Adapting indexing features to scene structures for relevant image retrieval

Scene structural matrices (SSMs) are 2D tables that can be used conveniently to capture the organization of a scene in a colour image when it has been adaptively partitioned and represented by a simplified binary space partitioning (BSP) tree. In this paper, we present a number of methods to populate the cells of SSMs. Well-established and popular image indexing features are used. Since the SSMs implicitly contain scene organization information and intermediate level concept, indexing the cells with these features is equivalent to adapting them to the structure of the scene. To show the effectiveness of our method, we conducted an experiment to retrieve landscape images, a type of image which can be described using an intermediate-level concept. Our results show that the new methods significantly outperform the traditional methods.