on Genetic Algorithm - Simulated annealing with new operators to perform global search and rapid convergence. “Some Methods for Classification and Analysis of Multivariate Observations.” Definition, Analysis and Applications of kmeans. 3.2.3 The efficiency of the algorithm.
According to the work of [
Authors
[
Sohler.
[
Títle and Commentary “New Efficient Strategy to Accelerate k-Means Clustering Algorithm”. Strategy to accelerate k-means algorithm, which avoids many calculations of distance, through a strategy based on an improvement to the partial distance algorithm (PD). “An Efficient k`-means Clustering Algorithm.” Based on the algorithm Rival, it penalizes competitive Learninig (RPCL). It does not require pre-allocation of the number of clusters. Two-step process. Pre processes and uses the prior information to minimize the cost function. “GAKREM: A novel hybrid clustering algorithm.” Eliminates the need to specify a priori the number of clusters.
Combines genetic algorithms and logarithmic regression
Màxima expectation.
“A Fast k-means implementation using coresets.”
Implemented version of Lloyd's k-means [
K-means based on distance L1. Proportionate restrictions incorporated in the calculation of centroids. “Developing a feature weight self-adjustment mechanism for a K-means clustering algorithm.” Improvement of the quality of k-means clustering via FWSA mechanism called "SelfAdjusment Feature Weight." Is modeled as an optimization problem. “Improvement the Efficiency and Efficacy of the K-means Clustering Algorithm through a New Convergence Condition”.
Improvement to the k-means algorithm by new convergence
conditions. Experimentally analyze the local convergence of
kmeans.
“Improvement of the K-means algorithm using a new
approach of convergence and its application to databases
cancer population." Title explicit.
[
According to [
Authors
[
Maleki, A.
[
Hall
Títle and Commentary “Geodesic K-means clustering”.Extends k-means by using a geodesic distance metric. Algorithm ensures resistance to outliers. “A fast and robust general purpose clustering algorithm.” It eliminates the effect of outliers through a process that considers real points as centroids. “A clustering technique for summarizing multivariate data”.(ISODATA). It performs dynamic estimation of K. Considers the effect of outliers in the process of clustering. 3.2.5 The definition of “means” limits the application only to numerical variables.
Several works have been developed that extend the application of categorical variables or others: k-means for
Authors
[
Rao, &Bhatnagar
[
Títle and Commentary “Genetic Algorithm for text clustering using ontology and evaluating the vality of various semantic simility measures.” Improving the kmeans algorithm by using a genetic algorithm that finds similarities conceptual. Based on ontology, thesaurus corpus for clustering of text fields. “K-means clustering algorithm for categorical attributes”. Title explicit. “Extensions to the k-means algorithm for clustering large data sets with categorical values.” Title explicit. 4 The Algorithm k-means on Matlab.
Experimental tests were conducted for K-means in the Matlab [
Tests for k-means in Matlab, used the well known UCI Machine Learning Repository
[
4 6 0.5
1.5 1 2 2.5
Fig.1 Representation of the Iris Data Set
For the experimental tests carried, the following data sets have been used: Iris, Glass
and Wine. This report presents the results for the Iris data set.
The Iris Data Set is a database of types Iris plant, which has No. of instances: 150
(50 in each class), No. of attributes: 4 (Sepal length, Sepal width, Petal Length, Petal
width) No. of classes: 3 (Hedges Iris, Iris versicolor, Iris virginica). One class is
linearly separable from the other two; the latter are NOT linearly separable from each
other. Based on data and classes defined in [
Test I4: >> [u,v,sumd,D]= kmeans(z,3,'display','iter'); The “Test I4” is an example of the test results on Matlab and kmeans for the Iris data set. Iter column represents the number of iteration, the phase indicates the algorithm phase, num provides the number of exchanged points, sum, provides the total sum of distances, inter% are the percentage of exchanged points in each iteration. 100 90 80 70 trce 60 n I .tso 50 P de 40 % 30 20 10 00 2 4 10 12
14 Fig. 2 % of Exchanged Points. Fig. 2 corresponds to the test I4 and the graphical representation of the exchange behavior at each iteration. Likewise Fig. 3 represents the behavior of the sum of distances for the same test. 350 t. isD300 e d uam250 S e lda200 t o T 150 100 500 2 4 Table 1 Summary of Results for Iris Data Set 5.1 Summary of Results for the Iris data set.
