Study On Relevance Feedback In Image Semantic Retrieval

With the universalization of information technologies and continuous improvement of information level, the amount of digital images increases rapidly. Therefore, in order to continually meet people's demand for capturing images, it is very urgent to study and develop some efficient and accurate retrieval technologies. The technologies of CBIR are lack of the capacity of capturing image semantic, so it is very necessary to make great effort to study the technologies of semantic-based image retrieval. Recently, relevance feedback has been an important means to capture images'semantic and narrow the "semantic gap". Thus, the key research work of this dissertation is the technologies of relevance feedback with the aim of improving the performances of image semantic retrieval.Main innovative contributions are as follows.1. By simple quantitative analyzing, the conclusion that the method of semantic network is lack of robustness at a multi-user circumstance is made. To address this problem, an algorithm of relevance feedback based on voting idea and semantic matrix is proposed; On the basis of the algorithm, a novel strategy is brought forward to add new images into the retrieval system, which can avoid great calculation of the similarities of visual characteristics between the new image and every image in the database. Thus, the calculation load of system can be cut down; In order to supply some more flexible and user-friendly ways to users, an effective method to process the query that can changed into the form of semantic vector is given.2. Relevance feedback technologies utilized in the field of traditional image retrieval are belongs to short-term learning strategy, which often discarded valuable user history feedback information. However, such history information often contains the information of users'perception to images'semantic. To solve this problem, a long-term learning strategy is study to improve the performance of retrieval. First, a framework of image retrieval with the function of feedback log learning is designed. Then, a method is proposed to record feedback log. After that, the semantic correlation degree between every two images is obtained through the analysis and exploration of feedback logs. On the basis of it, the semantic correlation matrix is constructed, and the images are clustered with K-means algorithm based on such matrix. Finally, the comprehensive similarity between two images can be got by weighted summation during the process of retrieval in order to make the retrieval results more closed to the understanding and perception of users.