Image Search by Latent Semantic Indexing Based on Multiple Feature Fusion

A new multiple feature fusion based latent semantic indexing (LSI-MFF) method has been presented to achieve improved image retrieval performance. The LSI-MFF method extracts different physical features, which come from not the whole image but its main objects, and constructs a multi-modal semantic space, each dimension of which represents a different feature component of the image. LSI-MFF can discover the latent relations among the images which are actually similar in semantic. Furthermore, semantic relevance feedback (SRF) information from the users, which implies "what the users choose is the best", is also integrated into LSI-MFF method to improve the feedback performance of the system. Experiment results demonstrate the good robustness of LSI-MFF. The prototype system VAST (visual & semantic image search system) based on LSI-MFF has shown that LSI-MFF has superior performance and is especially suitable for mass image database such as web environment.