Research On Single-modal And Cross-modal Retrieval Technology Based On Hash Method

In today's digital information age,multimedia data such as text,images,video,and audio flood the entire Internet.Due to the difference between multimedia data on attributes and characteristics,it is challenging to effectively manage and analyze these data,which promotes the emergence and rapid development of single-type data(single-modal)and multi-type data(cross-modal and multi-modal)analysis and pro-cessing technologies.Among them,retrieval technology is a typical one.Recently,the single-modal or cross-modal retrieval technology based on the hash method has become the mainstream research content in the field of image-text.In single-modal retrieval,the query data and the data in the database belong to the same modal,and cross-modal retrieval refers to the retrieved results from the database contain other modalities besides the query data modal.The hashing method can map high-dimensional data to low-dimensional Hamming space,and has the advantages of low storage consumption and efficient calculation.In order to meet the retrieval needs of large-scale data in today's society,the retrieval technology based on the hash method has attracted extensive attention.Despite the vigorous development of existing search technologies,there still are certain limitations.In single-modal hash retrieval,there are two types of dif-ficulties in processing robustness:1)Robustness measures,such as M-estimators,assume that noise follows a specific distribution and cannot be directly applied to discrete hash codes;2)Removing the noise of the original data can easily damage the neighborhood structure of data.In addition,cross-modal hash retrieval faces the problem of”semantic gap”,especially under supervised circumstances,the se-mantic expression and the semantic structure of the label between the modalities cannot be guaranteed to be consistent.In order to solve the above problems,the main work of this article is:1)In order to enhance the robustness of hash codes against noise,this paper proposes a single-modal hash retrieval method,named dual-graph regularized robust hashing(DGRH).Unlike existing robust hashing methods,this method directly uses the M-estimator to remove outliers in the dataset to ensure the robustness of the hash codes.Specifically,DGRH expects to recover the low-rank expression of the original data from the noise through the l₁norm,while using the dual-graph regularization to preserve the neighbor relationship between samples,and finally uses the recovered data to learn hash functions.In addition,this paper proves the robustness of hash codes through theoretical analysis,and the experimental results also prove that DGRH can achieve better results on three commonly used benchmark datasets.2)In order to narrow the”semantic gap”in cross-modal retrieval methods,this paper proposes a cross-modal hash retrieval method,named semantic-consistent cross-modal hashing(SCCH).This method embeds semantic labels into the common latent space,so as to ensure the learned latent semantic expressions and labels have the same semantic.In order to maintain samples and classes semantic consistency,SCCH maps each modal data to the same potential common subspace,aims to learn a separate hash function for each modal.And SCCH makes latent semantic expression of classes and hash codes are consistent through rotation transformation.In order to reduce the quantization error caused by approximating the real value to the binary code,SCCH maps the common latent semantic representation to the hash space through a rotation matrix,and then directly learns the hash code.In addition,this paper uses an efficient iterative optimization algorithm to solve the SCCH model.Experiments show that SCCH is superior to other representative methods on two commonly used multi-modal datasets.