Research On Multi-modal Multi-label Hashing Methods For Large Scale Data Search

Recent years,with an explosive development of internet technology and mobile devices in our country and in the world,the scale of data become larger and larger,data storage methods and types are also increasing.Multi-modal multi-label data processing becomes very important in life,i.e.web pages,news,etc.are usually expressed as a combination of text,pictures and video,each page also has several tags as its labels.For how to search for the appropriate content in multi-modal data,cross-modal retrieval has become an important issue.Because of its superiority in storage and computing performance,hashing method is especially suitable for solving those problems.Hashing method maps original data into hamming space,then compute hamming distances between hashing codes.So it can greatly improve the efficiency of data retrieval;meanwhile,it also improves the efficiency of spatial storage due to using of hashing codes.Hash methods can be divided into three categories according to the training data:supervised hashing,semi-supervised hashing and unsupervised hashing.Unsupervised hashing method uses unlabeled data to train the corresponding hash codes,and the supervised hashing method uses labeled data to improve the retrieval performance;Semi-supervised hashing method only use a part of supervised information,it has better performance than unsupervised hashing due to label information,meanwhile it lower data requirement because not all data samples need to have label.However,in real world,the above methods can not solve such problem well,so multi-modal hashing method have been proposed.Multi-modal hashing method aims to retrieve similar data in another modal through the use of hashing codes.As it is widely used in large-scale data retrieval today,it has a good application prospect and high academic value of multi-modal multi-label data.Hashing method and multi-label learning by machine learning attracts big attention of current Internet and computer technology industry.It has has wide application prospect in web search,image retrieval fields through its user experience.Different from traditional single-label learning,multi-label data contain multiple labels,but the application of single-label learning method can not get good results on such data set,multi-label learning is also more expensive and time-consuming.Hash method can reduce the retrieval time and space complexity.This paper aims to use multi-modal multi-label hashing method to search large scale data source.We design hashing function for multi-modal multi-label data sets,which aims to use image to search text or use text to search image.Many hashing methods seldom consider the information of multiple labels,most of them only use label information to build a simple similarity matrix,e.g.,the value of the corresponding matrix set 1 when two samples share one or more than one same label and 0 otherwise.Apparently,such methods cannot make full use of the multiple label information.Thus,we want to propose a model that have good performance if it can make full use of the information contained in multi-modal and multi-label data.Specifically,we assume that every label is associated with a hashing code in the projected Hamming space,and then we combine the binary codes of its labels to generate the hashing codes of a sample.At last we minimize the Hamming distance between similar pairs and maximize the Hamming distance between dissimilar pairs at the same time,so we can also learn a project matrix which can be used to generate binary codes for out-of-samples.After learning hashing code,due to the learnt hashing code is generated by a linear combination of the anchor,so we consider that the hashing code contains label information.Then we expand the application range of the hashing code,based on multi-label classify function,e.g.uses hashing code as the training sample of multi-label classification feature.Because the features are using a compressed binary code,so it also has the advantages of low cost and fast speed.Experimental results on three widely used data sets show that our method outperforms or is comparable to several state-of-the-art hashing methods.We also use hash code to do classification,the experiment proves that our method is effective.