Research On Supervised Hashing Methods For Large-scale Media Retrieval

Nearest neighbor search plays an important role in both research and application fields.However,with the explosive growth of multimedia data on the Internet and daily life,it gradually becomes impossible to find the exact nearest neighbor of a query within acceptable time and space complexity from such large-scale data.As a result,researchers turn to the idea of approximate nearest neighbor search that aims to seek a balance between retrieval accuracy and efficiency,and hashing based retrieval methods have attracted much attention for the fast retrieval speed and low memory consuming in the last two decadesAfter a period of development,now supervised hashing methods,which take the advantages of supervised information like semantic labels to better capture the similar-ity relationship among data instances,have become the main stream of hashing based retrieval methods.These supervised hashing methods have achieved promising perfor-mance,but there are still some important issues worth further studying.At first,the key of supervised hashing methods is how to make full use of supervised information to learn better hash codes and functions,which is still an open problem without per-fect solution.Besides,with the development of deep learning,many works have tried to design supervised hashing methods based on deep networks.However,whether su-pervised information could be used in a way different from the traditional supervised hashing methods when designing a deep one,as well as how to combine deep neural networks and hashing methods effectively to take the advantages of powerful feature extraction networks and experiences of hashing methods designing,still need further study.Finally,how to use the idea of supervised hashing to solve more difficult and important practical problems is also an important issue to be considered.Based on the above discussion,in this thesis,in-depth research on supervised hashing methods for large-scale media retrieval is conducted,and a novel supervised hashing method is proposed to deal with a more difficult but valuable practical problem,i.e.,fine-grained data retrieval.Specifically,the main works of this thesis include(1)A simple but effective strategy is introduced to combine deep feature extraction networks and existing traditional non-deep hashing methods,and design a new strategy,e.g.,hash code reconstruction to take the advantages of supervised information.Dual Deep Neural Networks Cross-Modal Hashing,i.e.,DDCMH,including three training stages and two deep networks as hash functions,is proposed.Specifically,an existing single-modal hashing method is first selected to generate the the initial binary codes for one modality,e.g.,text;thereafter,these generated codes are used as supervised infor-mation to train the deep network for the other modality such as image;finally,the image codes generated by the trained network are reconstructed according to a reconstruction procedure,and then used as supervised information to train a text network.This method could act as a framework to extend any single-modal hashing method to deep cross-modal hashing method and deal with cross-modal search task,and the performance of the proposed method would increase as better single-modal hashing method selected.(2)The traditional two-step learning strategy is improved in this thesis,which learns hash codes and hash functions separately like above introduced in DDCMH,and pro-pose a new method to make full use of semantic labels.A novel Two-stEp Cross-modal Hashing method,TECH for short,is presented for cross-modal retrieval tasks.As a two-step method,it first learns hash codes directly from semantic labels,while trying to exploit the label correlations in the label space;then in the second step,different from other methods that mainly focus on hash codes learning while ignoring hash functions,this method modifies and integrates the code reconstruction idea mentioned above to in-troduce semantic relationship into the hash function learning procedure.To the best of our knowledge,this method is the first hashing method that exploits label correlations,and also the first two-step hashing model that introduces supervised information in both step.(3)This thesis attempts to solve a more difficult but practical and valuable problem,e.g.,fine-grained hashing,which works on fine-grained datasets where the differences between instances may be very subtle,and propose supervised feature refinement based on the characteristics of fine-grained data.It is a new topic in the field of hashing based retrieval,and there are only very few related works now.At first,three key issues that fine-grained hashing should address is raised,i.e.,fine-grained feature extraction,fea-ture refinement as well as loss function designing,then propose an effective fine-grained hashing method with a double-filtering mechanism and a proxy-based loss function.The method adopts a well designed feature extraction network and inherits the two-step training strategy.The key of this method is the double-filtering mechanism that consists of a space filtering module for better fine-grained feature extraction as well as a feature filtering module for extracted feature refinement.The method is efficient,effective and easy to implement,and this research also demonstrates the potential of hashing based fine-grained retrieval.