Relevance Feedback Image Retrieval Which Based On Semi-Supervised Method

With the rapidly development of computer science and technology as well as the popularity of the Internet applications, It has produced a large amount of multimedia data, Which is presented in the form of pictures, videos and others in everyday life. Faced with such a large and rapidly growth of multimedia data, It has become an important issue that how to achieve fast retrieval and improve the retrieval accuracy in multimedia retrieval. On the basis of the previous work, we research depthly the main problem of image retrieval technology in the field of the multimedia retrieval.In this paper we give a image retrieval method which is based on image of interested region, and a relevance feedback retrieval method which is based on the semi-supervised learning is proposed on this basis. The retrieval method which is based on the image of interested region is a content-based retrieval method, and to retrieval image which through to extract the underlying features of image of interested region; The image retrieval method which is based on semi-supervised relevance feedback is an interactive retrieval method. Namely, on the basis of the initial search results, according to the user’s judgment, It designs two learner which thinking of co-training in the semi-supervised learning, and the method take full advantage of the unlabeled image to retrieval in the training process. The combination of these two methods can effectively improve the performance of image retrieval.In this paper, the research work as follows: Firstly, we give an overview of domestic and international research status and key technologies in image retrieval, including image low-level features (color features, texture features and shape features)extraction technology, the similarity measure and image retrieval performance evaluation method.Secondly, we study detailed the image retrieval technology, proposing an retrieval method which is based on the image of interested region. The method uses the harris corner detector to detect the images of points of interest, and then gives the convex hull of the image points of interest which is extracted by the convex hull of planar point set method, The area which is contained in the convex hull is the image of interested region. Finally, extracting color features and texture features of the image points of interest as the feature vectors to be retrieved. Through the experimental comparison and analysis, the use of this method to improve significantly the retrieval performance.Thirdly, we research the relevance feedback and semi-supervised learning which is in the machine learning. According to the idea of co-training, we give the paper of the ideology and algorithm description, and propose a relevance feedback method which is based on semi-supervised in the image retrieval. Use of a large number of unlabeled images to improve the accuracy of image retrieval, and to meet the needs of users.Experiments of this paper is in the Windows XP environment and design graphical user interface at aid of Matlab7.0software. It is to retrieve the10categories in the Corle image library of1000images. The experimental results show that we propose the retrieval method which is based on image of interested region and based on the relevance feedback retrieval method which based on semi-supervised have a good retrieval performance.