Research On Graph Based Hashing Algorithm For Multimedia Data Retrieval

Explosively growing multimedia data has brought tremendous pressure to the management,calculation and storage of the current computer systems.In the face of massive multimedia data,the storage,management and utilization of these data have been of great significance to the field of machine learning and computer vision.Retrieval is a basic application in the field of machine learning.The efficient retrieval of massive multimedia data is of great significance to the mining and utilization of multimedia data resources.Faced with a large variety of multimedia data such as image,video,audio,text,etc.,how to efficiently retrieve these data has become a hot research topic in academia and industry.For large-scale multimedia data,utilizing hash learning to generate hash codes can effec-tively reduce the data storage and calculation cost.Moreover,the efficiency of large-scale data retrieval system can be improved by reducing the data dimensions.At present,the research on hash learning method in the field of multimedia information retrieval has achieved initial results.However,the similarity measurement of the samples in original space and similarity information preserving techniques in hash learning process still need to be further studied.Furthermore,the application scenarios of hash learning methods still need to be further explored.In this dissertation,researches on the theory of graph-based hashing algorithm and its ap-plication in multimedia data retrieval are introduced.This dissertation mainly explores how to effectively maintain the similarity structure of the samples in original space to hamming space and extends the application scenarios of hashing methods.The main works of this dissertation are as follows:(1)This dissertation proposes a global similarity preserving hashing(GSPH)method,which preserves the global reconstruction similarities of samples in original Euclidian space to ham-ming space.This method can effectively detect the potential manifold similarity structure in original space and achieve an accurate description of global similarities among the samples in original space.In addition,GSPH provides direct mapping from original data to hash codes,which can simultaneously reduce the dimensional reduction error and binary quantization loss in the "two-stage" hash learning framework.By means of GSPH,original global similarity can be effectively preserved to hamming space.The experimental results validate that GSPH can effectively describe the potential manifold similarity structure in original Euclidian space,and maintain this structure to the hamming space.Based on GSPH,hash codes which reflect the similarities of samples in original Euclidian space can be explicitly generated.(2)The reconstruction similarity and the local distance geometry structure(distance based similarity)in original local space are utilized simultaneously in this dissertation,through which an accurate description of the local similarities among the samples in original space is achieved.Based on the local similarity,this dissertation proposes a local topology preserving hashing(LTPH)method.The graph structure constructed by LTPH can effectively improve the descrip-tion accuracy of the local similarity in original space.LTPH employs classification algorithm into the hashing learning process,and unifies the process of hash learning and the training pro-cess of classifier.With the facilitation of classifier,explicit mapping from original space to hamming space is constructed,which can effectively preserve the original local topology simi-larity to hamming space.The experimental results validate that LTPH can effectively generate hash codes that preserve the local similarity structure in original Euclidian space.(3)Research on the hashing learning method under multi-feature and multi-modality s-cenarios is conducted in this dissertation.The supervised information is utilized to construct discriminant similarity graph among samples in original space,based on which this dissertation proposes a discriminative bit selection hashing(DBSH)method.Different from the "Fusion +Encoding" framework that mainstream hashing learning methods adopted in multi-feature and multi-modality scenarios,DBSH utilizes the framework of "Encoding + Selection" to fully ex-ploit the information of different modalities from different perspectives.DBSH can be combined with the existing hashing learning methods to improve the utilization of existing hashing method-s.DBSH utilizes the supervised information to construct the bit selection criterion.At the same time,the alternating direction method of multiplier(ADMM)algorithm is utilized to overcome the difficulty caused by the discrete constraint in the bit selection process.With the facilitation of ADMM,the effectiveness of DBSH is improved,and the selected codes can fully maintain the discriminative similarities among samples.Considering the fast calculation speed of hash-ing codes,this dissertation further proposes an object recognition framework for multi-modality data,named "Hashing + Approximate Nearest Neighbor Voting(ANNV)"."Hashing + ANNV"utilizes the supervised similarity information among samples to improve the object recognition accuracy in multi-modality scenario.The experimental results validate that DBSH can effec-tively select hashing codes that preserve the discriminative similarities of samples in original space.The framework of "Hashing + ANNV" can achieve efficient and accurate recognition performance in multi-modality scenario.(4)Researches on the human motion sequence segmentation(HMSS)and human motion sequence retrieval(HMSR)are conducted in this dissertation.As one of the concentrated ex-pressions of multimedia data,realize HMSS is one of the basic tasks for semantic HMSR.In order to achieve efficient segmentation for human motion time series,this dissertation proposes a hierarchical HMSS framework based on hashing method.The framework firstly measures the human motion change degree in human motion sequence,and realizes the preliminary segmenta-tion of human motion time series.On this basis,not only the motion change degree of the human motion time series is considered,but also the process of human motion state change is studied.The hashing method is utilized to transform the human motion change process into a state vari-ation process,which can effectively reveal the inner correlation of motions in a same class.The inner correlation of the motions is expanded to effectively reduce the excessive segmentation during the HMSS process.On the basis of segmentation,the human motion sequence retrieval is studied and a hashing based human motion sequence retrieval(HBHMSR)method is proposed.HBHMSR utilizes the hashing method to extract key frames of a human motion sequence,which effectively improves the performance of HMSR.The experimental results demonstrate the ef-fectiveness of the proposed hierarchical HMSS framework and the HBHMSR algorithm.In summary,the main researches of this dissertation are graph-based hashing learning al-gorithm and its application in multimedia data retrieval.Three graph-based hashing learning methods are proposed and the applications of graph-based hashing method in image retrieval,human motion sequence segmentation and retrieval are studied and analyzed.The experimen-tal results validate the effectiveness of the proposed algorithm and its superiority over existing algorithms.