Research On Deep Hashing Retrieval Methods For Remote Sensing Images

With the continuous advancement of remote sensing technology,the quantity of remote sensing image data has rapidly increased.Massive remote sensing image data has gradually become an important support for national major needs such as urban planning,environmental protection,and national defense construction.In the current construction process of an efficient and high-quality remote sensing image information management system,how to quickly and accurately retrieve the required remote sensing images has become an urgent and important problem that needs to be solved.The aim of this dissertation is to explore two key issues in remote sensing image hash retrieval,namely remote sensing image feature learning and hash code learning.The excellent performance of deep neural networks in the field of computer vision has made them the main choice for image feature learning and hash code learning in remote sensing image hashing retrieval.However,existing deep hashing retrieval methods for remote sensing images still face many challenges: 1)existing feature learning methods do not fully consider modeling spatial context information;2)traditional supervised hash code generation captures pairwise similarity between data through pairwise learning,which is difficult to achieve global similarity measurement between data;3)the hash codes learned by unsupervised methods have insufficient discrimination;4)cross-modal retrieval methods are difficult to establish semantic correlations between different modal data.To address these issues,this dissertation conducts the following research:(1)To address the problem of eliminating redundant information and enhancing effective features in remote sensing image feature learning,this dissertation proposes a hierarchical fusion remote sensing image feature learning framework based on selfattention mechanism.A skip-layer self-attention module is designed for feature learning,using self-attention mechanism to suppress the interference of noisy pixels and enhance the utilization of effective information.To further prevent the loss of effective information in the hierarchical propagation process,the designed hierarchical fusion method more fully utilizes the features of each level.(2)To address the issue of global similarity measurement between data in hash code learning,this dissertation proposes a supervised remote sensing image hashing retrieval method based on Vision Transformer and center similarity learning.This method constructs an end-to-end structure suitable for joint learning of remote sensing image features and hash codes based on Vision Transformers.It uses a hash code learning method based on center similarity to establish hash centers and designs a loss function based on center loss to measure global similarity between hash codes.This loss function simultaneously considers center similarity constraints,hash code balance,and quantization loss generated during the transform from continuous feature to hash code.The loss function promotes hash codes with the same semantic information to have a closer distance in the Hamming space.(3)To address the issues of insufficient discrimination of the learned hash codes in unsupervised methods,this dissertation proposes an unsupervised remote sensing image retrieval method based on the Swin Transformer structure and contrastive hashing.This method adopts an unsupervised feature and hash code learning strategy based on contrastive learning and establishes a remote sensing image feature encoding network based on the Swin Transformer to further improve the feature learning capability.The proposed contrastive hashing loss function is used to constrain the hash code generation process,which consists of binary contrastive loss and distribution aggregation loss.The binary contrastive loss term serves as the main component of the loss function to ensure consistency between different views of the same image,while the distribution aggregation loss further promotes the convergence of feature distributions generated from different views of the same image towards the original image,achieving high-quality feature and hash code learning under unsupervised conditions.(4)To address the modality difference between remote sensing images and voice,this dissertation proposes a deep image-voice hashing method for remote sensing crossmodal retrieval.The feature and hash code learning of image and voice is combined into a unified framework.In addition,considering the similarity preservation and bit balance of hash codes,a new image-voice pairwise loss function is proposed to establish the semantic correlation between different modal data in the Hamming space,effectively reducing the impact of modality differences on retrieval performance.