Research On Semantic Segmentation Network Model Of High Resolution Remote Sensing Image

Semantic segmentation is one of the core contents of high spatial resolution remote sensing image analysis and understanding.Currently,deep learning methods are a research important issues in semantic segmentation of high-resolution remote sensing images.Based on deep learning methods,this dissertation explores the problem of semantic segmentation of high-resolution remote sensing images from three aspects.First,a deep learning based on semantic segmentation networks lead to loss of high-frequency information in remote sensing images and inaccurate edge segmentation.Aiming at this problem,design a dual decoupling semantic segmentation network model,which decouples the extracted two-level features into Edge features with high-frequency characteristics and Body features with low-frequency characteristics.The resulting Edge feature and Body feature maps are fused to improve the semantic segmentation performance of high-resolution remote sensing images.Furthermore,considering a loss function for Edge and Body is proposed to optimize the learning of ground features and their Edge and Body parts.Secondly,in view of the limitation of the convolution operation to capture the global feature representation,design a high-resolution remote sensing image semantic segmentation model based on the encoder-decoder structure of Swin transformer,and the Swin transformer model is used as the backbone network in the encoder stage to extract global features.In order to obtain multi-scale high-level semantic information,a atrous spatial pyramid pooling module is combined in the encoder stage to obtain multi-scale semantic features to alleviate some information loss caused by downsampling and subsequent upsampling.In the decoder stage,the low-level and high-level features are fused by skip connections to enrich the spatial information of the features,thereby improving the semantic segmentation accuracy of high-resolution remote sensing images.Finally,aiming at the problem that a large amount of label data is needed in the deep learning-based remote sensing image semantic segmentation task,design a high-resolution remote sensing image semantic segmentation model based on unsupervised multi-level domain adaptation.Different methods are used to reduce the gap between domains for the multi-level features of the target domain and the source domain,and different loss functions are used to optimize the parameter of the model.Experiments are carried out on two high-resolution remote sensing image datasets,Compared with the results of the existing semantic segmentation network models,and the effectiveness of this method is confirmed.