Research On Single-modal And Cross-modal Retrieval By Hashing Technology

With the development of computer and multimedia storage technology,people 's life is full of massive multimedia data,which includes image,text,voice,video,etc.How to quickly and effectively retrieve the required information from large-scale multimedia database becomes a challenging yet interesting issue,and unimodal retrieval and cross-modal retrieval are two typical and important research points to solve this.For the low storage consumption and fast retrieval speed,hashing has been widely studied and devel-oped in the field of unimodal and cross-modal retrieval.In this paper,we study unimodal and cross-modal hashing and propose new methods about them.Most of the existing unimodal hashing methods focus on similarity learning or quantization error minimiza-tion,but the discrimination of features still remains relatively unexplored.To address this,we propose 'Discrete graph hashing via affine transformation(ADGH)' and 'Discrimina-tive geometric-structure based asymmetric deep hashing(DGADH)'.The key problem in cross-modal retrieval is 'semantic gap'.To solve this,we respectively propose 'Collabo-rative subspace graph hashing(CSGH)','Xabel guided correlation hashing(LGCH)' and'Cross-modal sample-class semantic-consistent hashing(CSSH)'.Experiments on vari-ous datasets verify the superiority and effectiveness of the proposed models compared with the most relative methods.The innovations of this paper are as follows:(1)ADGH is proposed to learn discriminative graph embedding.It combines the strategy of joint learning and affine transformation in a new discrete hashing learning framework.The strategy of joint learning makes the learning process of graph embed-ding and binary hash codes promote each other.Moreover,affine transformation can ac-commodate both rotational angle and distance of graph embedding,while respecting the neighborhood structure among most samples.Experiments of image retrieval on three benchmark datasets show that ADGH outperforms the representative hashing methods in quantity.(2)DGADH defines an effective geometric structure to boost the discriminative ca-pability of real-value features while reducing quantization errors between real-value fea-tures and binary hashing codes in a novel manner.Experiments of image retrieval on four datasets show the effectiveness of DGADH against several well-established counterparts.(3)CSGH is put forward to perform a two-stage collaborative learning framework for cross-modal retrieval.This framework considers both the diversity of multi-modal features and the consistency of features in cross-modal retrieval.Moreover,CSGH en-deavors to keep the modality-specific neighborhood structure and the cross-modal cor-relation within multi-modality data through the Laplacian regularization and the graph based correlation constraint,respectively.Experiments of cross-modal retrieval on four datasets show the effectiveness of CSGH compared with the state-of-the-art cross-modal hashing methods.(4)LGCH is proposed to introduce label information into the hashing learning pro-cess.In detail,LGCH investigates an alternative way to exploit label information for effective cross-modal retrieval from two aspects:1)LGCH learns the discriminative com-mon latent representation across modalities under label guiding;2)LGCH introduces an adaptive parameter to effectively fuse the common latent representation and the label guided representation for hashing generation.Experiments of cross-modal retrieval on three multi-media datasets show LGCH performs favorably against many well-established base-lines.(5)CSSH can make full use of the similarity between samples and classes to keep the semantic consistency between them and meanwhile reducing the computation cost.In addition,CSSH maintains the semantic consistency of multi-modal data by connecting different modalities,real-value class features and shared hashing codes.Experiments of cross-modal retrieval on three datasets show the superiority of CSSH compared with other methods.