全卷积神经网络遥感影像道路提取方法

Road Extraction from Remote Sensing Image Based on Fully Convolutional Networks

针对人工选取简单特征提取道路效果不理想以及深度神经网络隐藏层信息应用较少的现状,提出一种基于全卷积神经网络的遥感影像道路提取方法。采用初始区域获取、中心线提取、中心线校正的工作流程对资源三号影像进行道路提取。首先自动标注训练样本,完成全卷积神经网络训练,借助卷积层等隐藏层提取的复杂特征获取道路区域;然后依据道路长宽比、形态学运算和格拉斯-普克(Douglas-Peucker,DP)算法完成干扰图斑滤除和断裂区域连接等工作;最后使用Zhang-Suen算法提取中心线,并利用网络首层卷积结果进行中心线校正。实验结果表明,该方法能借助自主学习的特征和网络隐藏层信息实现道路较好提取,不同实验区域中平均准确度在90%以上。

This paper proposes a road extraction method for remote sensing image. This method has three main steps, including initial region extraction,center line extraction and center line correction. Firstly, using fully eonvolutional networks to gain the initial region. And then,we perfect the initial region with some tricks,such as Douglas-Peucker algorithm. Secondly, vectorizing the region to extract the center line. Finally, using the result from the first layer in the networks to correct the centerline. The experimental results indicate that this method can get complete road centerline.