Study On Landmark Image Retrieval And Place Recognition

Owning to the progress of computer technology and the popularity of camera devices,there has accumulated a large amount of landmark and street view images on the Internet.Given a query image,how to efficiently and accurately retrieve images with similar content from large-scale image gallery becomes a pressing requirement in many real applications.This dissertation targets at the issue of landmark image retrieval and place recognition,and the research focuses on the following three folds: the burstiness problem in the Bag-of-Word(BoW)model based image retrieval approach,the deep learning based image retrieval approach and the research of matrix factorization based similarity search.The main contributions are given as follows:(1)An approach for the group burstiness problem is proposed.Current approaches for burstiness problem focus on tackling the intra-image and inter-image burst correspondences with one-to-many pattern,but ignoring the group burstiness problem with many-to-one pattern deriving from similar local features among the query image.We develop the group burstiness processing approach,which detects the group of burst features on the query image,and discounts the contribution of burst correspondences in votong similarity weight based on the group information.The experimental results on landmark image retrieval and place recognition demonstrate that the proposed approach can not only imporve retrieval accuracy,but also be compatible with existing burstiness processing approaches.(2)An approach for the burstiness problem in post-processing step is proposed.In landmark image retrieval,the burstiness problem still exists in the post-processing step even after the filtering of spatial verification.These burst correspondences are typically one-to-many or spatial clustering correspondences.To imporve the final retrieval accuracy,we propose an approach to tackle the burstiness problem in post-processing step,which utilizes the geometric and visual word information of local features to remove error correspondences with large difference from the global scale variation and redunlent correspondences of one-to-many matching,and reduce the contribution of correspondences with spatial clustering.Experimental results on landmark image retrieval demonstrate that the prposed approach can be compatible with different retrieval model and imporve the retrieval accuracy.(3)A landmark image retrieval approach based on the fusion of global and local RMAC(Regional Maximum Activation of Convolutions)deep feature is proposed.The RMAC based landmark image retrieval approach ignores the global information of query images,and the localization error will appear in the object localization step.We develop an approach based on the fusion of global and local RMAC.Firstly,the global RMAC from query image is used for initial filtering.Then,after object localization step,the global and local RMAC feature will be fused as the representation of images for reranking and query expanation.Experimental results with different deep features on landmark image retrieval demonstrate that the proposed approach can improve retrieval accuracy with low extra computation cost.(4)A post-processing approach for NetVLAD-based(Network Vector of Local Aggregated Descriptor)place recognition is proposed.In place recognition,the initial results retrieved by NetVLAD deep feature are correlated both in spatial location and visual content.To improve the recognition recall,we propose a post-processing approach based on the place fusion,which considers the reciprocal neighbor relations in spatial location and visual content between initial results and reranks them.Experimental results on place recognition verify the effectiveness of the proposed approach.(5)A similarity search approach based on locality-constrained linear coding(LLC)is proposed.Global feature based image retrieval typically relies on similarity search,but current sparse coding based similarity search approach consumes extensive off-line dictionary learning time.To handle this problem,we propose the LLC based similarity search,which directly employs k-means clustering for dictionary learning and LLC for data encoding.The LLC based approach can reduce dictionay training time and reduce data encoding error.Experimental results on landmark image retrieval show that the proposed approach can reduce off-line training time effectively and improve retrieval accuracy.