Research On Image Retrieval Technology Based On Distributed Deep Hash

With the explosive growth of unstructured data such as images,the large-scale semantic-based image retrieval has received extensive attentions.In such case,the deep hashing methods are proposed to extract the deep feature for hash codes with help of deep model,which can significantly enhance the retrieval accuracy and speed.However,a large amount of data needs to be collected for training on the data center,which can lead to not only high computational and communication cost but also the data privacy-preserving issue.Hence,this motivates the need for a distributed image retrieval model where the images distributed over different local databases can be trained in a unified form.Specifically,the main works are as following:First,the end-to-end distributed deep hash image retrieval algorithm is proposed to construct the deep hash model by using the local image dataset of the distributed nodes.During the training process,various nodes allow the interchange of parameters of output layer via the communication links.In such case,unified optimization of global parameters can be guaranteed by the Alternating Direction Method of Multipliers(ADMM)method.Furthermore,combination of the twice forward transmission and the back-propagation of deep learning enables an end-to-end deep hashing training.All these measures can ensure the effect of the unified training.Finally,the effectiveness of the distributed deep hash image retrieval algorithm has been verified on the open distributed database.On this basis,an end-to-end distributed hash multi-layer parameter optimization algorithm is proposed to sequentially optimize the multi-layer parameters by ADMM.At the same time,the twice forward transmission coupled with the back-propagation of deep learning achieves the target of the unified training effect.Hence,the generalization ability limited by fewer parameters of the output layer is addressed,which can guarantee the better generalization ability of our proposed method.Finally,experimental results on the open distributed database further demonstrate the feasibility of our proposed method.