Research On Subspace Based Image Retrieval And Classification Technique

With the fast development of the Network technology, content-based image retrieval has been a hot research point in the current information field. But because of the limitation of current image understanding technology and the famous semantic gap between high level concept and low level features for an image, the performance of the content based image retrieval is not good enough. In order to overcome these problems, researchers from different countries proposed many solutions, e.g. region based image description and relevant feedback.Subspace analysis is a kind of statistics methods. It aims at projecting sparse samples in a high dimensional space to a lower dimensional subspace by linear or nonlinear transform, so that the samples are close and effectively described.Considering the characteristics and structures, this paper develops research on content based image retrieval and classification based on subspaces technique.The main contributions of this dissertation are as follows:①In order to effectively descript the content of an image, a novel modified pseudo semantic model based on SVM ensemble learning is first built, which can give a good clustering of image semantic by projecting high dimensional vision description to a lower pseudo semantic subspace, meanwhile, the sample biased problem in SVM learning can be avoided.②In relevance feedback, in order to avoid the disadvantages of KNN and SVM active learning, the strategy of random grouping is proposed so that training samples are informative.③In relevance feedback, in order to improve the on-line studying ability, a ONPP based random subspace ensemble strategy is proposed. Semantic relations between images are obtained through manifold constructed by ONPP. To enhance the generalizing capability, the ensemble scheme is applied to individual weak classifier.④In image classification, in order to classify the samples nearby the hyperplane of SVM correctly, which are often get wrong classification by SVM, the manifold subspace, which show connotative semantic relations between samples, is introduced and by renewing these samples, better classification results are obtained.