一种多尺度卷积神经网络道路提取方法

A Road Extraction Method Based on Multiscale Convolutional Neural Network

针对神经网络训练过程存在分辨率不断降低和梯度消失的共性问题,提出一种基于多尺度卷积神经网络的遥感影像道路提取方法。首先,针对网络前向传播过程分辨率逐渐降低的问题,采用子影像训练网络模型,增强网络对细节信息的获取,然后应用多尺度卷积学习获取不同维度的分层特征,解决由分辨率下降导致的信息缺失问题;其次针对网络在反向传播阶段中出现的梯度消失问题,融入残差连接限制梯度过度更新,改善网络的深度受限问题;最后,针对网络深度和宽度的挖掘导致的网络训练效率问题,使用全局均值池化优化全连接层中大量的冗余数据。大量遥感影像实验结果表明,相对于U-Net网络和经典卷积网络,该方法在Accuracy和F 1值上均具有较大优势。

Owing to the commonality of decreasing resolution and disappearing gradient during the process of convolutional neural network training,we propose an improved method of multi-scale convolution neural network for road extraction in this paper.First of all,in view of the problem that the resolution of the network forward propagation process is gradually reduced,the sub-image training network model is adopted to enhance the network to obtain the detailed information;afterwards,we apply the multi-scale convolution learning method which solves the problem of information loss caused by the decrease of resolution by acquiring hierarchical features of different dimensions.Next,we aim at ameliorating the gradient disappearance of the network in the back propagation process.We incorporate residual connection to limit the over-updating of gradients and improve the depth-constrained problem which is attributed the success to the disappearance of gradients.Finally,global mean pooling is used to optimize a large number of redundant data in the full-connection layer for the efficiency of network training caused by the mining of network depth and width.The experimental results of a large number of aerial images show that there is a great advantage in the training time and road extraction accuracy compared with the U-Net network and traditional convolution network.In addition,the algorithm shows better connectivity and integrity for road extraction in images.