Generic and fully automatic content-based image retrieval using color

Abstract

In this paper, we propose a generic and efficient content-based image retrieval architecture. We compute “real” interimage distances for an initial subset of the images that are to be stored into an image database. For computing real interimage distances we use image content based on a low level feature. High-level image feature vectors are computed from the real interimage distances in such a way that the interimage distances are preserved in the feature space defined by the high-level features. These feature vectors are used to represent the images in the initial subset as well as to generate a training set. This training set is used to compute the feature vector of a query image during image retrieval and for deriving the feature vectors for images not in the initial subset. On-line retrieval is performed using the distances of the feature vector of the query image to feature vectors of the database images as estimates of the corresponding real distances. We have conducted experiments using color as the low level feature. Our results show a substantial reduction in the size of the feature space, which leads to highly efficient on-line retrieval. We also demonstrate that a high retrieval accuracy is achieved.