Collection Fusion for Distributed Image Retrieval

Searching information through the Internet often requires users to contact several digital libraries, author a query representing the information of interest and manually gather retrieved results. However, a user may be not aware of the content of each individual library in terms of quantity, quality, information type, provenance and likely relevance, thus making effective retrieval quite difficult. Searching distributed information in a network of libraries can be simplified by using a centralized server that acts as a gateway between the user and distributed repositories. To efficiently accomplish this task, the centralized server should perform some major operations, such as resource selection, query transformation and data fusion. Resource selection is required to forward the user query only to the repositories that are candidate to contain relevant documents. Query transformation is necessary in order to translate the query into one or more formats such that each library can process the query. Finally, data fusion is used to gather all retrieved documents and conveniently arrange them for presentation to the user. In this paper, we introduce an original framework for collection fusion in the context of image databases. In fact, the continuous nature of content descriptors used to describe image content, makes impractical the applicability of methods developed for text. The proposed approach splits the score normalization process into a learning phase, taking place off-line, and a normalization phase that rearranges scores of retrieved images at query time, using information collected during the learning. Fusion examples and results on the accuracy of the solution are reported