A Hybird Image Retrieval System with User's Relevance Feedback Using Neurocomputing

This paper aims at developing a hybrid scheme for intelligent image retrieval using neural nets. Each item in an image database is indexed by a visual feature vector, which is extracted using color moments and discrete cosine transform coefficients. Query is characterized by a set of semantic labels, which are predefined by system designers and associated with domain concerns. The proposed hybrid image retrieval (HIR) system utilizes the image content features as the system input, and the semantic labels as its output. To compensate the deficiency of semantics modelling, an on-line user’s relevance feedback is applied to improve the retrieval performance of the HIR system. The neural net acts like a pattern association memory bank that maps the low-level feature vectors to their corresponding semantic labels. During the retrieval process, the weights of the neural net are updated by an interactive user’s relevance feedback technique, where the feedback signal comprise the neural net actual output, semantic labels provided by users and the given query. A prototype HIR system is implemented and evaluated using an artificial image database. Experimental results demonstrate that our proposed techniques are promising.