Svm-based Active Feedback In Image Retrieval Using Clustering

With the rapid developments of computer technology and Internet, including digital images, the number of various kinds of multimedia data grows at a remarkable rate. Facing massive multimedia information, how to manage, organize and retrieval the useful information is becoming a very crucial and urgent problem, and also a major challenge. When the traditional annotation-based image retrieval methods have been unable to meet customer's demand, researchers have proposed content-based image retrieval technique, which retrieval by using visual features, such as color, shape, texture and other features.The success of content-based image retrieval (CBIR) is largely limited due to the gap between low-level features and high-level semantic concepts, which be called semantic gap. To bridge this gap, relevance feedback (RF) initially developed in text retrieval was introduced into CBIR since the 1990s and big success was achieved RF focuses on the interactions between the user and the search engine by letting the user provide feedback regarding the retrieval results. Researchers begin to look at this problem more systematically by formulating it into supervised learning problem, so, kinds of machine learning methods were introduced into the relevance feedback technology in many works. Recent work on RF often relies on support vector machines (SVM). Compared with other learning algorithms, SVM appears to be a good candidate for several reasons: generalization ability; without restrictive assumptions regarding the data; fast learning and evaluation for RF. With the idea of active learning, Tong proposed SVM-active algorithm. According to their criterion, good requests should maximally reduce the size of the version space, which can be approximately achieved by selecting the points near the SVM boundary. However, there may exist much redundancy between these examples, thus, the SVMs trained on these examples are usually biased and unstable.Visual feature selection has always been a hot domain in image retrieval research. When a user searches the database, his/her focuses on different feature subspaces are not equal to each other. In some cases, color may be the dominant subspace, while in some other cases shape may be more important. So, needed a way to better describe the relationships between the query and the attention the user pays on subspace.In order to improve the SVM-active selection criterion and a better modeling the interesting degree on different feature subspace, in this paper, we put forward an active learning algorithm. Color and texture are naturally considered as sufficient and uncorrelated views of an image; inspired by multi-view learning, we calculate the weight of color and texture feature subspace separately. SVM classifiers are learned in color and texture feature subspaces, respectively and the unlabeled data are classified. In order to reduce redundancy between these examples, a k-means based active selection criterion to select images for user's feedback is proposed. Experimental comparisons between our method and 1.Traditional SVM relevance feedback Algorithm; 2.Traditional SVMactive relevance feedback Algorithm show that the proposed algorithm has a higher accuracy, and has the better retrieval effect.