Research On Image Retrieval Algorithm Based On Online Hashing

Hash learning(also known as binary code learning)encodes high-dimensional data points into binary codes,so that the Hamming space can be effectively used to approximate the original high-dimensional space,and large-scale multimedia data can be quickly retrieved.In practical applications,search engines are usually required to index online image stream data,and online hashing algorithms emerge as the times require.Therefore,it is of great scientific and engineering significance to study the online hashing algorithm for hash coding of online streaming data and improve the learning efficiency and accuracy of the image retrieval model.But so far,online hashing is still an open problem,and the main challenge is that it is difficult to make a trade-off between model accuracy and learning efficiency.First,existing online hashing methods often rely on strong constraints to design robust hashing functions,such as orthogonality and similarity preservation,etc.,but require relaxation in the optimization process,resulting in decreased training efficiency.Second,existing online hashing methods lack global and local dual semantic information retention,which inevitably leads to low accuracy.This paper aims to improve the retrieval accuracy of the online hashing model,while ensuring the learning efficiency,so as to improve the efficiency of image retrieval.The main work is as follows:(1)A discriminative similarity-balanced online hashing algorithm named DSBOH is proposed.It preserves the similarity between newly arrived and already arrived samples in the original space in the hash space,uses the offline generated Hadamard matrix to form a codebook,and combines the global distribution and balanced similarity to generate a discriminant for image retrieval hash code.Firstly,an offline Hadamard matrix is constructed,and a codebook with corresponding code length is generated by using the locality-sensitive hashing algorithm to guide the generation of hash codes for multiple input samples;secondly,the inner product of hash codes is used to preserve the balanced similarity of the online data in the generated Hamming space;finally,DSBOH uses the alternate iterative method to solve the problem and gives the discrete optimal solution.Extensive experiments are carried out on three standard image datasets,CIFAR-10,MNIST and Places205,and the results show that compared with several advanced algorithms,DSBOH improves the accuracy of largescale online image retrieval,while shortening the training time,which is beneficial to realize fast online retrieval for large-scale images.(2)A manifold structure-based double semantic online hashing algorithm named MDSOH is proposed.Since the existing online hashing algorithms rely heavily on the constraint of pairwise similarity of data,the mining of the local manifold structure of the data is neglected.Therefore,MDSOH is committed to preserving the dual semantic information of global and local manifolds,aiming to reduce information loss,preserve more neighborhood information,and improve retrieval accuracy.Firstly,the pairwise similarity between the newly arrived data and the previous data is used to construct a global similarity matrix constrained hash code,so that the codeword retains the global data characteristics;at the same time,in order to preserve the local characteristics of the original data,for each data stream,use The anchor points obtained by clustering are constructed to construct an anchor point graph between the data samples and the anchor points to approximate the local structure of the original space and retain the low-dimensional manifold features.Extensive experiments on three benchmark datasets including CIFAR-10,MNIST,and Places205 demonstrate that the proposed algorithm effectively improves image retrieval efficiency compared with several existing advanced online hashing methods.