With regard to exchanges between groups that the algorithm makes in the tests conducted, it was observed that the most significant changes occurred from first to second iteration. For all cases, in the first step all points are located (100%), the third column includes the number of points to be exchanged in the second iteration and the fourth column is the percentage difference in the number of items exchanged between the first and second iteration.
According to the results in Table 1: It can see that for 150 points in the Iris database, and a set of 25 tests, the k-means algorithm in Matlab: Converge in an average of 7.2 iterations.
The average number of points exchanged during the second iteration was 18.84 points The percentage of points located on the second iteration in the corresponding group was 91.0% 6 Conclusions.
The results of the analysis for our sample work, allow us to establish a framework and analyze the theoretical study of the k-means algorithm Also we reseach and distinguish the different lines on which there is still a fertile field for investigation. As we can see several attempts at overcoming the shortcomings of the k-means algorithm have been done and different approaches in different disciplines have been proposed: Optimization, Probability and Statistics, Neural Networks, Evolutionary Algorithms, among others. The vast majority of contributions have focused on the first three lines of research identified in this study: The sensibility of the algorithm to initial conditions, convergence of the algorithm to a local optimum rather than a global optimum and the efficiency of the algorithm. Notes that challenges still to be resolved in such research and has been relatively little work done on the lines related to the implementation of the algorithm to other variables as well as treatment to outliers and noise.
Acording the tests conducted in Matlab, this laboratory showed that it is actually very
conducive to experimental testing, the implementation of k-means, allows to monitor
the performance of the algorithm through the information that can be deployed at
runtime, such as result of the objective function and the number of points exchanged
in each iteration.The results allow us to establish a framework to compare the
proposal improvement algorithm with the previous work. As part of this project and
to give continuity to previous work [
GAF-SVM Sergio Manuel Dorantes, Manuel Martín Ortiz, María J. Somodevilla, Jesús Lavalle Martínez, Ivo H. Pineda Torres
Facultad de Ciencias de la Computación, BUAP sergiomanuel@hotmail.com, {mmartin, mariasg, jlavalle, ipineda}@cs.buap.mx Abstract. Due to the amount of visual information that currently exists, there is a need to classify it properly. In this paper we present an alternative dual method for image categorization according to their texture content defined as GAF-SVM, this method is based in the use of Gabor Filters (GAF) and Support Vector Machine (SVM). To perform the image classification we rely on filtering techniques for feature extraction mixed with statistical learning techniques to perform the data separation. The experiments were carried out by taking a set of images containing coastal beach scenes and a set of images containing city scenes. A feature vector is obtained from applying a bank of Gabor Filters to the input images; the output feature space is then used as an input to the SVM Classifier. The Support Vector Machine is responsible for learning a model that is capable of separating the sets of input images. Experimental results demonstrate the effectiveness of the proposed dual method by getting the error classification rate to near 9%.
I. INTRODUCTION
The proposal for an alternative method of image classification requires analysis of
the methods presented so far concerning the area. Extracting visual information from
an image to obtain their most important features is essential for classification tasks;
over the years various approaches have been presented regarding this field of study
such as: color histograms, region-based classification and gray-level values of raw
pixels; though one solution has been to incorporate the texture analysis as main
feature descriptor. This is largely due to the fact that most surfaces on images contain
some kind of texture. In the recent years, texture analysis has been used for object
recognition, image interpretation, image segmentation and classification [
In recent papers like [
But bring out a formal definition of what a texture really is, it became a subjective
topic. As it was mention in [
Our proposal is based on the use of natural textures in real world images, for that reason the classification model must deal with more complex images in natural conditions.
