Image Retrieval Based On Relevance Feedback

With the quick development of Internet, text-based search already can't satisfy the advance requirement of multimedia searcher. The progress of image processing technology lead in new approaches of image retrieval: content-based image retrieval. It had a broad application outlook in General image retrieval image searching engine, multimedia filter in Internet and so on.The purpose of this paper is to develop a content-based image retrieval system. The system combined the technique of support vector machine with the mechanism of historic relevance feedback, and on the base of the constitution of this system the kernel algorisms of retrieval is researched.In this paper, we discussed the problems met in the approach of the mechanism of relevance feedback based on SVM in both academic and practical aspects. After analyzing the traditional training methods of SVM, it chosen an effective methode for the experiment system-SMO. Because different kernel function leads to different retrieval performance, we had chosen an appropriate kernel fuction and its parameter with experiments.On the basis of analysis of the CBIR technology, this paper indicated the primary feature of a retrieval algorithm: the features were represented as vectors; the space of each retrieval algorithm is independent; to each feature, weights were set for united retrieval. Afterward, this paper discussed how to design an image retrieval system based on relevance feedback, and established a experimental system.In this paper, we introduced an improved image retrieval approach based on HSVM. During the interactive procedure, the relevant and irrelevant images respect to the image marked by users both in current circulation and in the historical circulation are learned for constructing a HSVM classifier, with which we classify the images in database again. Experiments demonstrated that more relevant images can be found efficiently by the interactive method with a number of procedures even when the number of the samples in each circulation is a little. In addition, it has the generalization ability of limited training samples.