A Research On Semi-Supervised Deep Hashing With A Bipartite Graph

With the explosion in the volume of image data,it has been raised as a big chal-lenge about how to index and organize these data efficiently and accurately.Hashing is advantageous due to its fast query speed and low memory complexity.Traditional hashing methods leverage hand-crafted features to represent images,which exists per-formance bottleneck.Recently,with the great success of deep learning,researchers propose deep hashing methods which use neural networks to learn features,yielding remarkable performance compared to traditional methods with hand-crafted features.However,most of existing deep hashing methods are designed for the supervised sce-nario and require a large number of labelled data.In order to reduce the cost of labelling data and making full use of unlabelled data,we propose a novel semi-supervised hashing method for image retrieval,named Deep Hashing with a Bipartite Graph(BGDH)in this paper.To explore the underlying data structure,BGDH exploits a large number of unlabelled data and limited number of labelled data,and constructs a bipartite graph which contains similarities among in-stances to discover the hidden data structure,based on which an embedding is generated for each instance through an unsupervised graph embedding learning approach.Then,we feed raw pixels as well as graph embeddings to a deep neural network,and learn deep representations and hash codes with a supervised learning approach.BGDH is an end-to-end training framework that simultaneously optimizes both supervised and unsupervised loss.Since graph embeddings of instances are learnt from the bipartite graph,the basic BGDH is unable to predict hash codes for instances which are not appeared in training data.To address this limitation,we further propose an inductive variant of BGDH.In-ductive BGDH adds a sub-network to connect raw pixels and graph embeddings,and updates the parameters in the sub-network during training.However,the performance of graph embeddings is limited due to the shallow network architecture in inductive BGDH.Thus,we design an improved inductive BGDH approach.The improved ap-proach uses an aggregator function,which iteratively updates the graph embedding for each node by leveraging information in neighbours,to generate better deep graph em-beddings and further improve the performance of hash codes.Experimental results on two real image retrieval datasets show that our proposed approaches based on BGDH framework outperform state-of-the-art hashing methods.