Incremental hybrid bayesian network in content-based image retrieval
Baice Li, Senmiao Yuan · 2006
A learning system is one of the future directions of the evolution of content-based image retrieval (CBIR) system. Relevance feedback (RF) is a technique that enables systems to learn from users. In the past few years, this technique has been used as an effective solution for content-based image retrieval. Based on information theory, this paper proposes an incremental hybrid Bayesian network approach -IHBN, which processes the examples with decision tree. Only for those examples that cannot be processed are resorted by Bayesian network. The model constructed by IHBN can extend easily. Its size and structure change dynamically while learning. Its distinct incremental learning mechanism can not only make inductive learning possible while lacking of domain knowledge, but also depress the noise sensibility of the learning algorithm. This approach retains the interpretability of Bayesian network and decision tree, while resulting in classifiers that outperform both constituents. We propose this approach in content-based image retrieval system with relevance feedback. By comparing the performance of our approach with most conventional approaches in CBIR with relevance feedback, the experimental results demonstrate that the IHBN approach is more feasible and effective in the content-based image retrieval.