Research On Theories And Methods For Large-scale Image Hash Learning

Nowadays the dramatic growth of multimedia data from the widespread adoption of mobile phones and the Internet have posed a great challenge to data storage and information retrieval.The research of hashing technology benefits from the high retrieval efficiency and low storage cost for large-scale image retrieval and can extract the semantic features to characterize images to some degree,thus exhibits important theoretical and practical values.Hashing technology has attracted growing attention during the past few years and is the basis and the key step for large-scale image retrieval task.However,there exists a huge “semantic gap” between hash codes learned from the low-level visual features and high-level semantics,thus it is significantly challenging to design effective hash learning algorithms for improving the performance of image retrieval.This dissertation aims at learning the high-quality binary codes by starting from learning to hash and focuses on studying the following methods including unsupervised hashing methods,symmetric supervised hashing methods,cross-modal supervised hashing methods and asymmetric supervised hashing methods.The specific research content and main contributions of this context are summarized as follows.1.Since most existing unsupervised hashing methods ignore the global topological structure of the data set,we propose a manifold-ranking embedded order preserving unsupervised hashing method.This method incorporates manifold-ranking embedding,hypercube quantization,information theoretic regularization and consistency between manifold ranking and hamming ranking into a joint optimization framework,and we develop an efficient alternative optimization algorithm for solving this discrete optimization problem.2.Since existing hash methods suppose that the data distributions satisfy the manifold assumption that semantic similar samples tend to lie on a low-dimensional manifold which will be weakened due to the large intraclass variation,we propose a novel unsupervised hashing algorithm based on the nonnegative matrix factorization technology.This method integrates the high-level feature representations learned from nonnegative matrix factorization,bits balance and bits independence,and the out-of-sample extension term into a joint optimization framework,and we develop an efficient alternative optimization algorithm for solving this discrete optimization problem.3.Since existing deep learning to hash methods seek to solve the single retrieval task within one stream framework or jointly solve the retrieval task and the classification task within two stream framework which does not make full use of the semantic information to generate compact and discriminative hash codes,we propose a multi-task learning architecture for deep semantic hashing,which incorporates the retrieval task and the classification task within one-stream framework.Finally,these two tasks are investigated into one-stream deep learning framework which improves image retrieval performance.4.Since existing deep cross-modal supervised hashing methods do not well preserve the discriminative ability and the global multilevel similarity in hash learning procedure,we propose a global and local semantics-preserving based deep supervised hashing method for cross-modal retrieval.This method integrates the local semantic structure preserving term for capturing inter-modal correlations,the global semantic structure preserving term for capturing intra-modal correlations and a consistent regularization term for generating a unified hash codes for different modalities into an end-to-end learning framework,which can generate local and global semantics-preserving hash codes and improve the performance of cross-modal retrieval.5.Since it is typically time-consuming to train the symmetric hashing methods,and these methods can hardly take full advantage of the supervised information in the largescale database,we propose a novel discriminative deep metric learning approach for asymmetric discrete hashing approach for supervised hashing learning.This method integrates an asymmetric strategy with a deep metric learning method and improves the compatibility between the discrete encoding procedure for database images and the feature learning procedure for the query images with single deep network via end-to-end learning framework.Extensive experiments on various benchmark datasets show that the proposed asymmetric deep hashing method outperforms the existing hashing methods.