A New Approach for Interactive Image Retrieval Based on Fuzzy Feedback and Support Vector Machine Malihe Javidi 1 , Baharak Shakeri Aski 2 , Hale Homaei 3 , H.R.Pourreza 1 1 Department of Computer Engineering, Ferdowsi University of Mashhad, Mashhad, Iran 2 Islamic Azad University, Ramsar Branch, Iran 3 Islamic Azad University, Mahmoodabad Branch, Iran {malihejavidi, Shakeriaski.b, hale.homaei}@ gmail.com, hpourreza@um.ac.ir Abstract In this paper, we introduce an efficient content- based image retrieval system based on fuzzy relevance feedback. Conventional Content Based Image Retrieval (CBIR) systems that use Relevance Feedback (RF), want user to mark retrieved images as relevant or irrelevant, while this determination is difficult for images which are rich in semantic. As a result, this system integrates the log information of user feedback using a soft feedback model to construct Fuzzy Transaction Repository (FTR). The repository remembers the user’s intent and therefore, provides a better representation of each image in the database. The semantic similarity between the query image and each database image can then be computed using the current feedback and the semantic values in the FTR. Furthermore, the SVM is applied to the session-term feedback in order to learn the visual similarity. These two similarity measures are normalized and combined together to form the overall similarity measure. Experimental results using a COREL database demonstrate the effectiveness of the proposed method. 1. Introduction With the development of the Internet, and the availability of image capturing devices, the size of digital image collection is increasing rapidly and thus efficient image searching, browsing and retrieval tools are required by users. Content-based image retrieval (CBIR) is a process of retrieving a set of desired images from a collection of images based on visual contents present in the images, such as color, texture, shape or spatial relationship. Extensive experiments on CBIR systems show that the retrieval accuracy of today’s CBIR systems remains relatively unsatisfactory [1]. CBIR systems interpret user information needs based on a set of low-level visual features extracted from the images. However, these features may not correspond to the user’s interpretation and understanding of image contents. In order to improve the retrieval accuracy of CBIR systems, the focus of research has been shifted from designing sophisticated low-level feature extraction algorithms to reducing the ‘semantic gap’ between low-level features and high-level semantic concept [1]. So to reduce the gap, different techniques were introduced. Using object ontology to define high-level concepts [2], supervised or unsupervised learning methods to associate low-level features with query concepts [3]-[4] and introducing Relevance Feedback (RF) into retrieval loop for continuous learning of users’ intention [5]-[7] are some of these techniques. Among these techniques, RF is a powerful tool. It was introduced to CBIR, with the intention to bring user in the retrieval loop in order to reduce the semantic gap. In this technique different approaches are used to learn the user’s feedback. A typical approach is to adjust the weights of low-level features [6]-[9]. The re-weighting method considers the discriminating power of different features and enhances the contribution of features that best identify the relevant examples marked by the user. Another method is called Query Point Movement (QPM) [5]-[10]. QPM which improves the estimation of the query point by moving it towards positive examples and away from the negative examples. Recently Machine learning techniques such as SVM are also used for concept learning [11]. SVM is often utilized to capture the query concept by separating relevant images from irrelevant ones. Generally, the labeled samples provided by the user are limited, and such small training data set will result in weak classification of database images (as relevant/irrelevant). In [12], the D-EM (Discriminant- EM) is used to solve this problem. Traditionally, the user is restricted to binary classification to determine whether an image is “fully relevant” or “totally irrelevant”. Therefore, a single CIMCA 2008, IAWTIC 2008, and ISE 2008 978-0-7695-3514-2/08 $25.00 © 2008 IEEE DOI 10.1109/CIMCA.2008.176 1205 Authorized licensed use limited to: Ferdowsi University of Mashhad. Downloaded on September 2, 2009 at 14:38 from IEEE Xplore. Restrictions apply.