结合上下文信息与多层特征融合的遥感道路提取
Remote Sensing Road Extraction Combining Contextual Information and Multi-Layer Features Fusion
陈果 1胡立坤1
作者信息
- 1. 广西大学电气工程学院,广西 南宁 530004;广西大学先进测控与智能电力研究中心,广西 南宁 530004
- 折叠
摘要
现有的U-Net虽然为遥感图像道路提取提供了较为理想的解决方案,但由于其缺乏对全局信息的关注,模型对于上下文信息的提取能力不足.为了进一步提高道路提取的准确度与完整度,提出一种结合上下文信息与多层特征融合的context&multilayer features-UNet(CMF-UNet),该模型利用金字塔特征聚合模块融合多层特征,并引入多尺度上下文信息提取模块用于加强上下文信息捕获能力.在Massachusetts Roads和CHN6-CUG两个数据集上进行实验验证,结果表明,所提方法能够有效提升道路分割精度,相较于原U-Net,CMF-UNet在Massachusetts Roads数据集上的召回率、F1分数和交并比分别提升了5.77个百分点、2.02个百分点和2.62个百分点,在CHN6-CUG数据集上的召回率、F1分数和交并比分别提升6.47个百分点、1.53个百分点和2.04个百分点.
Abstract
Although the existing U-Net provides an ideal solution for remote sensing road extraction,its lack of attention to global information leads to the model's insufficient ability to extract contextual information.In order to further improve the accuracy and completeness of road extraction,context&multilayer features-UNet(CMF-UNet),which utilizes a pyramid feature aggregation module to fuse multi-layer features and introduces a multi-scale contextual information extraction module to enhance the contextual information capture capability,is proposed.Experimental validation is conducted on two datasets,Massachusetts Roads and CHN6-CUG,and the results show that compared with U-Net,CMF-UNet improves recall,F1-score,and intersection over union on the Massachusetts Roads dataset by 5.77 percentage points,2.02 percentage points,and 2.62 percentage points respectively;on the CHN6-CUG dataset,recall,F1-score,and intersection over union are improved 6.47 percentage points,1.53 percentage points,and 2.04 percentage points,respectively.
关键词
图像处理/U-Net模型/多尺度上下文/注意力机制/条带池化Key words
image processing/U-Net model/multi-scale contextual information/attention mechanism/strip pooling引用本文复制引用
基金项目
国家自然科学基金(61863002)
广西壮族自治区重点研发计划(桂科AB21220039)
出版年
2024
NETL
NSTL
万方数据