Semantic Segmentation Of Remote Sensing Image Based On Deep Learning

Semantic segmentation of artificial features in high-resolution remote sensing images is one of the research hotspots in remote sensing imaging technology,which can be widely used in geological surveys,satellite navigation,and accurate target positioning.The remote sensing imaging process has the influence of noise interference and obstacle occlusion.When using traditional methods to semantically segment artificial features in high-resolution remote sensing images,the accuracy of segmentation is not high,and the amount of calculation in the segmentation process is also large.The convolutional neural network in deep learning has strong feature extraction and parameter calculation capabilities,good robustness to noise and affine transformation,and good generalization ability to scene changes,which can improve the accuracy of image semantic segmentation,and reduce the amount of parameter calculations.Therefore,the use of deep learning methods for semantic segmentation of artificial features in high-resolution remote sensing images has important research and practical application significance.Firstly,this paper introduces the research background of remote sensing image semantic segmentation and the current research status at home and abroad,expounds the theoretical basis and traditional methods of image segmentation and image semantic segmentation;outlines the structure of deep learning and convolutional neural networks,loss functions and commonly used training strategy.Secondly,the methods of semantic segmentation of remote sensing images based on deep learning are studied,effectively fuse the global information in large-scale images with the rich detailed information in local image sub-blocks at the depth feature level.The details are as follows:（1）A semantic segmentation method for remote sensing images based on GLNet and HRNet is proposed.On the basis of GLNet,this method uses HRNet to replace the Res Net backbone in the global branch to obtain side output feature maps with higher resolution and stronger characterization capabilities.The original feature map sharing strategy is cancelled,and local branches are trained independently to eliminate the confusion caused by the feature maps in the global branch.In addition,a multi-level loss function is used to optimize the network.Experimental results show that the method is better than GLNet in terms of segmentation accuracy and mean absolute error.（2）An image semantic segmentation method based on HR-GLNet and semantic flow is proposed.This method introduces the semantic flow alignment method of semantic flow in the network structure of HRNet,so that when feature maps of the different scales in the network are connected to each other,the semantic information can be aligned with each other.The method is used in the GLNet network framework,and the experimental results show that the method is better than GLNet and the network framework based on GLNet and HRNet in terms of segmentation accuracy and mean absolute error.Finally,Summarize the content of the full text and look forward to the future research directions.