Research Of Relevance Feedback In Content-Based Image Retrieval

In recent years, content-based image retrieval (CBIR) has become an active research area. But the performance of (CBIR) systems based on low level features is poor due to the limitation of current image understanding technology and subjectivity of human perception. Relevance feedback is a promising tool to improve retrieval performance, in which human is involved in the retrieval process. This dissertation focuses on two novel relevance feedback techniques: a relevance feedback technique in image retrieval with analyzing feedback log; the method of Bayesian decision theory relevance feedback based on minimum risk, and proposes a long image semantic learning mechanism based on relevance feedback.The main researches of this thesis are followed from the extraction of image feature; the analysis of relevance feedback log; Bayesian dynamic memory; and the long-term learning of image semantic, the main work includes:1. The extraction of image feature and the quantization of color: A new variance analyzing based color quantization method is analyzed; the method is compared with other five quantization methods in terms of the root mean squared error. Experimental results show that the variance based color quantization method produces results that are far superior to other popular image quantization algorithms. the color space model, color feature extraction, color similarity measure is Described. Color coherent histogram is used for color feature extraction; it combines Color and spatial information.2. The analysis of relevance feedback log: Some problem in supervised learning is discussed; the current technology of log-based relevance feedback image retrieval is analyzed deeply, In the feedback process, the feedback history log is analyzed and indexed, and collaborative filtering algorithms is introduced for expanding the example of feedback sample, The speed and efficiency of system has been improved through online analyzing feedback history information.3. The adaptive Bayesian memory: The feedback idea based on mobile query point is analyzed, and the Bayesian decision theory is introduced in detail. The Bayesian decision theory relevance feedback method based on minimum risk is proposed, in which producing memory information through an adaptive mechanism, and applying minimum risk to satisfy users with the Retrieval results.4. The long-term learning of image semantics: Semantic information in the relevance feedback is analyzed, and a long learning mechanism for image semantics based on relevance feedback is proposed. A minimum risk determination mechanism is introduced to tag the image semantic. The semantic matrix can be update in relevance feedback.