Hashing based large-scale image retrieval has been getting more publicity recently,and many related centralized hashing methods have been proposed to achieve excellent performance.However,in some scenarios of applications,huge volume of data is stored in a distributed way.The cost of transmitting and computing all these data for training is colossal.A novel distributed hashing method is expected to address this issue,where the centralized training can be implemented by the exchange of relatively few parameters.The Alternating Direction Method of Multipliers(ADMM)method can be used to tackle the distributed convex optimization problems,and can be applied in the areas of large-scale machine learning,image retrieval,and computer vision.This paper studies an ADMM-based distributed hashing search methods.The objective of our proposed method can be described as follows.First,vector quantization can be used to minimize quantization error by training coding matrix.Furthermore,the classification error of hashing codes is minimized to improve the semantic accuracy of hash codes.In the scenario of distributed learning,the ADMM method can be used as a solution to the above issues.Specifically,a large task is decomposed into multiple sub-tasks that can be implemented individually on a separate computing node.Then the encoding matrix is obtained by an exchange of messages.Hence,the distributed hashing-based method can be developed for image retrieval.The experimental results on CIFAR-10 and NUS-WIDE datasets validate the superiority of this method.The distributed deep hashing methods is proposed to further improve the performance of distributed image retrieval.Specifically,a latent layer is designed to be placed between the feature layer and the output layer in the convolutional neural network.Then the binary constraint is combined with the classification error to generate the hash codes.First,the pre-trained process is implemented on each computing node.Then the ADMM method is employed to obtain the approximate global optimized classifier by the consensus optimization.Finally,the global optimized classifier is utilized to update the parameters of the latent layer to generate corresponding hash codes.The back propagation method is applied to update the parameters of the feature layers.Experimental results on the CIFAR-10 dataset demonstrate the effectiveness of the proposed method. |