A New Content-Based Image Retrieval System Using Parzen Relevance Feedback and KullbackLeibler Divergence

Abstract

In content-based image retrieval (CBIR), many multimedia applications use visual distance to find a collection of images which share the same properties. However, visual distance between two images is often not suitable to semantic distance between the same images. In fact, the semantics term refers to the way how people interpret the image content. Currently, it is difficult to find good correspondences between high-level image semantics and low-level image features which create a “semantic gap”. In this paper, we propose a new relevance feedback method which reduces the semantic gap between images. The key steps of our process are the following: At first, we compute the visual distance through the KullbackLeibler Divergence (KLD). Then, we apply the Relevance Feedback to enhance the retrieval effectiveness by using three different machine learning algorithms: Gaussian Mixture Model (GMM), Support Vector Machines (SVM) and Parzen classifiers; thus, we learn relevant and irrelevant images according to user selection. Experimental results on 5000 images from the COREL database show that comparing to traditional approaches, Parzen classifier is effective and can significantly improve retrieval rates.