Active Learning Mehtod For Content-based Image Retrieval

With the development of Internet network and multimedia technology in the 20th century, there is more and more multimedia data to be processed. As the database storage increasing, however, it is a challenge how to orgnize, express, store, manage, query and retrieve massive multimedia data for traditional database technique. Content-based Image Retrieval (CBIR) is a hot research topic in multimedia information process. Image, Vedio, as the most intuitive and visual expression among informative data, how to query and retrieve effectively becomes an urgent problem in multimedia information process.Visual features such as color, shape and texture are used to find similar images in Content-based Image Retrieval (CBIR). It is helpful for those features to find the same semantic images, but there are still lots of problems to be solved in CBIR, such as sementic gap, feature extraction and retrieval method etc. The common method to retrieve images in CBIR is to combine Relevance Feedback and Classification. In this paper, Active Learning [21, 22, 23] method is used in Relevance Feedback [16]. The new Active Learning methods "Positive sample Enhanced Angle-diveristy Algorithm" and "URL-based Positive sample Enhanced Algorithm" are proposed for solving imblance problem in CBIR, which means the number of positive sample is far less than the negative ones. In addition, this paper uses SVM as classifier for image retrieval as its good performance. Our paper analyzes some kinds of SVM algorithms and points out the shortcoming of SVM whose ranking is unreasonable for CBIR. Finally a new ranking method for SVM is proposed.The experiments are conducted on two datasets, Stardard Corel dataset which contains 10000 images and Web dataset which contains more than 11000 images downloaded from internet. Our experiments prove that our active learning method and new ranking method for SVM algorithm have better performance in CBIR compared with traditional method.