基于改进DeepLabv3+的遥感影像道路提取算法  

Road Extraction Algorithm for Remote Sensing Images Based on Improved DeepLabv3+

作　　者：王谦 何朗[1] 王展青[1] 黄坤[1] WANG Qian;HE Lang;WANG Zhanqing;HUANG Kun(School of Science,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学理学院,武汉430070

出　　处：《计算机科学》2024年第8期168-175,共8页Computer Science

基　　金：国家自然科学基金(62176194)。

摘　　要：道路提取可以帮助人们更好地理解城市环境,是城市交通和城市规划等方面的重要部分,随着深度学习与计算机视觉的发展,利用基于深度学习的语义分割算法从遥感影像中提取道路的技术趋于成熟。针对现有的深度学习道路提取算法存在的提取速度慢和容易受背景环境因素干扰而产生漏分割、不连续等问题,提出了一种基于ECANet注意力机制和级联空洞空间金字塔池化模块的轻量化算法CE-DeepLabv3+。首先,将主干特征提取网络更换为轻量级的MobileNetv2,减少参数量,提高模型的执行速度;其次,通过增加空洞空间金字塔池化模块的卷积层进一步扩大感受野,再级联不同特征层来增强语义信息的复用性,从而加强对细节特征的提取能力;再次,加入ECANet注意力机制,抑制背景环境中的干扰因素,聚焦道路信息;最后,采用改进的损失函数进行训练,消除了道路与背景样本不均衡对模型性能产生的影响。实验结果表明,改进算法的性能优良,与原始DeepLabv3+算法相比,在分割效率、分割精度上有较大的提升。Road extraction can help us better understand the urban environment and is an important part of urban transportation and planning.With the development of deep learning and computer vision,the use of deep learning-based semantic segmentation algorithm to extract roads from remote sensing images has become increasingly mature.However,existing deep learning road extraction algorithms suffer from slow extraction speed and susceptibility to background environmental factors,resulting in missed segmentation and discontinuity.To address these issues,a lightweight algorithm called CE-DeepLabv3+based on ECANet attention mechanism and cascade atrous spatial pyramid pooling module is proposed.Firstly,the main feature extraction network is replaced with the lightweight MobileNetv2 to reduce parameter volume and improve model execution speed.Secondly,the convolution layers of the atrous spatial pyramid pooling module are further expanded to increase the receptive field,and different feature layers are cascaded to enhance semantic information reuse,thereby improving the ability to extract fine-grained features.Thirdly,the ECANet attention mechanism is added to suppress environmental interference and focus on road information.Finally,an improved loss function is used for training to address the impact of road and background sample imbalance on model performance.Experimental results show that the improved algorithm achieves excellent performance,with significant improvements in segmentation efficiency and accuracy compared to the original DeepLabv3+algorithm.

关 键 词：语义分割 遥感影像 道路提取 注意力机制 DeepLabv3+ 级联空洞空间金字塔池化 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]