Multimedia Retrieval by Means of Merge of Results from Textual and Content Based Retrieval Subsystems

Resumen

The main goal of this paper it is to present our experiments in ImageCLEF 2009 Campaign (photo retrieval task). In 2008 we proved empirically that the Text-based Image Retrieval (TBIR) methods defeats the Content-based Image Retrieval CBIR “quality” of results, so this time we developed several experiments in which the CBIR helps the TBIR. The TBIR System [6] main improvement is the named-entity sub-module. In case of the CBIR system [3] the number of low-level features has been increased from the 68 component used at ImageCLEF 2008 up to 114 components, and only the Mahalanobis distance has been used. We propose an ad-hoc management of the topics delivered, and the generation of XML structures for 0.5 million captions of the photographs (corpus) delivered. Two different merging algorithms were developed and the third one tries to improve our previous cluster level results promoting the diversity. Our best run for precision metrics appeared in position 16th, in the 19th for MAP score, and for diversity value in position 11th, for a total of 84 submitted experiments. Our best and “only textual” experiment was the 6th one over 41.