一种遥感影像道路提取的改进U-NET模型

Remote sensing images improved U-NET model for road extraction

针对U-NET网络在提取遥感图像道路的过程中出现精度损失和训练速度缓慢的问题,提出了一种改进的U-NET语义分割网络模型。先在编码器和解码器中间加入以级联方式连接的扩张卷积层,在扩大中心部分特征点感受野的同时保留特征图的空间信息,解决了U-NET网络在进行道路提取时精度损失的问题;再针对采用数据增强进行数据集扩充而导致的训练收敛变慢的问题,使用了ELU激活函数,改善了网络训练效率低的问题。实验表明,本文U-NET网络对道路的提取精度得到了明显的提升,可以准确提取出遥感图像中的道路要素且效率较高,并且在与传统U-NET的对比中优势较为明显,其各项指标正确率、召回率和F1综合指标分别达到86.79%、83.12%和84.91%,较传统U-NET分别提升2.06%、1.83%、1.94%,具有较好的应用前景。

To improve the accuracy and training speed in U-NET network road extraction from remote sensing images,a new U-NET semantic segmentation network model is proposed.First,a convolutional layer is connected in cascade between the encoder and the decoder to expand the receptive field of the central feature point while retaining the spatial information of the feature map.This increases the accuracy of the U-NET network in road extraction.In the slower training convergence caused by data augmentation for data set expansion,ELU activation function is used to improve the network training efficiency.Finally,experiments show that the U-NET network has a significant improvement in the road extraction accuracy,by accurate extracting road elements in remote sensing images.There are more advantages in comparison with traditional U-NET obviously.The accuracy rate,recall rate and F1 comprehensive index reached 86.79%,83.12%and 84.91%,respectively,2.06%,1.83%,1.94%higher than traditional U-NET,with a good application prospect.