Research On Distributed Fast Supervised Discrete Hashing Algorithm

With the development of Internet technology and the widespread popularity of social media,information and data sources in many areas of society have been enriched and expanded,such as digital libraries,security images,remote sensing systems,medical images,etc.In the information age,data has become an indispensable resource for social development.Nowadays,facing of the trend of large-scale explosive growth of data,how to use modern information technology to handle such huge data has become a significant and challenging Sexual issues.The image retrieval method based on hash has become a research hot spot for researchers at home and abroad because of its small memory footprint,fast speed and good retrieval performance.The hash-based image retrieval method aims to represent image data with high dimensional features and high storage capacity in the form of hash codes,which not only reduces the storage space but also improves the efficiency of image retrieval.Among the existing hash learning technologies,the hash learning technology based on supervised learning uses supervised information for hash learning,and achieves higher retrieval accuracy,such as Supervised Discrete Hashing(SDH),Fast Supervised Discrete Hashing(FSDH),etc.SDH technology returns the hash code to its corresponding label in the objective function,while FSDH technology returns each label to its corresponding hash code space.However,in many real-world applications,a large amount of data is usually distributed on multiple machines,such as security data and communication network data.Therefore,in order to overcome the limitation that centralized hashing can only process data on a single machine,Supervised Distributed Hashing(SupDisH)was proposed.SupDisH extended SDH technology to a distributed framework for learning,and achieved good results.In this paper,combined with FSDH technology,a distributed fast supervised discrete hash algorithm(DFSDH)is proposed.Specifically,FSDH is introduced into a distributed framework in which all nodes share a centralized hash learning model.At the same time,consistency constraints are introduced on the nodes in the distributed network to ensure that all nodes perform hash learning in parallel by sharing parameters.For a single node,high-quality binary hash codes and hash functions are obtained by adopting an alternating iteration process.In terms of complexity comparison,the communication complexity of the DFSDH algorithm and the SupDisH algorithm remains the same,but the computational complexity of the former is much higher than the latter.Finally,the experimental results on the two public datasets CIFAR-10 and MNIST show that DFSDH has slightly improved retrieval accuracy compared with most centralized supervised hashing methods and existing distributed hashing methods.In contrast to the training efficiency,the training efficiency of the DFSDH algorithm is much higher than the training efficiency of the SupDisH algorithm,which also successfully validates the lower computational complexity of the DFSDH algorithm.In general,the DFSDH algorithm proposed in this paper has better competitiveness than most centralized supervised hashing methods and existing distributed hashing methods.