Research On Image Hashing Retrieval Methods Based On Semantic Transfer

The explosive growth of images leads to an increasing demand for efficient image retrieval.Hashing has become a popular retrieval technique.It can greatly reduce the memory consumption of storage data and accelerate the process of retrieval by projecting the original high-dimensional data into compact binary codes and converting the time-consuming distance calculation in the retrieval process into the Hamming distance calculation between hash codes.Among the various types of hashing,unsupervised hashing is more scalable than supervised one because it does not rely on expensive semantic labels.Therefore,unsupervised hashing is suitable for practical image retrieval systems and has attracted a lot of research interest.Particularly,benefiting from the booming development of deep learning,unsupervised deep image hashing methods have been greatly promoted in terms of retrieval performance in recent years.However,due to the lack of effective semantic supervision,deep hashing models are difficult to optimize and bring considerable training costs.As a result,the retrieval accuracy and training efficiency of existing unsupervised deep hashing are still limited.It becomes crucial to find free additional relevant semantics as the auxiliary supervision for training processes of deep image hashing and enhance the discriminative capability of image hashing models.To solve the above problems,this thesis delves into four aspects: social tag semantic transfer,image structure semantic transfer,web image semantic transfer,and neighbor semantic transfer,and proposes four image hashing methods based on semantic transfer.The research conducted in this thesis is outlined in more detail below:(1)To improve the image representation capability and retrieval performance of unsupervised hashing models,we formulate a unified scalable deep hash learning framework that explores the weak but free supervision of discriminative user tags that are commonly accompanied by social images.It jointly learns image representations and hash functions with deep neural networks and simultaneously enhances the discriminative capability of image hash codes with the refined semantics from the accompanied social tags.Further,we propose a discrete hash optimization method for directly solving the hash codes and avoiding binary quantization errors.Experiments on two standard social image datasets demonstrate the superiority of the proposed approach in terms of retrieval performance.(2)To fully capture the structural semantics of images and overcome the challenges of optimizing unsupervised deep hashing without explicit semantic supervision,we propose a lightweight augmented graph network hashing method that employs an effective two-pronged strategy.For one thing,we extract the inner structure of the image as the auxiliary semantics to enhance the semantic supervision of the unsupervised hash learning process.For another,we design a lightweight network structure with the assistance of the auxiliary semantics,which greatly reduces the number of network parameters that need to be optimized and thus greatly accelerates the training process.Experimental results show that the proposed method achieves significant performance improvement compared with the state-of-the-art unsupervised deep hashing methods in terms of both retrieval accuracy and efficiency.Notably,on MS COCO dataset,our method achieves more than 10% improvement in retrieval precision and 2.7x speedup on training time compared with the second best result.(3)To exploit the semantics available in web images and achieve cross-domain semantic transfer,we propose a webly supervised image hashing method that learns image hash codes from web resources.We first train a concept prototype learning network on the web images.Further,we design a lightweight Siamese network architecture and a dual-level transfer mechanism to efficiently transfer well-trained network parameters and the prototype codes learned from source web images to the target images.Experiments on two widely-tested image datasets show the superiority of the proposed method in both retrieval accuracy and training efficiency compared to state-of-the-art image hashing methods.(4)To address the issue that most existing methods pay little attention to the query content analysis during the online retrieval stage,we propose an online query expansion hashing method.Our approach incorporates complementary self-expansion and neighborhood-expansion strategies to learn semantically invariant feature representations of images and adaptively transfer the semantics of neighboring samples to the corresponding images.Moreover,our semantic expansion and hash code generation are unified in an end-to-end framework to guarantee retrieval efficiency.Experimental results demonstrate that the proposed method achieves superior retrieval accuracy and comparable retrieval efficiency compared with the state-of-the-art methods.Particularly,on MS COCO dataset,this method outperforms the sub-optimal results by about 6%.