Online Matrix Factorization Hashing For Nearest Neighbor Search

Internet and multimedia technologies develop rapidly.This phenomenon has led to an explosive growth in data,and the modality of data is more diverse.How to efficiently manage and use information resources,dig out valuable information content,and promote the further development of society has become a widespread concern.Hash-based nearest neighbor search technology,because of its fast calculation speed and small storage space,provides an effective method for efficient management and application of large-scale information resources,and has attracted attentions of researchers.Currently,hash-based nearest neighbor search techniques mostly are batch-mode methods.The problem of batch-mode methods is that when the system obtains new data that is inconsistent with the historical data features or has inconsistent distribution.At this time,if hash functions are not updated,hash functions cannot learn the characteristics of the new data,resulting in a decrease of search accuracy.If hash functions are updated,new data and historical data need to be retrained,resulting in large computational and storage overhead.To solve these problems,this paper proposes online matrix factorization hashing for nearest neighbor,which makes hash functions incrementally update according to the new data,without storing a large amount of historical data,effectively reducing the calculation and storage overhead of nearest neighbor search tasks,improved search accuracy.The main research contents as following:(1)Aiming at problems of batch-mode hashing,this paper proposes an online matrix factorization hashing for large-scale image retrieval.When acquiring new data,the method decomposes the feature matrix of new data and applies the projection matrix to map it to the low-dimensional space.The error caused by the projection and matrix factorization uses Frobenius norm to measure,and the objective function is optimized and updated.The system stores matrix calculation results corresponding to the current data for the next update.In the process of hash learning,the method only needs to store calculation results of the historical data for updating the model without storing a large amount of historical data.The results of nearest neighbor search experiments on standard datasets show that the single-modal online hash search method based on matrix factorization is more accurate than existing batch-mode methods and online hashing methods.(2)In order to solve the online update problem of the hash function in cross-modal search tasks,this paper proposes a cross-modal online hash search method based on matrix factorization.When acquiring new data information,the method maps the paired data of different modalities to the same latent semantic space,and uses the alternating iterative optimization method to optimize the objective function to obtain a projection matrix and a low-dimensional coefficient representation matrix,and the quantized coefficient representation The system stores the calculation results of current data for the next update.The experimental results on standard datasets show that the proposed method has higher accuracy with the accumulation of data,and has higher accuracy than the existing cross-modal batch-mode hashing methods and online hashing methods.(3)An efficient optimization algorithm for online hashing based on matrix factorization is proposed.The online learning process in above methods uses the alternating iterative method to optimize the objective function.This method has more iterations in the optimization process and reduces the efficiency of updating the hash model.To solve this problem,this paper proposes an approximate Newton method to optimize the objective function.In the optimization process,the Newton method is used to calculate the Newton direction,and the descending direction of each iteration is determined.In order to simplify the inversion process of the Hessian matrix,the approximation of the inverse matrix of the Hessian matrix is solved by the diagonal approximation method.The optimization method is applied to the above methods,and compared with the method of alternating iterative method optimization,it can be verified that the method can effectively reduce the number of iterations of the optimization process while improving the search accuracy.In summary,this paper proposes online matrix factorization hashing for nearest neighbor search from two aspects of single-mode data and multimodal data,which effectively reduces the storage pressure of batch-mode hashing methods.In addition,in order to reduce the number of iterations of the hash function optimization process,this paper proposes an approximate Newton method to further improve the computational efficiency of the online hashing.Extensive experiments demonstrate that proposed methods outperform most state-of-the-art hashing methods.