结合超像素与Transformer的高分辨率遥感图像语义分割算法OA

A Transformer-based semantic segmentation model SegFormer-MSFF is proposed for automatic and accurate semantic segmentation of high-resolution remote sensing images.In order to solve the problem that when the existing Transformer-based method performs grid slicing and serializes the image,it may result in discontinuous semantic segmentation results between image slices,a sub-network with lightweight multi-scale superpixel information extraction is introduced based on convolutional neural network.The boundary information of the image is preserved through multi-scale dynamic superpixel information,and it is fused with the middle semantic features at all levels to enhance the Transformer's ability to capture the continuity between image slices.A joint loss function is proposed to achieve dynamic superpixel segmentation,which weights the superpixel segmentation loss and the semantic segmentation loss to obtain the objective function of the overall model,allowing the overall model to simultaneously explore the semantic and superpixel features of the image.Experiments on the high-resolution aerial remote sensing images datasets Vaihingen and Potsdam showed that the mean intersection over union of SegFormer-MSFF is 1.4%and 0.9%higher than that of SegFormer.