多尺度特征增强的街景绿色景观分割方法  

Multi-scale feature-enhanced method of green landscape segmentation for street views

作　　者：程勇[1] 王沂萱 任周鹏[2] 王军[1] 顾雅康 CHENG Yong;WANG Yixuan;REN Zhoupeng;WANG Jun;GU Yakang(School of Software,Nanjing University of Information Science&Technology,Nanjing 210044,China;State Key Laboratory of Resources and Environmental Information System,Institute of Geographic Sciences and Natural Resources,Chinese Academy of Sciences,Beijing 100101,China)

机构地区：[1]南京信息工程大学软件学院,南京210044 [2]中国科学院地理科学与资源研究所资源与环境信息系统国家重点实验室,北京100101

出　　处：《测绘工程》2025年第1期11-21,共11页Engineering of Surveying and Mapping

基　　金：国家自然科学基金资助项目(41975183,41975184,42071377)。

摘　　要：针对街景图像中景观复杂多样且多种景观相互遮挡,绿色景观分割效果存在相似景观错分、边界分割模糊、细节丢失等问题,提出一种多尺度特征增强的城市绿色景观分割网络。在编码部分改进多尺度残差网络提取上下文信息以区分相似景观,同时构建多级特征聚合增强模块增强目标特征的边缘细节信息。增加双注意力机制,在局部特征上建模丰富的上下文联系。最后,将多级特征聚合增强模块同样引入解码器,并融合多层级特征来提高目标信息的恢复能力完善边缘信息。在公共街景数据集Cityscapes与自制数据集StreetData的消融实验表明,该网络与基础网络相比,平均交并比分别提高2.96%和5.57%。此外,在两个数据集上进行对比实验,该网络较对比模型平均交并比分别高1.25%~5.29%和1.52%~6.95%。定量分析与实验结果表明,该方法能够有效识别街景的绿色景观,实现高精度的城市绿色景观数据提取。To address the challenges arising from the complex and diverse nature of landscapes in street view images,such as misclassification,blurry boundary segmentation,and loss of details,we propose MFDNet,the Multi-Scale Feature-Enhanced Urban Green Landscape Segmentation Network.In the encoding stage,we utilize an improved multi-scale residual network to extract contextual information and distinguish between similar features.Concurrently,we introduce a feature enhancement module to improve the edge and detail information of target features.To capture rich contextual dependencies,we incorporate a dual-attention mechanism to model local features effectively.Moreover,we integrate the feature enhancement module into the decoder,allowing for the fusion of multi-level features to enhance the recovery of target information and refine edge details.Through ablation experiments conducted on the Cityscapes dataset and our homemade dataset StreetData,we demonstrate that MFDNet achieves an average improvement of 2.96%in intersection ratio and 5.57%in merger ratio compared to the base network.Furthermore,comparison experiments on the two datasets highlight the superior performance of MFDNet over the comparison model,exhibiting higher average intersection ratios of 1.25%~5.29%and merger ratios of 1.52%~6.95%.These experimental results confirm the effectiveness of MFDNet in accurately identifying green landscapes in street views and extracting urban green landscape data with high precision.

关 键 词：深度学习 街景图像 多尺度特征增强 城市绿色景观 语义分割 

分 类 号：P237[天文地球—摄影测量与遥感] TP391[天文地球—测绘科学与技术]