The 2-D Gabor filters (2D-GF) have certain properties that make them suitable for
textural identification in many ways: 2D-GF have tunable orientation and radial
frequency bandwidths, tunable center frequencies, and optimally achieve joint
resolution in space and spatial frequency. The demodulated Gabor channel envelopes
generally contain only low spatial frequencies which are optimally localized in both
domains [
Gabor filter based methods have been successfully applied for a variety of machine
vision application, such as texture segmentation [
This paper is organized as follows: in section II it is mention the related work we based on to develop this article, in section III it is made a detail description of the proposed method, in section IV it is an explanation of the way the input data is processed as well as the Gabor filter’s parameters selection, in section V the details of the SVM classifier parameters, and in section VI the experimental results. II. RELATED WORK The classification of images has been studied from various approaches, most of all through the mixing of methods, one for texture extraction and one for the classification process.
In [
The first step to complete the proposed method is to extract the texture features with a bank of Gabor filters applied to each input image, and then take the filter’s output to form a training dataset to feed the SVM classifier.
III. PROPOSED METHOD
In order to accomplish the image classification we rely on filter based techniques to
perform texture feature extraction mixed with statistical learning theory techniques to
achieve the image data separation. Gabor filters were selected to extract texture
features from images due to their resemblance to the human visual system [
A. Gabor Filters
A number of authors have used a bank of filters to extract local images features [
A 2-D Gabor filter is a linear filter whose impulse response is defined by a harmonic function multiplied by a Gaussian function. In the spatial domain can be defined as follows: ψ (x, y) =
f 2 e−(γf 22 x′2 +ηf 22 y′2 ) πγη
⋅ e j2π fx′ x′ = x cosθ + y sinθ y′ = −x sinθ + y cosθ (1)
Where f is the central frequency of the filter,θ the rotation angle of the Gaussian major axis and the plane wave, γ the sharpness along the major axis, and η the sharpness along the minor axis (perpendicular to the wave). In the given form, the aspect ratio of the Gaussian is λ=η/γ. This four parameters (f,θ,γ,η) define the shape of the filter, and by changing them we can detect different textures.
The normalized 2-D Gabor filter function has an analytical form in the frequency domain. u′ = u cosθ + v sinθ v′ = −u sinθ + v cosθ
Filter Design vs. Filter Bank
There exist two aspects regarding the implementation of Gabor filters, on one hand
the filter bank approach and in the other hand the filter design approach [
A filters bank allows the analysis of an image in a single pass way at several
frequencies and in several orientations at once. According to the characteristics of our
model, the use of a filters bank is the solution choice of deployment, although it could
mean an increase, in the computational processing, this is not significant. The design
of a Gabor filter bank consists, in general, in the selection, for each filter, of the
proper values of the following parameters: frequency, orientation, γ andη, the last two
parameters known as the smoothing parameters [
In this research it is defined a bank with up to 3 orientations and up to 2
frequencies, resulting in a bank with maximum 6 output filters, allowing us to
accurately detect a texture among a large set of images. This decision was made based
on the studies presented in [
Using many different orientations and scales (frequencies) ensures invariance;
objects and some textures can be recognized al various different orientations, scales
and translations [
C. Support Vector Machines
Support Vector Machines (SVM) were introduced by Vapnik as a powerful learning
tool based on statistical learning theory, a Support Vector Machine is a binary
classifier that makes its decision by constructing a linear decision boundary or
hyperplane that optimally separates data points of the two classes in feature hyper
space and also makes the margin maximized [
SVM starts from the goal of separating the data with a hyperplane, and extend this
to non-linear decision boundaries using the kernel trick [
Where x represents a point (a vector), w represent the weight (also a vector). We want to choose w and b to maximize the margin, or distance between the parallel hyperplanes that are as far as possible while still separating the data. The hyperplane must separate data such as:
wT x + b = 0
T w xk + b > 0 for all xk of a class y
T w x j + b < 0 for all x j of another class
If data are separable in this way, there will probably be more than one way to do it.
