Research On Semantic Consistency And Matrix Factorization Based Cross-Modal Hashing Retrieval

One important characteristic of big data is multimodality,and with the arrival of big data era,the retrieval between cross-modal data becomes a potential demand such as using images to retrieve the related text documents.Cross-modal hashing methods use hashing functions to translate query data into binary codes in hamming space,i.e.hashing codes,and uniform the whole data of each modality in form,then retrieval between hashing codes in hamming space takes place of cross-modal retrieval,reducing storage consumption and accelerating retrieval speed.Furthermore,hashing codes preserve similarities among original data in general,which include intra-similarity and inter-similarity.Similarities preserving is an important part of cross-modal hashing methods and the starting point of research in this thesis.However,the current methods only-use low-level features to measure similarities among data,overlooking the importance of semantics and being harmful to reduce semantic gap as well as improve retrieval accuracy.Human beings distinguish or judge objects from the semantic point,so the real relationship between two points is decided by semantics.When low-level features include noise or don't have strong discrimination,the usage of semantic similarity contributes to the generation of hashing codes that have enough discrimination as well as improvement of retrieval accuracy.This thesis uses semantics to measure intra-similarity and inter-similarity and proposes two cross-modal hashing methods,i.e.,semantic consistency based cross-modal hashing(SCCH),semantic consistency and matrix factorization based cross-modal hashing(SCMFH).The experiments on two mainstream datasets verify the proposed methods.The main content and contributions of this thesis:(1)SCCH only uses semantics to measure similarities among data,reducing computing complexity as well as the semantic gap between hashing codes and high-level semantics,ensuring similarities among hashing codes and similarities among original data have consistency from the point of semantics.Hashing functions translate data into hashing code by linear projections and binarization.(2)Both low-level features and high-level semantics are used to measure similarities that are indicated by graph Laplacian matrix among data in SCMFH,reducing semantic gap between low-level features and high-level semantics as well as semantic gap between hashing codes and high-level semantics.The method constructs a shared abstract space for each modality data by matrix factorization,realizing data's abstract expression which are translated into hashing codes by quantization.Based on hashing codes,hashing functions are got by learning hyperplanes in binary classification.