Research On Image Retrieval Based On Deep Learning Hashing

With the development of Internet technology,the number of images has been greatly increased.And this phenomenon has played an important role in the development of computer vision.At the same time,it is an urgent problem to solve the issue of how to find the required image from the large-scale datasets.And the query time is expensive because the nearest neighbor searches the query sample and each training sample in the large scale image retrieval.In order to solve the above problems,hashing method is proposed.However,traditional hashing methods use low-level features to generate hash codes,but low-level features cannot represent the image information in detail.In recent years,many deep hashing methods have achieved good results since deep neural network can better obtain image information.Therefore,deep learning is suitable for large-scale image retrieval.This paper focuses on deep learning hashing for large-scale image retrieval.The main work and innovation of this paper are as follows:Firstly,this paper summarizes the basic knowledge of deep learning and introduces several typical deep learning models.And it also introduces the basic knowledge of hashing method and several typical hashing methods.Besides,for popular deep learning hashing methods,we summarize theirs main ideas.Secondly,for existing deep hash methods,hash-code learning based on deep feature representations is a shallow learning procedure,which cannot fully exploit deep feature representations to generate hash codes.In this paper,we propose a supervised end-to-end deep network architecture to learn features and hash codes together for large-scale image retrieval,named Deep Learning Supervised Hashing(DLSH).The proposed method devises a deep architecture by combining Convolutional Neural Network(CNN)and Recurrent Neural Network(RNN).DLSH can leverage the advantages of CNN and RNN to capture nonlinear structure of inputs so as to generate compact and discriminative hash codes.The proposed method not only makes the interaction between deep feature representations,but also uses deep neural network to generate hash codes based on deep feature representations.At the same time,a new objective function is proposed to preserve the semantic similarity and balancing property of hash codes,ensuring the probability that each bit becomes 0 or 1 is the same.The proposed method performs better than other state-of-the-art hashing methods on MNIST,CIFAR-10,CIFAR-20 and YouTube Faces datasets.Thirdly,existing hashing methods only use semantic features to generate hash codes by shallow projection but ignore texture details.In this paper,we proposed a novel hashing method,namely Hierarchical Recurrent Neural Hashing(HRNH),to exploit hierarchical RNN to generate effective hash codes.HRNH can learn similarity preserving hash functions for images by using image pyramid representation that contains texture details and semantic information simultaneously.And HRNH can exploit directly convolutional feature maps as input to preserve the spatial structure of convolutional feature maps.In addition,a new loss function is designed to maintain the semantic similarity and balanceable property of hash codes,and simultaneously considers the quantization error of binarizing the continuous embeddings into the discrete binary codes.The proposed method performs better than other state-of-the-art hashing methods on MNIST,CIFAR-10,CIFAR-20 and YouTube Faces datasets.