Deep Hashing For Semantic Image Retrieval

With the rapid development of the Internet,online image resources are growing at an explosive rate.In order to retrieve relevant images efficiently according to the needs of different users in such a large-scale data,retrieval algorithms related to hashing came into being.However,in most of the existing hashing algorithms,the image feature extraction and subsequent hash function learning are isolated,which makes features hard to be well adapted to the hash function learning,affecting the retrieval effectiveness.Some hashing methods based on deep learning alleviate this problem,but still face a series of problems such as the difficulty of optimization and training,and the redundancy of generated hash codes.In this paper,we propose a novel deep hashing method,called deep hashing for semantic retrieval(DHSR),to achieve complete end-to-end learning by using deep convolutional neural network as feature extractor and also hash function learner.In this way,image feature extraction and hash function learning are integrated into one stage,so that these two tasks can improve each other in the joint process.To this end,we design a loss function combining quantization error with two types of supervised information,including point-wise labels and pair-wise labels,in order that the convolutional neural network can generate hash codes which not only maintain the original relative position of samples,but also learn better individual semantic features.Furthermore,we devise a hierarchical retrieval strategy using compound hash codes based on the neural network.Specifically,we can generate two sets of hash codes with only one training session,and combine advantages of hash table lookup and hash code ranking.In addition,we use the divide-and-encode module instead of fully-connected layer to generate less redundant hash codes,and adopt methods like ensemble and fine-tuning procedure,along with a series of techniques to improve the stability and accuracy of our method.Extensive experiments on three large scale datasets MNIST,CIFAR-10 and NUS-WIDE show that our DHSR method can outperform other state-of-the-art hashing methods in image retrieval applications.