| For the massive amounts of multimedia data such as images and videos,one of the most fundamental problems is that automatically and quickly finding the content of in-terest from the images,i.e.,content-based image retrieval.Because of the intelligent,real-time and precise requirements of image retrieval,it is a very challenging task that designing high-performance content-based image retrieval algorithms.This dissertation focuses on key technologies of content-based image retrieval,including image repre-sentation(how to design efficient and discriminative image features),feature fusion(a powerful method of improving feature performance),and image similarity metrics(how to design efficient and accurate similarity strategy).The main contributions of this dis-sertation are summarized as follows:(1)For large-scale image retrieval,in the classical Bag-of-Word(BoW)model,a larger visual vocabulary can guarantee the retrieval accuracy but lead to a low recall,whereas a medium-sized vocabulary can improve the recall.Therefore,we propose an image retrieval algorithm based on separable visual vocabulary and sparse representation,which can retain both the advantages of large vocabulary and small vocabulary.In order to reduce the quantization error and improve recall,a sparse representation model based on non-negative orthogonal matching pursuit algorithm is proposed,which simultaneously quantizes each feature to multiple visual words.To further improve the retrieval accuracy,BoW image representation based on local gradient features and global structure feature of each image are extracted and fused in similarity measurement stage.The experimental results show that the proposed image retrieval algorithm based on the separation visual vocabulary and sparse representation achieves higher retrieval accuracy.(2)An image retrieval algorithm based on compressed sensing feature fusion is pro-posed.In the calculation process of the traditional Vector of Locally Aggregated De-scriptors(VLAD),since the contribution of each local descriptor to the VLAD vector is not uneven,the residual norm changes obviously,which will directly affect the image similarity.To this end,we propose a weighted VLAD model to balance the contribu-tion of each local descriptor to the VLAD vector.Furthermore,the deep feature and the weighted VLAD vectors based on local gradient features and local color features are fused to improve the comprehensiveness of image representation.In order to obtain a bet-ter feature fusion result,compressed sensing is introduced for the three features fusing,so that different features can be projected into a same subspace,which not only reduces the dimension,but also obtains better retrieval results.The experimental results on image databases show that the proposed image retrieval algorithm based on compressed sensing feature fusion achieves higher retrieval accuracy.(3)Two image retrieval algorithms based on composite visual phrase and multi-level image block joint similarity matching are proposed.1)A local composite descriptor which can simultaneously describe gradient and color information of the region of interest is proposed,and a composite visual phrase is formed by fusing deep feature representing the high-level semantics and the local composite descriptor in a same block to improve the retrieval accuracy.2)Using multi-level image partition and multi-feature extraction,block index is created for each block feature to improve matching efficiency.Considering the similarity between block feature vectors and the similarity between block indexes,a new method based on multi-level image block joint similarity matching is proposed to reduce the loss of creating block index to ensure matching accuracy.Finally,image similarity is calculated by feature weighted fusion strategy.The experimental results on image databases show that the proposed image retrieval algorithm based on multi-level image block joint similarity matching achieves higher retrieval accuracy. |
PDF Full Text Request