Research On Semantic Segmentation Technology Of Remote Sensing Images Based On Contextual Information Aggregation

Remote sensing technology is increasingly used in fields such as autonomous driving,environmental monitoring,and disaster warning.Semantic segmentation provides detailed information about images compared to traditional image classification and object detection techniques.Convolutional neural networks must capture pixel-level contextual information for accurate segmentation results.This thesis aims to explore effective methods to enhance the accuracy of semantic segmentation for remote sensing images by utilizing pixel contextual information.Two network models are proposed to better utilize pixel-level contextual information,namely the Target Context Network(TCN)and Context Guided Network(CGN).The TCN mainly focuses on how to use pixel contextual information to improve the accuracy of semantic segmentation.This model combines pixel contextual information with pixel position relationships in the image,builds a target context module,and improves the VGG16.Comparative experiments on the publicly available datasets Potsdam and Jiage show that the TCN model has better performance than the SFNet network,with an average Intersection over Union(Io U)improvement of 0.21% and 0.72% on two datasets,respectively.Through analysis of experimental results on the target context,it was found that the network still has the problem of insufficient recognition accuracy for small objects and background information.In order to further improve the accuracy of semantic segmentation of remote sensing images,this thesis proposes a Context Guided Network.This network model integrates semantic information module,edge information module,and context guidance module,and better utilizes contextual information through the guidance network to improve recognition accuracy for small objects and background information.The experimental results show that the intersection over union of the proposed network on the Potsdam and Jiage datasets was respectively increased by 1.29% and 1.32% compared to SFNet,further improving the accuracy of semantic segmentation of remote sensing images.The effectiveness of the two proposed network models was assessed through experiments conducted on publicly available remote sensing image datasets,including Potsdam and Jiage.