Research On Cross-modal Retrieval Method Based On Average Approximate Hashing

The rapid development of the Internet has brought a lot of multimedia data,and retrieval of multimedia data has become the need of most people.As the name suggests,cross-modal retrieval is a technology that can retrieve data in different modalities.With its effective and efficient properties,it has been widely considered in the retrieval of large-scale multimedia data.Then the hash technology can convert massive data into binary code through projection matrix,which greatly reduces the storage space and retrieval time and also promoted the crossmodal retrieval technology.The current mainstream hashing method is to learn a common subspace by using the collective matrix factorization method to further obtain a unified hash code,so that heterogeneous data can be retrieved through it.But this method still has many shortcomings.First of all,most of these methods only focus on preserving the locality of the data,and ignore the preservation of reconstruction residuals in the process of matrix decomposition.Secondly,when the cross-modal data is projected into the common semantic space,the energy loss of the data is not considered.Thirdly,due to the different properties of different modal data,it is unreasonable to project heterogeneous data directly into a unified semantic space.Therefore,unlike most cross-modal retrieval methods based on collective matrix factorization,this paper ultilize label information to construct similarity graph,and then preserves the features of heterogeneous data in different subspaces by PCA-like method.Then,different subspaces are approximated to each other by global similarity graph.Finally,a average approximation method is proposed to integrate different semantic spaces into the same subspace which can generate hash codes.The main contributions of this method are summaried as follows:1)AAH method integrates the locality,residual and energy preserving into a graph embedding framework.In the process of reconstruction and dimensionality reduction of the original data,the reconstruction residuals of the modalities are preserved.In the reconstruction,AAH introduces PCA-like strategy to ensure that the reconstructed semantic space can maintain the main energy of the data.2)AAH proposes to project heterogeneous data into different semantic spaces,so that different semantic spaces can preserve the properties of different modalies.Then AAH makes two semantic spaces close to each other globally and locally,so that different semantic spaces can be close to each other,which can get the unified hash codes;3)Then AAH ultilizes the average approximation strategy to get the average value of the semantic space of different modalities approximated to the unified hash codes,so that the learned unified hash codes can preserve the information of different modalities at the same time.Experiments on three widely used standard databases show that the proposed AAH outperforms several state-of-the-art cross-modal hashing methods.