基于改进U-Net网络的城市道路提取方法研究  

Research on Urban Road Extraction Method Based on Improved U-Net Network

作　　者：谢天怡 梁曦文 徐昇[1] XIE Tianyi;LIANG Xiwen;XU Sheng(College of Computer Science and Technology,Nanjing Forestry University,Nanjing 210037)

机构地区：[1]南京林业大学信息科学技术学院,南京210037

出　　处：《计算机与数字工程》2023年第3期650-656,共7页Computer & Digital Engineering

基　　金：国家自然科学基金青年科学基金项目(编号:62102184);江苏省自然科学基金青年科学基金项目(编号:BK20200784);中国博士后科学基金面上项目(编号:2019M661852);江苏省高等学校大学生创新创业训练计划项目“基于深度学习的遥感图像研究”(编号:202110298091Y)资助。

摘　　要：针对现有方法对高分辨率遥感图像的道路信息提取精度有限的问题,论文选用PyTorch作为深度学习框架,在卷积神经网络(CNN)的基础上选用U-Net模型进行改进,使用ResNet50残差模块替换原模型的卷积部分,增加模型感受野并解决梯度弥散问题,从而增加网络深度来提取目标的深层特征。实验部分使用南京玄武湖附近道路数据集作为马萨诸塞州道路数据集的补充,对比了不同道路场景下改进模型的提取效果,并利用准确率和召回率结合提取结果评价改进模型的性能,总结其适用的场景。实验结果显示,论文改进模型在马萨诸塞州道路数据集上的不同场景中的提取效果较好,而在分辨率更高的南京玄武湖附近道路数据集中的车辆或遮挡物较多路段,道路提取结果出现不连续现象。基于马萨诸塞州道路数据集,改进模型提取结果准确率约为0.8,召回率约为0.75。其中,相较于早期的U-Net网络结构,准确率提高48.14%,召回率提高19.04%;相较于近年有学者提出的改进TGCA-Net网络结构准确率提高19.40%,召回率持平。To address the problem of limited accuracy of road information extraction from high-resolution remote sensing imag⁃es by existing methods,this paper selects PyTorch as the deep learning framework,chooses the U-Net model to improve on the ba⁃sis of convolutional neural network,uses ResNet50 residual module to replace the convolutional part of the original model,increas⁃es the model perceptual field and solves the gradient dispersion problem,so as to increase the network depth to extract the deep fea⁃tures of the target.The experimental part uses the road dataset near Nanjing Xuanwu Lake as a supplement to the Massachusetts road dataset to compare the extraction effect of the improved model under different road scenes,and evaluates the performance of the improved model using the accuracy and recall combined with the extraction results to summarize its applicable scenes.The exper⁃imental results show that the improved model in this paper has better extraction effects in different scenes on the Massachusetts road dataset,while the road extraction results show discontinuities in the road dataset with higher resolution near Nanjing Xuanwu lake where there are more vehicles or occlusions.Based on the Massachusetts road dataset,the accuracy of the improved model extrac⁃tion results is about 0.8 and the recall rate is about 0.75.Among them,compared with the earlier U-Net network structure,the accu⁃racy is improved by 48.14%and the recall rate is improved by 19.04%.Compared with the improved TGCA-Net network structure proposed by some scholars in recent years,the accuracy is improved by 19.40%and the recall rate is the same.

关 键 词：城市道路提取 高分辨率遥感图像 U-Net ResNet50 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]