基于SVM的图像内容检索研究-电路与系统专业论文.docxVIP

ABSTRACTAB ABSTRACT AB STRACT With“interactive systems’’and“human-in-loop”structures．relevance feedback is an important tool to improve the performance of content based image retrieval．It can bridge the semantic gap between high—level concepts and low—level visual features． Especially,SVM based relevance feedback greatly improve the performance of retrieval system with its good generalization． However’the SVM·RF’S performance may become poor because of the following three reasons：1)SVM classifier is unstable with small training samples；2)SVM’S optimal hyper-plane may be biased when the positive feedback samples are much less than the negative feedback samples；31 Over-fitting due to that the feature dimension size is larger than the size of the training set；4)Algorithm’S time cost is strictly limited，as the user takes part in the retrieval process．In response to these problems， this paper carries out the following research： Firstly,the key technologies ofCBIR described and analyzed． Secondly,a new algorithm is proposed．Combining multiple features and training multiple SVM classifiers，it gets a better grasp of image retrieval subjective intent． Thirdly,an asymmetric bagging based fuzzy support vector machine(AB-FSVM) is proposed．An asymmetric bagging is made to negative samples，and then based fuzzy theory and SVM，the retrieval images gotten．This solve the small training samples and samples’asymmetry problems．Besides，with relatively less feature dimension size，over-fitting problems can be eased．Experiments show that compared with existing algorithms，the retrieval performance has been greatly improved with only a slight increase in the time-consuming． Based on the above work，a local database·oriented image retrieval system is built．And the performance ofthe proposed algorithm is verified． Key Words：content based image retrieval(CBIR)，relevance feedback(RF)，support vector machine(SVM)，fuzzy support vector machine(FSVM) 中国科学技术大学学位论文原创性声明本人声明所呈交的学位论文，是本人在导师指导下进行研究工作所取得的成 中