Large-scale Media Retrieval Based On Learning To Hash

In recent years,with the rapid development of Internet technology and digital de-vices.the amount of media data has become larger and larger.making the large-scale media retrieval task more and more critical.Due to the fast query speed and drasti-cally reduced storage,hashing methods are qualified to do this task and have attracted substantial attention.Recent research shows that supervised hashing has demonstrated better accuracy than unsupervised hashing in many real applications.Although many supervised hashing methods have been proposed with promising results,there are sev-eral challenges remaining to be addressed.For example,most of them use an instance-pairwise similarity matrix,whose size is the square of the amount of data.Because of using such large matrix,these methods may need large time and space cost and they are not conducive to deal with large-scale retrieval.For another example,some hashing methods relax the binary constraints of hash codes,optimize the continuous problem and then quantize the real-valued solution into binary hash codes.However,the errors caused by relaxation may degrade the performance.In this paper,we conduct in-depth research on large-scale media search based on learning to hash and propose four novel supervised hashing models.The main contri-butions of this paper are as follows:(1)After observation and analysis,we find that instances,which belong to the same class,should have the same hash codes.Based on this,we only need to learn class-level hash codes for each class and preserve the similarity among different classes.In this way,the training process is no longer dependent on the amount of data and can be done with low time and space cost.And this method can be easily extended to large-scale media retrieval task.(2)The widely-used objective function of supervised hashing involves an instance-pairwise similarity matrix,whose size is the square of the amount of data.This matrix makes the time complexity of these methods too large to do the large-scale media re-trieval task.Some methods leverage sampling technologies to avoid direct usage of this matrix.However,sampling will lead to information loss.In order to solve this problem,we replace a hash code matrix in the objective function,and then introduce an inter-mediate term whose size is only linearly related to the data amount.This term is the product of the instance-pairwise similarity matrix and the label matrix,and is a constant term that can be pre-computed before optimization.By using this intermediate term directly in the optimization,we avoid the high time complexity and large space con-sumption caused by the pairwise similarity matrix.And we use a discrete optimization algorithm to obtain hash codes,so that the huge quantization error can be avoided(3)Although we can use an intermediate term to avoid the direct usage of instance-pairwise similarity matrix,whose size is the square of the amount of data,this interme-diate term is still linearly related to the amount of data.When the amount of data is large enough,the space cost will still be very high.In the fourth chapter,we propose another supervised hashing method.Similarly,this method also introduces an intermediate ter-m that helps to avoid the direct usage of the instance-pairwise similarity matrix in the optimization process.And this method uses no sampling techniques that will result in information loss.The size of this intermediate term is independent on the amount of data,only related to the number of classes and the dimension of the feature.Thus,it is much smaller than the instance-pairwise similarity matrix.In addition,the procedure of hash codes learning not only benefits from the similarity matrix,but also takes features of the data into account.And we design a discrete optimization algorithm to optimize all bits of hash codes in one step.Therefore,the method proposed in this chapter can learn the hash codes accurately and efficiently.Both experimental and theoretical anal-ysis show that this method is superior to the existing supervised hash methods and the method proposed in the previous chapter.(4)In addition to search media data within a single modality,cross-modal media retrieval is also an important research direction and has a wide range of applications in real life.Therefore,we have designed a cross-modal supervised hashing method.The method fully considers the manifold structure and semantic information of the data,and uses the discrete optimization strategy to ensure that hash codes can be accurately learned.By conducting experiments on widely-used datasets,the experimental results show that the proposed cross-modal hashing method is very effective.