Bayesian Probability Model Based on Region and Relevance Feedback
贝叶斯框架下基于区域的相关反馈算法

Many researchers have found it can improve the retrieval performance by combining region-based representation and relevance feedback technology. Since the previous works have ignored the probabilistic distribution of regions in the same semantic class, it is hard to represent the semantic information effectively. In this paper, Bayesian probabilistic model based on region and relevance feedback is proposed. The probability model of image similarity can be constructed via the Bayesian classifier obtained by on-line learning and its certainty based on the least error probability of the nearest region in relevant images set. When it comes to the non-parameter density estimation technique for characterizing the region feature distribution, it also takes the collective distribution into consideration because of inaccurate segmentation. Thus, the posterior distribution of region feature can be estimated accurately, and the experimental results demonstrate its effectiveness.