改进U-Net模型的无人机影像在建道路分类  

Classification of Road Under Construction in UAV Image Based on Improved U-Net Model

作　　者：胡荣明[1] 魏青博 竞霞[1] 廖雨欣 任乐宽 HU Rongming;WEI Qingbo;JING Xia;LIAO Yuxin;REN Lekuan(College of Geomatics,Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]西安科技大学测绘科学与技术学院,西安710054

出　　处：《遥感信息》2023年第5期8-15,共8页Remote Sensing Information

基　　金：国家自然科学基金项目(42171394)。

摘　　要：针对无人机影像背景复杂,城市在建道路分类易被相似目标、建设设施等信息干扰的问题,提出了基于改进U-Net模型的无人机影像在建道路提取模型。为获取更深层次的边界细节信息,采用Res2net结构分阶替换原有U-Net网络的卷积层,提高网络下采样深度;增加CBAM双注意力机制模块引于各分块特征信息之后,对空间和通道进行重新校准,强调道路特征,校正模型参数;引入改进的Dense ASPP模块,与前层次的细节信息拼接,增强道路区域上下文信息的获取能力。结果表明,所提出的改进U-Net网络训练的提取模型在精确率、召回率、F1分值、平均交并比等评价指标上,均优于传统的U-Net、DeeplabV3+、HRnet等网络模型,可有效提取建设道路各阶段信息,针对在建道路项目的施工进度监测提供方法支持。Aiming at the problem that the background of UAV image is complex and the classification of urban roads under construction is easily disturbed by similar targets,construction facilities and other information,an improved U-Net model based on UAV image road under construction extraction model is proposed.In order to obtain deeper boundary details,Res2net structure is used to replace the convolution layer of the original U-Net network in order to improve the network down-sampling depth.The CBAM dual attention mechanism module is added to introduce the block feature information,and then the space and channel are recalibrated,emphasizing the road features,and correcting the model parameters.The improved Dense ASPP module is introduced to splice with the details of the previous level to enhance the ability to obtain the context information of the road area.The results show that the improved U-Net network training extraction model proposed in this paper is superior to the traditional network models such as U-Net,DeeplabV3+and HRnet in terms of accuracy,recall rate,F1 score,average intersection and combination ratio,and can effectively extract the information of each stage of road construction and provide method support for the construction progress monitoring of road projects under construction.

关 键 词：无人机影像 语义分割 深度学习 注意力机制 道路提取 

分 类 号：P237[天文地球—摄影测量与遥感] TP79[天文地球—测绘科学与技术]