A novel active learning approach for SVM based web image retrieval

Abstract

There is a great deal of research conducted on hyperplane based query such as Support Vector Machine (SVM) in Content-based Image Retrieval(CBIR). However, the SVM-based CBIR always suffers from the problem of the imbalance of image data. Specifically, the number of negative samples (irrelevant images) is far more than that of the positive ones. To deal with this problem, we propose a new active learning approach to enhance the positive sample set in SVM-based Web image retrieval. In our method, instead of using complex parsing methods to analyze Web pages, two kinds of "lightweight" image features: the URL of the Web image and its visual features, which can be easily obtained, are applied to estimate the probability of the image being a potential positive sample. The experiments conducted on a test data set with more than 10,000 images from about 50 different Web sites demonstrate that compared with traditional methods, our approach improves the retrieval performance significantly. © Springer-Verlag Berlin Heidelberg 2007.