Research Of Image Retrieval Based On Relevance Feedback

With the emergence of large-scale digital image library, people turn their attention to study the information in the image itself as the search index of the image contained, content-based image retrieval CBIR (Content Based Image Retrieval) came into being, which is widely used in medical, military, aerospace fields and so on. Study found that it is difficult to determine to use which methods and which methods of feature extraction in the large-scale retrieval process. This paper based on extracting image feature vectors combine neural network ensemble and dynamic weight adjustment with the relevance feedback respectively to improve the accuracy of image retrieval.Following the work:First of all,Original Bagging algorithm is improved. Original algorithm select training set randomly and repeatedly to train the individual network. But combined with relevance feedback in image retrieval will cause that the individual networks trained are invalid, due to there could be an imbalance in the distribution between positive and negative (satisfied and dissatisfied) image user marked in initial search results. The improved Bagging algorithm can select training set respectively from positive and negative images randomly and repeatedly, which can avoid the "one-sided" phenomenon of original algorithm effectively. Experimental results show that the method improves the effectiveness of the individual networks.Secondly, the clustering algorithm (a cluster algorithm based on Accessibility Matrix, abbreviated as AM) based on accessibility matrix is proposed to cluster individual networks, which can get the number of clusters automatically. The algorithm is based on graph theory and microhabitat theory in genetic algorithm. First of all, according to Euclidean distance between the sample points, the close sample points is connected which can simulate and constitute the initial microhabitat. Secondly, the microhabitats which have the same node are combined. In implementation process, an accessibility matrix of undirected graph can be obtained according to graph theory, accessibility vertexes belong to a microhabitat, which is a class. Then using this selective cluster ensemble algorithm to image relevance feedback retrieval, experimental results show that retrieval precision and recall ratio has been greatly improved.Third, one feature combination algorithm of dynamic weights adjustment based on SIFT (Scale Invariant Feature Transform) and global color histogram is proposed. First, using SIFT and global color histogram to extract feature vectors of images, first ranking can be got in descending order of similarity. After that, utilizing the users'relevant retrieval, a new formula of calculating the weights of features was presented, which doesn't increase the quantity of computation and takes into account the subjectivity of the user. Experimental results show that the method improves the results of retrieval comparison with the method of using single feature and permanent weights.