International Journal of Innovative
Computing, Information and Control ICIC International c°2009 ISSN 1349-4198
Volume 5, Number 3, March 2009 pp. 677—688
SEMANTIC MATCHING AND ANNOTATION OF IMAGES BY SELF-ORGANIZING MAPS
Hsin-Chang Yang1 and Chung-Hong Lee2
1Department of Information Management
National University of Kaohsiung Kaohsiung, 811, Taiwan
2Department of Electrical Engineering
National Kaohsiung University of Applied Sciences Kaohsiung, Taiwan
Received February 2008; revised July 2008
Abstract. Image retrieval has attracted lots of attention from both researchers and practitioners. Different methodologies as well as commercial systems have been proposed and developed to tackle this task. Most of these systems are based on one or both of two major image retrieval schemes, namely annotation-based image retrieval and content-based image retrieval. The former is simple and accurate, provided some annotations have been added to the images. However, such annotations are often missed in most of images available in large datasets such as the WWW. To tackle this deficiency, we pro-pose a method that could automatically annotate images with some keywords that could feasibly describe the semantics of the images. A set of training images as well as their annotations are trained to find the relationships between images as well as between key-words. New image could then be annotated and retrieved according to such relationships. Our preliminary experiments suggest promising result in both image annotation task and image retrieval task.
Keywords: Image annotation, Image retrieval, Self-organizing map
1. Introduction. Recently the task of image retrieval has received a great deal of at-tention from the web community since there are so many useful images on web pages. In fact, image retrieval has been studied for decades in library science and computer sci-ence communities. Image retrieval is a branch of information retrieval whose task is to retrieve some pieces of information (the documents) to meet a user’s information needs according to certain (semantic) relevance measurements. Currently most information re-trieval systems retrieve documents based on their ’contents’. That is, they measure the relevance between the query and a document according to internal representations or de-rived features. Such representations or features will vary for different document styles and retrieval schemes. For text retrieval systems, the contents are often represented by a set of selected keywords that are intended to capture the semantics of the documents. Many studies have successfully represented the semantics of text documents [1]. For im-age retrieval systems, the representation of imim-age content generally fall into two types. The first is to represent an image with a set of keywords that could describe the content of the image. Such keywords are used as annotations for an image and thus we may call this type of representation scheme the annotation-based image representation and call the task of retrieval by annotation the annotation-based image retrieval. The other rep-resentation contains a set of visual features extracted from the image that hopefully may
