Relative entropy-based feature matching for image retrieval

Increased interest in content-based storage and retrieval of images and video frames has been stemmed from its potential applications in multimedia information systems. Various matching methods have been proposed in the literature, including histogram intersection, distance method, and reference fable method. A comparison of these three techniques has proved that the reference table method is the best in terms of retrieval efficiency. However, the drawback of this method is that it requires a pre-defined set of reference feature (color, in particular) which can approximately cover all features (colors) in the selected application. While this condition may be satisfied in some applications, in situations where there are continuing additions and/or deletions to the database and where knowledge of features in the images is not available a priori, such a technique will not produce very reliable results. The reference feature or color table method requires a representative sample of all images stored in the database in order to select the reference feature or color table. For example, such a priori knowledge is impossible to obtain in a trade-marks database. To alleviate the reference table requirement, recent works suggest the use of unsupervised feature matching based on color-clustering, which is a computationally expensive approach. In this study, we propose an image retrieval method based on the relative entropy (E-rel), known as the Kullback directed divergence. This measure is nonnegative and it is zero if and only if two distributions are identical; i.e., perfect match. E-rel has only one minimum for every comparison. This offers a unique criterion for optimization with low computational complexity. It also provides a thoughtful view for the type of data distribution in the sense that the whole range of data distribution is considered in matching and not only some moments. The algorithm described here has been tested on an imaging database system, consisting of 100 various images of different object and texture scenes stored in a content addressable stack. The efficacy of retrieval is presented by listing the retrieval results using different query images. The experimental results show that the relative entropy is effective for ordering the images of a database system in accordance with the similarity between their gray-level distributions.