Image segmentation is an important basic link of remote sensing interpretation.High-resolution remote sensing images contain complex object information.The application of traditional segmentation methods is greatly re...Image segmentation is an important basic link of remote sensing interpretation.High-resolution remote sensing images contain complex object information.The application of traditional segmentation methods is greatly restricted.In this paper, a remote sensing semantic segmentation algorithm is proposed basedon ResU-Net combined with Atrous convolution. The traditional U-Net semanticsegmentation network was improved as the backbone network, and the residualconvolution unitwas used to replace the originalU-Net convolution unit to increasethe depth of the network and avoid the disappearance of gradients. To detect morefeature information, a multi-branch hole convolution module was added betweenthe encoding and decoding modules to extract semantic features, and the expansionrate of the hole convolution was modified to make the network have a bettereffect on the small target category segmentation. Finally, the remote sensing imagewas classified by pixel to output the remote sensing image semantic segmentationresult. The experimental results show that the accuracy and interaction ratio of theproposed algorithm in the ISPRS Vaihingen dataset are improved, which verifiesits effectiveness.展开更多
文摘Image segmentation is an important basic link of remote sensing interpretation.High-resolution remote sensing images contain complex object information.The application of traditional segmentation methods is greatly restricted.In this paper, a remote sensing semantic segmentation algorithm is proposed basedon ResU-Net combined with Atrous convolution. The traditional U-Net semanticsegmentation network was improved as the backbone network, and the residualconvolution unitwas used to replace the originalU-Net convolution unit to increasethe depth of the network and avoid the disappearance of gradients. To detect morefeature information, a multi-branch hole convolution module was added betweenthe encoding and decoding modules to extract semantic features, and the expansionrate of the hole convolution was modified to make the network have a bettereffect on the small target category segmentation. Finally, the remote sensing imagewas classified by pixel to output the remote sensing image semantic segmentationresult. The experimental results show that the accuracy and interaction ratio of theproposed algorithm in the ISPRS Vaihingen dataset are improved, which verifiesits effectiveness.
