Combining Configurational and Statistical Approaches in Image Retrieval

We develop a framework for combining configurational and statistical approaches in image retrieval. While configurations have semantic description power, the explicit representation of an image by a set of configurations lacks the vector space structure from which the statistical feature-based representations have benefitted. That makes concept learning and prediction harder. Our framework treats configurations analogously to words occurring in a document. It combines a configuration-based approach with statistical approaches to take advantage of both the semantic description power of the former, and the simple vector-space structure of the latter.