On the problems of using local descriptors for large images

Content-based retrieval in image databases uses a-priori extracted features and performs individual matching with the generated features from the query image. The most similar results determine a final ranking. While this approach is suitable for smaller sized images, which are adequately described by one set of features, it falls short for large scenes, e.g. satellite images. In this case several set of features have to be extracted at different positions within the image. A study showed that the selection of appropriate locations is not straightforward and that more suitable techniques result in a significantly increased computational burden. Therefore this paper proposes the utilisation of feature functions that approximate the expected feature value at any arbitrary image position. To incorporate not just a local context but also a globally dependent description for each location, radial basis functions were used. The approach proved to provide a higher precision in terms of the retrieval and a reduced storage requirement for the features.