Training templates for scene classification using a few examples

We investigate a method for extracting simple, flexible, relational templates that capture the color, luminance and spatial properties of classes of natural scene images from a small set of examples. We have built an interactive system that allows the user to build his own templates by selecting a set of example images and letting the system extract a set of templates that capture the common relations within the set of images. Given a small set of example images, the classification system works by automatically building flexible templates that define color, luminance and spatial relations between image patches in these examples. These extracted templates can be matched against the entire database to obtain other images that belong to the query-class. The system also gives the user the option of using the results of the initial classification in order to refine the query and perform a more selective search if needed. The system uses very low resolution images to extract the templates from a set of examples. It has been used successfully to retrieve a few classes of natural scenes from the COREL photo library. Our experiments show that the algorithm is fast, requires little storage and works reliably in this domain