Research On Salient Object Detection Of Remote Sensing Images Based On Attention Mechanism And Multiscale Features

Salient object detection(SOD)in remote sensing images(RSIs)is an important topic in remote sensing information processing.Despite the remarkable progress in SOD for natural scene images,it is difficult to directly extend the success of SOD for natural scene images to RSIs due to the cluttered backgrounds and variations of salient object scales and types in remote sensing images.In this thesis,a series of methods are proposed to address the above limitation.The proposed methods can provide high significance for practical applications of SOD for RSIs.The main research contents of this thesis are listed as follows:(1)For the challenge of variations in salient object scales,a dense concatenation multiscale feature network is proposed.The network extracts multiscale features through progressive concatenation,which can effectively extract multiscale information and suppress the interference of variations in salient object scales.Besides,A dense concatenation multiscale module is designed to fuse information between different scales,thus achieving better feature representation.Extensive experiments demonstrate the effectiveness of the proposed method.(2)For the challenge of variations in salient object types,an attention pyramid decoder network is proposed.The network embeds a global attention mechanism in the encoder structure to capture the relationship between different target types and hence alleviate the interference of variations in salient object types.In addition,the network adopts a pyramid decoder structure to progressively integrate feature information at different levels and hence realize multilevel feature information aggregation.Finally,a bidirectional residual refinement module is constructed at the end of the network to further enhance the structural integrity of salient objects.Extensive experiments demonstrate the robustness of the proposed method.(3)For the challenge of cluttered backgrounds in RSIs,a global perception network is proposed.The network first captures global and local attention information through the global and axial attention mechanisms.The global and local attention information are mutually enhanced for better attention generation.Besides,the network simultaneously aggregates features at different levels and controls the information transfer at each level,effectively filtering out redundant features.Extensive experiments demonstrate the superiority of the proposed method.