针对全卷积神经网络多次下采样操作导致的道路边缘细节信息损失和道路提取不准确的问题,本文提出了多尺度特征融合的膨胀卷积残差网络高分一号影像道路提取方法。首先,通过目视解译的方法制作大量的道路提取标签数据;其次,在残差网络ResNet-101的各个残差块中引入膨胀卷积和多尺度特征感知模块,扩大特征点的感受野,避免特征图分辨率减小和道路边缘细节特征的损失;然后,通过叠加融合和上采样操作将各个尺寸的道路特征图进行融合,得到原始分辨率大小的特征图;最后,将特征图输入Sigmoid分类器中进行分类。实验结果表明:本文方法的提取精度优于经典全卷积神经网络模型,准确率达到了98%以上,有效保留了道路的完整性及其边缘的细节信息。

This paper aimed to solve the problems of road edge detail information loss and inaccurate road extraction due to multiple downsampling operations of the fully convolutional neural network. Thus, a road extraction method of GF-1 remote sensing images based on dilated convolution residual network with multiscale feature fusion is proposed. First, numerous labels for road extraction are generated through visual interpretation. Second, dilated convolution and multiscale feature perception modules are introduced in each residual block of the residual network, namely, ResNet-101, to enlarge the receptive field of the feature points without reducing the feature map resolution and losing the detailed edge information. Third, through superposition fusion and upsampling operations, the road feature maps of various sizes are fused to obtain the feature maps of the original resolution size. Finally, for classification, the feature maps are input into the Sigmoid classifier. The experimental results indicate that the proposed method is more accurate than the conventional fully convolutional neural network models, with the accuracy rate being more than 98%. The proposed method effectively preserves the integrity and detailed edge information of the road area.