| With the rapid development of the mobile Internet,digital imaging data has shown an unprecedented growth trend in recent years,which has brought severe challenges to large-scale image retrieval tasks.As a retrieval tool with excellent performance,deep hashing has been developed rapidly in recent years,but still faces some problems,such as the time performance and retrieval accuracy of unsupervised hashing that cannot meet actual needs.In the multi-label scenario,the correlation between tags is not sufficiently mined,and the supervised hash ignores the guiding role of image depth features.Based on the above observations,this work has conducted further research in the field of deep hashing,and designed and implemented a large-scale image retrieval system with complete training,retrieval,and performance analysis functions.The main contributions of this work are outlined as follows:First,to improve the time and accuracy performance of unsupervised hashing,this work proposes unsupervised discriminative deep hashing with locality and globality preservation(UD~2H).UD~2H constructs a reasonable similarity relationship utilizing the semantic information of the data itself in the output space of the deep network.The guidance of global clustering structure enables the hash function to generate more discriminative hash codes,and the addition of the neighborhood similarity relationship improve the cohesion of similar samples.By introducing an asymmetric learning strategy,UD~2H accelerates the optimization of the deep network and reduces the training time.At the same time,UD~2H introduces a clustering loss into the loss function to improve the global clustering results and realize the joint optimization of the similarity relationship and the hash function.Second,aiming at the problems of insufficient label correlation mining in multi-label scenarios and insufficient utilization of deep features in supervised scenarios,this work proposes deep feature attended deep hashing for multi-label image retrieval(DFADH).DFADH utilizes the co-occurrence frequency between tags in multi-tag scenario to construct a similarity relationship that applies to both single-tag and multi-tag scenarios.Combining the label-based similarity and the deep-feature-based similarity,DFADH proposes a corrected similarity matrix to describe the similarity relationship between multi-label samples and effectively utilizes the image semantic information contained in the depth feature.At the same time,DFADH proposes a loss function that combines the likelihood loss and the similarity difference loss,which improves the effect of the generated hash code on maintaining the semantic similarity between the original data.Third,this work designs and implements a practical system for large-scale image retrieval.The system encapsulates the two hash algorithms proposed in this work and the excellent hash algorithms that have emerged in recent years.The system provides complete data import,parameter setting,result viewing,and performance analysis functions.The independent training module of the system enables trainable hash algorithms to adapt to diversified data sources in actual retrieval tasks,and the performance analysis module facilitates the user’s analysis and selection of hash algorithms in different scenarios. |
