Decoding image semantics using composite region templates

The authors present a method for decoding image semantics using composite region templates (CRTs). The CRTs define prototypal spatial arrangements of regions and features in the images. The system classifies unknown images by matching the strings of regions extracted from the images to the templates in a CRT library. They describe the process for generating the CRTs from photographic images by automatically segmenting the images into color regions. They demonstrate that the system performs well in classifying images from ten semantic classes and in searching for images in a large collection.