DSMC:全局—局部特征混合编码的遥感影像道路分割模型OA

Aiming at the problems of fragmented segmentation results and error segmentation of small objects when existing methods being used to segment the internal road structure of remote sensing images,we proposed a remote sensing image road segmentation model based on global-local feature hybrid encoding.We used deep separable convolution and shift window TransformerV2 to hybrid encode the details of roads and global context information.In the decoding stage,we employed content-aware reassembly of features and hybrid local channel attention units to filter negative samples and perform fine-scale upsampling of small-size features,and used unified focal loss as the training loss function to guide the model to learn road sample features through parameter weighting.Experimental results show that the F1-scores of this model on the public datasets CHN6-CUG and DeepGlobe Road are 90.68%and 91.35%,respectively,which are 4.26%and 3.81%higher than those of Swin-UNet model.Compared with PSPNet,UNet++and other models,the proposed model exhibits the best performance.