Among all the possible existing hyperplanes, SVM selects the one in which the
distance between the hyperplane and the closest data is the widest possible [
When working with a dataset that is not linearly separable, it is necessary to turn to
the use of a kernel function. The kernel function allows the SVM to form non-linear
boundaries [
Now instead of considering the input vectors {x1,…, xn} it is considered the transformed vectors {φ(x1),…, φ(xn)} as shown in figure 1. By doing this substitution, it is obtained a SVM raised in a new space (this is called the ‘kernel trick’), it is important to mention that in practice the implementation of this nonlinear technique consumes the same amount of computational resources of its linear equivalent. Fig. 1. Using the Kernel to transform (map) the input data space.
The general problem that SVM want to resolve is to search, for a given learning
task, with a finite amount of data, an appropriate function that helps to carry out a
good generalization, which results from a proper relationship between the accuracy
achieved with a particular training set and the ability of the model [
The use of the ‘Radial Basis Function’ (RBF) kernel is based on the fact that this kernel is basically suited best to deal with data that have a class-conditional probability distribution function approaching the Gaussian distribution, like the texture present on the input images. It maps such data into a different space where the data becomes linearly separable. The kernel function is defined as follows:
A disadvantage concerning this kernel is that is difficult to design, in the sense that it is difficult to obtain an optimum value for its parameter σ (sigma) and choose the corresponding C that works best for a given problem. The fact that certain combinations of σ and C make the SVM highly sensitive to training data also contributes to the error rate of the RBF-based SVM.
One of the advantages of the RBF kernel is that given the kernel, the weights, the number of support vector and the support vectors itself are automatically obtained as part of the training procedure, i.e. they don’t need to be specified by the training mechanism.
IV. SETTING THE EXPERIMENTS As part of the experiments it was decided to work with two sets of images, one set consisting of coastal beach scenes, and the other set consisting of city scenes images. The processing of input images is done in order to reduce computational complexity.
The first set is conformed by 128 images of beach scenes content, the second set is conformed by 128 images of city scenes content, a total of 256 images.
The input images after processed end up being 8-bit per pixel grayscale images of dimension 128x128, working with just one channel reduces de number of convolutions. The output of each filter is obtained by the convolution of the input image with a Gabor filter. The process is shown below:
G(x, y) = I (x, y) ⊗ψ (x, y) (7) where G(x, y) is the output of the filter I (x, y) is the original image ψ (x, y) is the Gabor filter
This computation can theoretically be done in the spatial domain however the
Gabor filter is usually narrow. The filter is usually much larger in the frequency
domain and thus less affected by aliasing effects due to sampling. It is thus more
convenient to do all the computation process in the frequency domain. The
convolution is then reduced to a simple and efficient point-wise multiplication of the
Fourier transforms [
The family of Gabor filters selected to set up the filter bank for the experiments in the frequency domain are: v′ = −u sinθ + v cosθ
I (u, v) = ∫−∞∞ ∫−∞∞ i(x, y)e−i2π (ux+vy)dxdy where i(x, y) is the original input image (11)
After the transformation, normalization is applied to the output image in order to avoid effects by illumination.
At the end of normalization, we have a certain number of square matrices per filtered image; each matrix dimension is 128x128. The number of square matrices depends on the number of orientations and frequencies concerning the filters bank, in our experiments the filter bank consist of 2 frequencies and 3 orientations, so the number of output matrices is 6.
The convolution of the input image with the Gabor filter is performed. In domain frequency the convolution is represented in a point-to-point multiplication of the transformed image with the Gabor filter.
