Research And Application Of Feature Filtering Strategy In Relevance Feedback

Along with the popularization of the Web technology and the development,dissemination,and application of information technology , multimedia information mainly include of image,voice,and video become mainstream of information communication and service rapidly. The object and method of modern information processing have changed more. A great deal of all kinds of information have been gathered,transmitted,circulated and applied, we are stepping in a information society rapidly. The appearance and application of large-scale image database cry for effective retrieval mechanism. And Content-Based Image Retrieval (CBIR) is the sticking point to solve the problem. Due to the difficulty of semantic extraction, CBIR is still in common use. However, the similarity between low-level features may not necessarily reflect the semantic similarity and the user's perception among images. Since, human is the final user of retrieval system, it's very important to grasp the user's subjective perception among image contents by interactive technologies. In order to embed user's model into image retrieval system, relevance feedback mechanism has been introduced into CBIR domain in recent years.Relevance feedback has been shown as a powerful technique for interactive CBIR. Although a number of state-of-the-art techniques have been devoted to relevance feedback, such as Support Vector Machines (SVMs), existing techniques still have many drawbacks and limitations, including (1) paying little attention to the insufficient training samples problem; (2) assuming training samples are simply from one positive class and one negative class; and (3)'Curse of Dimension' and 'Semantic Gap' problems; (4) requiring many rounds of feedback to achieve satisfactory results. This thesis investigates the learning techniques on relevance feedback to address the above problems. It also proposes effective algorithms to improve them through different perspectives.We first present a minimal distance rank scheme to attack the insufficient training samples problem. The MDR method based on the nearest neighbor classification idea but uses relative ranks, and treats each positive feedback image as a single query. Experimental results are reported to show the effectiveness of the suggested scheme in a very limited training samples situation.Second, traditional learning techniques for relevance feedback usually assume training samples are drawn from one positive class and one negative class. We argue that it is more practical and reasonable to consider the relevant samples coming from multiple positive classes and the irrelevant samples coming from different negative classes. Based on this relaxation, a novel supervised-cluster-based relevance feedback algorithm constructed is presented in this thesis. Under the negative examples supervised, this algorithm merges positive examples into corresponding clusters according to their low-level featuresFurthermore,'Dimensionality Curse'and'Semantic Gap'are still important problems facing by CBIR. Making use of feature selection to reduce dimension is a necessary process of most image retrieval systems. To overcome these problems, a new feature selection approach, feature filtering strategy based on negative examples supervised is proposed to find the commonness of the positive examples, and then extract reduced feature set to fit the user's perception among them. Contrastive experimental results show that, it can not only speed up the similarity matching, but also help to bridge the 'Semantic Gap'.Experimental results show that, our method can solve the drawback of traditional algorithm. Furthermore, much better performance is observed, especially when using very limited feedback examples and very few rounds of feedback.