Semantic Segmentation for Remote Sensing Images Based on Swin-Transformer and Multiscale Feature Refinement

Thanks to the development of deep learning, semantic segmentation of remote sensing images has made great advances. However, because of the widespread complex background and variable imaging conditions, the feature consistency of the same classes and the variability of different classes are more difficult to capture, leading to the existing methods still have the space for improvement. Thus, we combine the local convolution operation and swin transformer to design a novel semantic segmentation method (STFR). The proposed STFR can improve the multiscale feature representation and capture the global context information, which consists of three steps. Firstly, the multiscale features of the image is extracted by swin transformer; then, a parallel structure is constructed based on depth-wise separable convolution to refine the deep semantic information; finally, the semantic segmentation result can be obtained. Experimental results on a public remote sensing semantic segmentation dataset indicate the the proposed STFR can obtain a great performance.