Lightweight Semantic Segmentation Network for Remote Sensing Urban Scenes Based on Selective Kernel

In the context of remote sensing images, semantic segmentation networks must possess robust global information extraction capabilities due to the presence of indistinct object edges and low segmentation accuracy between similar classes. Consequently, this paper introduces a semantic segmentation network SC-SKNet based on selective kernel as decoder for realtime remote sensing city scene segmentation. The network uses ResNet18 as the encoder and incorporates a selective kernel attention module in the decoder, integrating large kernel convolution to enhance the network's capacity for processing both global and local information while achieving a balance between segmentation speed and accuracy. Experimental results demonstrate that SC-SKNet not only exhibits accelerated processing speeds but also significantly enhances segmentation accuracy compared to current advanced lightweight networks. The method was tested on the Vaihingen and Potsdam datasets, and the average mIoU and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{mF}_{1}$</tex> scores reached <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$83.3 \%$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$87.2 \%$</tex>, respectively, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$91.0 \%$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$93.0 \%$</tex>, respectively. The overall accuracy OA is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$91.5 \%$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$91.7 \%$</tex>, respectively, and the inference speed can reach 127.2FPS when processing <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1024 \times 1024$</tex> resolution images on NVIDIA RTX 4090 GPU.