Cross-Media Retrieval Based on Synthesis Reasoning Model

To gain better cross-media retrieval performance,it is crucial to mine the semantic correlations among the heterogeneous multimedia data. In this paper,we adopt the synthesis reasoning model as the underlying mechanism to mining the multimedia semantics for cross-media retrieval. We construct the synthesis reasoning sources according to the multimedia object low-level features and the reasoning source intensity field according to the multimedia co-existence information. A series of multimedia semantic spaces are built by spectral method after synthesis reasoning. The cross-media retrieval is performed on a per-query basis by which different retrieval methods are adopted for different queries. Both short term and long term relevance feedback are learned to introduce the new multimedia objects into the multimedia semantic spaces which were not in the training set,to refine the reasoning result. Experimental results show that the proposed methods can be used to accurately mine the multimedia semantics and the approach of cross-media retrieval is accurate and stable.