Research On Semantic Segmentation Methods Of High Resolution Remote Sensing Image Based On Deep Learning

With the rapid development of remote sensing satellite technology,intelligent analysis of remote sensing images is of great significance in the fields of security protection,environmental monitoring,military reconnaissance and so on.Semantic segmentation of high-resolution remote sensing image is the basis of remote sensing image analysis.However,traditional image segmentation methods need to design features manually,with low accuracy and poor generalization ability.With the development of neural network,the performance of image segmentation is no longer dependent on the depth of semantic feature extraction,and the performance of image segmentation is greatly improved by using the development of neural network.The main research contents of this thesis include the following three aspects:1.Semantic segmentation of high-resolution remote sensing images based on full convolution network and UNet network.This thesis analyzes the mechanism of full convolution network（FCN network）and UNet network.There are three different hop structures in the decoding part of FCN network,which are 8 times up sampling,16 times up sampling and 32 times up sampling respectively.In this thesis,three FCN networks with different up sampling structures are compared in semantic segmentation of high-resolution remote sensing images.UNet network is another classic network model of image semantic segmentation.This thesis analyzes the shortcomings of UNet network and makes a comparative experiment on the improved UNet network.2.Semantic segmentation of high-resolution remote sensing image based on multilevel context information aggregation.High resolution remote sensing images contain rich contextual information.For pixel by pixel segmentation tasks,capturing rich context information plays an important role in improving the segmentation results.This thesis builds a multi-level context information extraction network,and uses UNet network as the basic feature extraction network.In addition,in order to further obtain multi-scale information,ASPP module is added to the hop connection of UNet network.Features of different scales will shift to different degrees in the process of continuous convolution down sampling.When aggregating multi-level context information,this thesis uses the feature aggregation module to align the features first and then aggregate them.3.Semantic segmentation of high-resolution remote sensing images based on multilevel semantic feature fusion and hybrid attention mechanism.This thesis analyzes the different levels of semantic features in high-resolution remote sensing images,including low-level semantic features and high-level semantic features.In addition,multi-scale feature can also be regarded as a high-level semantic feature.Low level semantic features usually refer to the edge,texture and other features of ground objects,while high-level semantic features contain rich context information.In order to obtain multi-level semantic features,this thesis sets up a dual path network to extract two levels of features respectively,and then uses an efficient feature fusion module to fuse and interact the two levels of features.In addition,in order to make the network pay more attention to the features that are meaningful to the segmentation target,this thesis analyzes the mechanism of common attention mechanism methods,and designs an improved attention module to improve the performance of semantic segmentation.