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

This dissertation discusses some key problems on content-based image retrieval techniquesfocusing on the gap between low-level features and high-level semantics. A relevancefeedback scheme for image retrieval is proposed, it employs iteration logistic regression andBayesian (ILRB), emphasizing on the distribution model of features and probability distributionon the image space. Based on the discussion above, an integrity image retrieval system isdescribed, and its effectiveness is proved in the experiments.Problems we discuss in this paper include:First, it models the elements of feature in image database. Most people look one image interm of semantics of the image, however, the color, texture and shape features of low-levelhave not the direct relation of the semantic of image, some feature can reflect part of semantic,but some not. So the feature of image and the similarity according to the feature can't representthe perception of people. Aim to this problem, the logistic regression is proposed tomodel the feature of image, according to the selection of user to select the feature better (adjustthe weight of elements in feature), so the retrieval system can select feature automatically,it solves the problem of difference between the low-level feature and the high-level semanticwell.Second, the paper models feature space in image database. Bayesian method is used toget the posterior probability of the images in database, and according to the posterior probabilityit can estimate the predictive probability of them, and then ranks the image in term ofthe predictive probability in the next feedback.