Research And Implementation Of Image Retrieval Based On Feedback And Mutilple Feature Extraction

Nowadays,with the rapid grows in image database and the development of computer technology,the superiority of image retrieval technology as an efficient information retrieval method is more evident.This thesis mainly studies the technologies of multi-features extraction,image feature quantification and modeling in CBIR as well as the application of the technology of relevance feedback.Multi-features representation can make up the shortcomings about not enough feature description in single image feature.The performance of multi-features representation largely depends on the strategy of the selection about image feature and the technology about image feature quantification.Therefore,this thesis mainly studies on how to build an efficient and effective model for multi-features representation and quantification.In addition,relevance feedback(RF)technology is also an important research part in CBIR.The main idea of RF is letting the user participate in CBIR system and ameliorating image retrieval result by successive iteration of annotation.However,the problem about limited feedback information in a feedback action cannot be ignored.This thesis is engaged in a lot of work design,an algorithm for augmenting user's feedback information and making up for the lack of insufficient training data to achieve a good retrieval performance.In the section of experiment,this thesis makes use of two open source databases for the performance testing and the experimental result have proved the proposed can achieve a good retrieval accuracy.The main content of the thesis include:(1)In this thesis,some basic theories and technologies of CBIR are summarized.This thesis gives a brief introduction about the basic theory and the development of CBIR,the basic flow of image retrieval,the technologies of image feature quantification and dimensionality reduction as well as the method about RF.(2)This thesis designs a novel algorithm of image feature extraction by utilizing SIFT,HOG and Opponent-Color model after a carefully analysis of multi-features fusion technologies.In order to maintain the independence between the various low-level features,this thesis designs to describe image feature separately at first and then fuses each feature into a complete image feature descriptor.In the process of feature quantification,this thesis firstly uses the Bag of Feature(BOF)for imagefeature modeling and the LLC for local feature coding.Finally,the technology of pooling and space pyramid model are used to achieve higher level feature modeling.In order to verify the performance of the proposed image feature extraction algorithm,several experiments results have conducted on two open-source image databases to demonstrate the superiority of the proposed method.(3)The thesis proposes a SVM-based classification model for RF.In the process of classification model training,the insufficiency of training samples would cause the instability of model.In order to obtain more training samples,a feature subspace partitioning algorithm(FSP)and a pseudo-labeling strategy are proposed to utilize unlabeled samples to train the SVM model.In addition,this thesis proposes an active learning based selection strategy for most informative images selection to solve the little finity of feedback information in each retrieval round and reduce the workload of user's labeling.Experimental results from two famous image dataset show that average accuracy obtained by the proposed method is 15% higher than baseline.Moreover,by comparing with another three SVM-based RF algorithm,this algorithm can get the query balance point with minimum number of feedback rounds.(4)Experimental analyses have been conducted for some parameters which would affect the performance of the proposed algorithm.In our work,the codebook size of bag of feature and the threshold of pseudo-labeling selection function are two uncertain factors in retrieval performance evaluation.For the consideration of reliability,several experiments conducted in two open-source are carefully analyzed respectively,and the optimal value are chosen for our work.