The Research On Image Retrieval Technology Based Relevance Feedback

With the development and application of multimedia technology and internet technology, various resources on the network become more and more richer. In order to meet various application requirements, image retrieval gets more and more popular, the image retrieval technology grows an important research subject.To meet the users'needs of image retrieval, content-based image retrieval in image retrieval has become an important research in image retrieval, which starts to extract color, texture, shape, space and other features from images, then match the corresponding features between images from image library and the image user want to retrieve, at last gives the same or similar images retrieved in contents to the user. It is the core for content-based image retrieval, how to accurately express and use image features achieving efficient image retrieval.In content-based image retrieval system, the image's contents are performed by the color, texture information and some low-level features. However, these underlying characteristics can not reflect the similarity of the high-level concepts in human visual perceive. Based on image feature extraction and measurement, This paper researchs the image retrieval based relevance feedback between the image bottom features and high-level semantic gap. By the interactive between system and users, relevance feedback is a technology through learning to improve the retrieval systerm's performance. This paper proposes a weighted distance method for relevance feedback mechanism, and image's weighted value is the standard deviation ratio of the feature values between the images in the database and associated image user selected. Making the relevance feedback technology applying for not only weight of independence each other but also for the weight of incremental update. On the other hand, the weights of features in turn can improve retrieval results using different features. The results show that the method has better adaptability and effectiveness. The research is accomplished in the ground of MATLAB 7.0 and SQL SERVER 2000, also adopts the average precision and progress to evaluate retrieval performance.