Research On Cross-modal Hashing Methods Based On Similarity Preserving

With the advent of the era of big data,big data presents the characteristics of large-scale and various modal forms,which brings new challenges to the field of information retrieval.Faced with a large amount of high-dimensional multi-modal data,how to quickly and accurately retrieve data from different modalities as a hot issue has attracted the attention of many researchers.The cross-modal hashing methods solve the problem of mutual retrieval of different modal data by constructing models to map heterogeneous data of different modalities into homogeneous hash codes.Cross-modal hashing is suitable for large-scale cross-modal retrieval problems due to its low storage cost and fast retrieval speed.A large number of cross-modal hashing methods have emerged recently,but they still have some shortcomings:Most current methods maintain the intra-modal similarity based on the original characteristics of the data,while overlooking the semantic-based intra-modal similarity.Some methods that use semantic supervised information convert semantic labels into paired similarity matrices,which reduces the discriminative ability brought by label category attributes.In addition,most methods obtain hash codes by directly discarding the discrete constraints.And it will result in a large quantization loss and make the hash codes inaccurate.To solve the problems mentioned above,two cross-modal hashing methods are proposed.To solve the problem based on similarity maintenance,Supervised Discrete Anchor Graph Cross-Modal Retrieval Hashing(SDGCH)is proposed.It utilizes semantic category information and graph method to learn unified hash codes.We introduce anchor points to reduce the complexity of the Laplacian matrix for constructing the graph.Based on the anchor graph method,the inter-modal similarity based on semantic information is maintained by using semantic label projection.In order to solve the discrete problem of hash codes,two optimization methods are proposed.One is to solve the problem by bitwise discrete optimization using the alternating maximization method,and the other is to directly solve the hash codes by introducing intermediate variables.Using these two discrete optimization frameworks improves retrieval accuracy and efficiency.Aiming at the problems of similarity preservation and label discrimination,Label Consistent Locally Linear Embedding based Cross-modal Hashing(LCLCH)is proposed.It utilizes Locally Linear Embedding to maintain the manifold structure of the original data.And it utilizes semantic labels to construct a common truth space and then quantized into a Hamming space.Therefore,it not only maintains the potential intra-modal correlation of heterogeneous modal data,but also maintains label consistency.In order to further ensure the effectiveness of the hash codes and reduce the quantization loss caused by relaxing the discrete constraints,LCLCH combined with the iterative quantization method obtains the discrete binary hash codes directly.The mutual retrieval experiments of image and text modalities are performed on single-label dataset Wiki and multi-label datasets MIRFlickr and NUS-WIDE,and the feasibility of the methods are verified by comparison experiments.