Research On Deep Hashing Search Algorithm Based On Domain Adaptation

Deep hashing methods aim to learn and extract binarization features of samples,namely hash code,using deep learning architecture.Binary codes have been widely applied in the area of largescale image retrieval due to storage and computation efficiency.However,in the deep hashing methods,the supervised information is required and the distribution of target samples is expected to align to that of the training samples.In practical applications,when the label information is lacked and the alignment of target and source distributions is not guaranteed,the retrieval performance has been always degraded due to a lack of the generalization ability.Therefore,domain adaptive technology has been widely studied recently to solve the above problems.In this article,the domain adaptive deep hashing based on multi-kernel maximum mean discrepancy is studied.Specifically,The Res Net-50 pre-trained network is used to construct a deep hashing model framework.All the hidden representations of the fully-connected layers are mapped into the reproducing kernel Hilbert spaces,and the multi-kernel selection method is used to reduce the deviation between the source domain and target domain.In the meantime,the hashing constrains are presented to enhance further the performance.The experimental results on the Office-31,Image CLEF-DA,and Office-Home datasets show that compared with traditional deep hashing algorithms,our composed method has the superior retrieval performance in the scenario of no supervised information and misaligned distribution.On this basis,the domain adaptive deep hashing based on minimum entropy maximum diversity method is studied,where the regularization term of maximum category diversity is utilized to dissolve the trivial solutions arised from the minimum entropy.The comparative experiments on the open data set demonstrate that our proposed method can further improve the retrieval performance in the scenario of domain shift.