Image retrieval based on dynamic-state Bayesian networks

In this paper, a new content-based image retrieval (CBIR) methodology is proposed to overcome certain drawbacks inherited in previously proposed Bayesian network (BN)-based CBIR methods. It incorporates short-term relevance feedback learning (SRF) and long-term relevance feedback learning (LRF) methods based on dynamic-state Bayesian network (DSBN). Firstly, the multi-layers BN is constructed according to the prior knowledge. Secondly, a training phase is conducted for updating BN parameters through using the feedback information. The training cycles can be stopped until the retrieval accuracy is adequately high. The newer BN parameters are generated for conducting further image clustering in a LRF scheme. Thirdly, SRF-LRF interactive refinement is applied in order to optimize the pre-trained SRF BN parameters, leading to the improved retrieval accuracy of the whole DSBN-based CBIR scheme. Experimental results on 10,000 images demonstrate the effectiveness of the proposed methodology.