Research On Supervised Hashing Method

With the rapid development of Internet technology and the widespread popu-larity of mobile devices,multimedia data is showing an explosive growth trend.How to quickly retrieve the neighbors of the query data in a large-scale database has become more and more important.Hash learning,as one of the main meth-ods of approximate nearest neighbor retrieval,has attracted wide attention from researchers due to its good performance in retrieval speed and storage overhead.Hash learning can be roughly divided into two categories:the unsupervised and the supervised.Compared with unsupervised methods,supervised hashing meth-ods can effectively improve the accuracy of retrieval by introducing semantic labels,pairwise similarity,ranking relations and other supervised information,thereby becoming the main direction of hash learning research.Although super-vised hashing methods have got many achievement recently,there are still some deficiencies and challenges in this field.For example,how to make full use of la-bel information and learn complex hash codes discretely with acceptable training time and storage overhead;how to compress the hash code length of large-scale samples to further reduce storage overhead and retrieval time;how to deal with the instability of the hash learning algorithms due to parameter initialization,local optimal solutions and unreasonable parameter settings,thus improving the accuracy of the hash learning algorithms.To tackle the above problems and challenges,this thesis explore supervised hashing methods and proposes six hash learning models consisting of single-modal methods and cross-modal methods.First of all,to deal with the insufficient utilization of semantic labels and the low accuracy of short-length hash codes in single-modal hashing,this thesis propose supervised discrete hashing with mutual regression and supervised short-length hashing:(1)To cope with the insufficient utilization of semantic label information and the instability of the discrete hash learning process,this thesis propose a method named supervised discrete hashing with mutual regression.This method utilizes single projection to describe the mutual regression relationship between the se-mantic label and corresponding hash code.The mutual regression can preserve the pairwise similarity,thereby making the learned hash code more stable and accurate.(2)To cope with the short-length hash codes with weak classification abil-ity,serious information loss and poor accuracy,this thesis propose a method called supervised short-length hashing.The proposed method integrates robust and mutual regression,matrix factorization,discrete optimization and balance constraints to learn more stable and accurate short-length hash codes.Moreover,given a single-modal hashing algorithm,a self-improvement method is proposed to boost its performance:(3)Since hash learning easily suffers form local minimum and instability due to unreasonable settings of hyperparameters and optimization methods,this the-sis propose a self-improvement framework for hash learning algorithm and design a method for linear model hashing.The proposed method can improve the sta-bility of given linear model hashing algorithm without adding any constraints or penalties to original algorithm.In addition,the proposed framework can boost the research of deep hashing methods.Furthermore,combining the aforementioned single-modal hashing methods and the model boosting framework,this thesis propose a method named rein-forced short-length hashing:(4)To further improve the performance of short-length hashing,a method termed reinforced short-length hashing is proposed.For enhancing the classifi-cation ability of short-length hash codes,the proposed method adopts mutual regression to effectively use semantic label,while embedding the pairwise sim-ilarity into the process of hash code learning through asymmetric strategy.In addition,the proposed method also proposes a model boosting strategy by in-tegrating the bit balance constraint and bit uncorrelation constraint to further optimize the distribution of hash bits.Last by not least,inspired by the aforementioned single-modal hashing meth-ods,this thesis propose two cross-modal hash learning methods:(5)Since the discrete cross-modal hashing is unstable and time-consuming in a way,this thesis propose a method named fast discrete cross-modal hashing with regressing form semantic labels.The proposed method learns the affine transformation from the semantic label to corresponding hash code,which can alleviate the Hubness problem,improve the stability of regression and reduce the time complexity significantly.(6)To further improve the performance of cross-modal hashing,a cross-modal hashing algorithm based on a two-step model is proposed.The proposed method first learns the similarity-preserving hash representation based on mutual regres-sion,and then learns the modality-specific projections that can maintain the consistency between heterogeneous feature distribution and semantic similarity for out-of-sample extension.The proposed method can effectively improve the precision in cross-modal retrieval tasks.