Research On Discrete Cross Modal Hashing Method Based On Ranking

The rapid development of information technology promotes the progress of human society,but also brings new challenges to scientific research.In the era of large data,data not only presents the characteristics of large volume and high dimension,but also tends to be more diversified in data forms.Under the impact of large-scale data,the traditional nearest neighbor retrieval technology has been unable to give consideration to the accuracy and efficiency of retrieval.Therefore,the concept of approximate near-est neighbor retrieval is proposed,aiming at sacrificing a certain accuracy in exchange for a faster response time.Hashing retrieval,as one of the representative techniques of approximate nearest neighbor retrieval,proposes the solution to the storage problem of large amount of data and the computational cost problem caused by large amount of data and high data dimension in large-scale data retrieval.At present,there are cor-responding hashing methods for single-modal retrieval task and cross-modal retrieval task.Cross-modal hashing retrieval is more challenging to maintain the similarity not only within the modals but also between the modals.In the existing cross-modal hash-ing methods,many of them build the similarity matrix and embed it into the hashing learning framework,or learn the mapping relationship between label matrix and hash code matrix,and seldom consider the relative similarity relationship between data,that is,the ranking information of samples.Ranking information is related to the sequence of returned samples in the retrieval list and represents the degree of similarity with the query sample.Therefore,the utilization of ranking information is helpful to further improve the retrieval performance.Considering the importance of ranking informa-tion,this thesis proposes two ranking based discrete cross-modal hashing methods for unsupervised and supervised scenes respectively.First,for unsupervised learning scenarios,this thesis proposes a anchor based multi-modal manifold ranking learning framework,by mining multi-modal data itself manifold structure information to establish a ranking score matrix,and anchor point sampling strategy is adopted to avoid the difficulty of training on large data sets which is common in the similarity matrix based hashing method.Then,the ranking informa-tion contained in the score matrix is used to construct a ranking based graph hash modal to learn hash codes and hash functions.Secondly,for supervised learning scenarios,in this thesis,semantic labels of data are added to the original unsupervised manifold ranking framework as the supervisory information to supervise the learning of score matrix.In the hash learning stage,this thesis also designs two strategies for preserving the global similarity and bit similarity of hash codes respectively to learn more accurate hash codes.In addition,discrete optimization schemes are designed respectively for the two proposed cross-modal hashing algorithms,which reduce the quantization error caused by the relaxed hash codes and further improve the accuracy of the learned hash codes.The main contributions of this thesis are summarized as follows:?In this thesis,a discrete cross-modal hashing method based on manifold ranking is proposed to mine and capture the manifold ranking information of data and integrate it into the graph hash learning.A ranking based graph hash learning framework is established.?In this thesis,a ranking based supervised discrete cross-modal hashing method is proposed,which combines the manifold structure of data and label information to learn a ranking score matrix,and two effective similarity preserving strategies are designed to achieve the global similarity and bit similarity preserving of hash codes.?For the two proposed cross-modal hashing methods,this thesis gives detailed discrete optimization schemes without the need for relaxation of hash codes.?Through a large number of experiments on several multi-label data sets,this thesis verifies the effectiveness of the two proposed cross-modal hashing methods and further tests the stability and convergence of the modals.