Research Based On An Active Relevance Feedback Mechanism In Content-Based Image Retrieval

In order to narrow down the semantic gap between the low-level image features and user's high-level query concepts, and ultimately improve the efficiency and accuracy of content-based image retrieval system (CBIR), this paper presents a novel relevance feedback algorithm based on the Logistic Regression model (LR). In this paper, user preferences are added to the algorithm. Based on modeling of user preferences as a probability distribution, the algorithm can calculate the relevance probability of an image belonging to the set of those selected by the user. And it ranks the images according to their probability. The process is repeating until the user is satisfied with the query results or the target image has been found.Implementing this relevance feedback mechanism will face a key problem that the labeled sample size, that is, the number of the training examples is typically smaller than the dimension of the feature space of the images, the Logistic Regression model can't adjust parameters of an entire feature very well. The problem of scarcity of labeled (training) examples in the feedback process is effectively addressed by meaning of two aspects in this paper.Firstly, we will divide the feature space into several small subsets. We use the Iterative Logistic Regression model (ILR) to adjust the elements inside each subset. In addition, Active Learning algorithm (AL) is introduced to select the most informative samples to tag by users and add them into training set. This method not only can enlarge the training sample size, but also these samples selected are best for classifier and optimizing the classification result.Experimental results are shown that the performance of the retrieval system is greatly improved by the proposed method.