SGCV-Net: An Efficient and Lightweight Semantic Segmentation Network for High-Resolution Remote Sensing Images based on Semantic Guidance and CNN-ViT

The semantic segmentation technology can deeply analyze the land-coverage information in high-resolution remote sensing image (HRRSI), achieve accurate classifications and recognitions at the pixel level, and then provide a strong data support for various applications, such as urban planning, disaster management, and environmental monitoring. However, existing lightweight semantic segmentation networks often encounter challenges such as insufficient multi-scale feature extraction capabilities, blurred boundary contours, and a large number of parameters and computations. To address these issues, an efficient and lightweight semantic segmentation network based on the semantic guidance, the convolutional neural network (CNN), and the visual transformer (ViT), named SGCV-Net, was proposed. In the proposed SGCV-Net, an efficient and lightweight encoder with a novel depth-wise and partial convolution pair (DPP) module and a novel Sobel-enhanced MobileViT (SMViT) block was firstly designed to effectively extract semantic features. Then, a neck layer with a novel semantically guided data-dependent upsampling asymptotic feature pyramid network (DUs-AFPN) was designed for accurate multi-scale feature fusion. Finally, a minimal multilayer perceptron (MLP) layer was employed as the decoder. The proposed SGCV-Net was comparatively tested on the Gaofen image dataset. The experimental results showed that the SGCV-Net achieved the highest mean intersection over union (mIoU) with only 3.68 M parameters and 7.61 GFLOPs and the mIoU reached 83.16%.