Once the filter output is obtained, G(x,y) needs to be transformed back to its spatial representation using the Inverse Fourier Transform in 2-D.
g(x, y) = i(x, y) ⊗ψ (x, y) - Spatial Domain G(x, y) = I (x, y) ⋅ Ψ(x, y) - Frequency Domain i(x, y) = ∫−∞∞ ∫−∞∞ I (u, v)ei2π (ux+vy)dudv where
I (u, v) is the image in frequency domain
When convolution is performed some results are not useful especially if the image does not contain textures that respond meaningfully to the filter selected parameters. To reduce the problem all the outputs obtained by convolution are summed up to remove the results that are not relevant and to enhance those that helps to detect texture regions; this also helps to reduce dimensionality of the feature space by having one square matrix as an output, same size of the input image.
At this point we have one matrix per input image, reducing the dimensionality of input data. Each matrix is used to build up a feature matrix, which is going to serve as an input of the SVM classifier.
To complete the convolution of the input image with each one of the Gabor filters
we take only the real part of the output filtered image. As mentioned in [
Re(G(x, y)) (12)
Then we modified each output matrix of dimension 128x128 to construct the feature matrix. We take each matrix and transform it in a 1x16384 vector, each vector is then piled up with the next transformed matrix to form the feature matrix.
Finally we have a feature matrix of dimension 256x16384 which serves as input to the classifier.
V. SVM CLASSIFIER The goal of experimentation is to obtain a training model through SVM, which can be capable of separate a set of input images. Once we have the feature matrix with processed and filtered images we proceed with the SVM classification procedure. According to the nature of the classification process we need to define a training dataset, so the classifier could learn a model, and a test dataset, that let us test the learned model. The training dataset is conformed by 75% of the input dataset, and the test dataset by the remaining 25%. The selection criteria to build up the training dataset and the test dataset are done randomly. In fig. 2(a) is shown an example of coastal beach scene images, and in fig 2 (b) an example of city scenes images, which the classifier will try to separate. Fig. 2. Example of images used in the experiments. (a) Beach scene images, (b) City images.
The experiments were performed using SPIDER [
In table I there is a list of kernels available in SPIDER, its formula and its parameters. k (x, y) = (x ⋅ y + 1)d
− x− y 2 k ( x, y) = e 2⋅σ 2 k ( x, y) = It is used the “RBF” kernel to execute the experiments with different sigma values. Another parameter used by SPIDER is the ‘soft margin parameter’, C, which penalizes the training errors. This value is set to 1000 in all the experiments.
Iteratively, the sigma values were changed until a significant error rate is obtained. The test results for the learned algorithm are presented in table II.
As it can be seen in table II the sigma value which represents the lower percentage error is σ = 35, with an error rate of 9.37%.
RBF σ=21 σ=22 σ=23 σ=24 σ=25 σ=26 σ=27 σ=28 error error VII. CONCLUSIONS Extracting texture features by a Gabor filter bank and classify the filter outputs via the Support Vector Machines offers an excellent accuracy rate, 90.63% of the input images are correctly classified according to their class, belonging to beach scenes class or city scenes class.
The article proves the efficiency of using a dual model, first to extract de texture features and then classify them with SVM.
REFERENCES
[
Pattern Analysis and Machine Intelligence, vol 21, Issue 4,, pp. 291 – 310, Apr 1999.
[
Proceedings Of World Academy of Science, Engineering and Technology, Vol. 5, April 2005.
[
Available: http://www.cse.iitk.ac.in/~amit/courses/768/00/rajrup/, last visited: April 20, 2009.
[
Lab, Dept. of Ind. Eng. and Mgmt., Yuan-Ze University, Chung-Li, Taiwan, 2000.
[
Computer Vision, Vol. 1, 2002.
[
University of Technology, Finland, Nov. 2003.
[
Optical Engineering, Vol. 35, pp. 2852 – 2863, 1996.
[
Report 100, Lappeenranta University of Technology, Dept. of Information Technology, 2005.
[
California, 2004.
[