Image Retrieval Based On Hashing Learning

In recent years,a large scale of multimedia data has been generated with the rapid development of the Internet.How to quickly obtain useful information from large-scale multimedia data is important and large-scale image retrieval in the field of multimedia has received extensive attention.The goal of large-scale image retrieval is retrieving images similar to the target image from the massive image database by using approximate nearest neighbor search techniques.In approximate nearest neighbor search techniques,hashing learning plays a critical role.Hashing learning aims to map the high-dimensional and real-value data into the low-dimensional and binary space.And in the binary space,the retrieval based on hamming distance can be used to replace the retrieval based on Euclidean distance in the real space.Therefore,hashing learning takes the advantages of fast retrieval speed and low storage costs.Based on different learning modes,this thesis has proposed a series of more advanced hashing learning methods for large-scale image retrieval tasks to improve the performance of image retrieval.The hashing methods proposed in this thesis begin with supervised learning and unsupervised learning and futher extend from traditional hashing learning to more advanced end-to-end deep hashing learning.The main works and contributions of this thesis can be listed as follows:Firstly,we propose a new supervised hashing strategy in traditional hashing learning—supervised information reconstruction,i.e.,recoding the label information.This strategy can make the newly learned features retain the original semantic similarity and generate robust binary code.This thesis introduces Hadamard Code to recode the given supervised information,which makes the recoded information more adaptable to hashing learning.Combining this strategy with at least squares regression model and extreme learning machine model,we put forward two new supervised hashing methods,which have achieved excellent experimental results.Secondly,from the aspect of projection and quantization step in hashing,we study how to reduce the quantization loss for optimizing the learning objectives of unsupervised hashing learning.On one hand,we propose a balanced projection method to avoid the quantization loss caused by unbalanced variance of projected features in hashing learning.By clustering the dimensions of original features into a given number of clusters,our method can reduce the number of dimensions,and generate features with a balanced variance distribution.On the other hand,this thesis proposes a quantization loss function based on rate-distortion theorem in information theory for solving the quantization of unbalanced projected features.We learn an optimal quantization and encoding scheme based on distortion minimization of rate-distortion theorem,via quantizing each dimension of real-value features by using the variable bits encoding.The performance of these proposed unsupervised hashing methods outperforms previous hashing methods.Thirdly,by exploring a new distance measurement,we propose signal-to-noise ratio(SNR)distance to replace the original Euclidean distance measurement in deep metric learning,so that the intra-class distance is further closer and inter-class distance is farther in learned deep features.Based on deep metric learning and SNR distance,this thesis proposes a new loss function in deep supervised hashing learning to further constrain the consistency of similarity of the learned deep features and the given semantic information,so as to obtain more discriminative features.In this thesis,we propose a deep metric learning method and a deep hashing learning method based on SNR distance measurement.Compared with previous works,these proposed methods can greatly improve the accuracy of image retrieval.In general,this thesis mainly studies the hot issues in image retrieval based on hashing learning and introduces our six works involving three aspects.These works in this thesis have been proven to show a significant performance improvement on image retrieval over the previous methods in the well-known public image retrieval datasets.