Study On Kernel Based Hashing For Cross-modal Retrieval

With the continuous maturity of Internet technology and the continuous popularization of social software,the application data generated in social networks shows a geometric increase.In the face of massive data,how to rapidly and accurately retrieve the sample data of different modalities has always been the focus of many scholars.Cross-Modal Hashing technology encodes multimodal data into a binary form,and it has been widely used in large-scale multimedia data retrieval because of the advantages of fast retrieval speed and convenient storage.In this dissertation,two methods to achieve high-precision cross-modal hash retrieval are proposed which based on kernel methods.Experiments on two public datasets,Wiki and NUS-WIDE,demonstrate the effectiveness of the two proposed algorithms.The main contents of this dissertation are as follows:1.We propose a multi-kernel cross-modal hashing algorithm which based on maximizing the semantic correlation.Firstly,the original spatial features are mapped to the kernel space by using a plurality of kernel functions.Secondly,the corresponding hash function is learned by establishing the maximization of the semantic correlation.In the learning process of hash function,the method of sequence learning is adopted,which reduces the influence of the orthogonal constraint on the coding distinction in the objective function.Finally,the learned hash function generates the corresponding hash codes in each modality.2.We propose a multi-kernel cross-modal hashing algorithm which based on kernel matrix factorization technique.Firstly,the kernel function is used to map the original space features into the kernel space,and the collective matrix factorization technique is used to study the common semantic features between different modalities.Secondly,using the learned semantic features and the similarity matrix as the classifier label information to train the classifier for the instance data of different modalities.Finally,the parameter matrix of the classifier is used as the mapping matrix of hash function to learn the hash function,and then obtain the respective hash codes under different modalities.