改进的UNet3+网络高分辨率遥感影像道路提取
Improved UNet 3+network high-resolution remote sensing image road extraction
周家厚 1普运伟 2陈如俊 1邓云龙 1周鑫城 1李俊1
作者信息
- 1. 昆明理工大学国土资源工程学院,昆明 650093
- 2. 昆明理工大学国土资源工程学院,昆明 650093;昆明理工大学计算中心,昆明 650051
- 折叠
摘要
为解决UNet3+网络随深度加深出现大量融合冗余操作以至于模型训练时间过长而导致在道路提取中运用较少的问题,对UNet3+网络进行改进,通过删减UNet3＋的网络层级使用Bottleneck模块替换原有网络中的卷积层,保留网络特征融合能力的同时降低网络复杂度,并引入混合注意力机制优化模型,构建了 一个新的网络模型.将改进方法与几种典型的道路提取模型做对比.实验结果表明:(1)所提方法相较于Unet3+网络在、Recall、IOU、ACC四个指标上分别提升了 4.72％、2.46％、4.84％、2.01％,均优于对比算法;(2)对比几个经典的特征提取模型,改进的模型具有更好的识别效果,在道路提取的精度、连接性、完整性等方面均表现出优越性.
Abstract
To solve the UNet3+network with depth deepening a large number of fusion redundant operation that the model training time is too long and resulting in road extraction using less problems,the UNet3+network improve-ment,by cutting UNet3+network hierarchy using Bottleneck module to replace the convolution layer in the original network,retain the network feature fusion ability and reduce the network complexity,and introduce hybrid attention mechanism optimization model,build a new network model.The improvement method is compared with several typical road extraction models.The experimental results show that:(1)compared with Unet3+network,the proposed method improves by 4.72％,2.46％,4.84％and 2.01％respectively,all better than the comparison algorithm;(2)com-pared with several classical feature extraction models,the improved model has better recognition effect,and phenoty-ping in the accuracy,connectivity,integrity and other aspects of road extraction.
关键词
深度学习/注意力机制/UNet3+/道路提取/跳跃连接Key words
deep learning/attention mechanism/UNet3+/path extraction/skip connection引用本文复制引用
出版年
2024
NETL
NSTL
万方数据