SGEU-Net:用于从高分遥感影像中提取道路的空间分组增强注意力网络

Spatial Group-wise Enhanced U-Net for Road Extraction from High-resolution Remote Sensing Images

道路提取是现代路网规划的重要组成部分。近来,许多深度学习方法已被应用于该领域。然而,由于车辆以及树木和建筑物阴影的遮挡,在保持连续性的同时准确提取道路区域仍然是一个问题。论文提出了一种新型的道路提取网络-空间分组增强网络(SGEU-Net),由两个部分构成:一个改进的U-Net编码器-解码器网络和空间分组增强(SGE)注意力模块。SGE模块可以明显改善不同语义子特征在组内的空间分布,产生更可观的统计差异,增强语义区域的特征学习。改进的算法在马萨诸塞州道路数据集上进行实验,结果表明,与当前先进算法相比,所提算法提高了从遥感图像中提取道路的效果。

Road extraction is an integral part of modern road network planning.Recently,many deep learning methods have been applied in this field.However,it is still a problem to extract the road area accurately while maintaining continuity due to the oc⁃clusion of vehicles and the shadows of trees and buildings.This paper presents a novel road extraction network,the Spatial Group-wise Enhanced U-Net(SGEU-Net),builts on two parts,which are an improved Encoder-Decoder U-Net and the Spatial Group-wise Enhanced(SGE)module.The SGE module can significantly improve the spatial distribution of different semantic sub-features within the groups and produce a more considerable statistical variance,enhancing feature learning in semantic regions.The improved algorithm is experimented on the Massachusetts road dataset,and the results show that the proposed algorithm im⁃proves the extraction of roads from remote sensing images compared with current state-of-the-art algorithms.