SAA-UNet:特征信息融合网络的遥感图像分割
SAA-UNet:Remote Sensing Image Segmentation Based on Feature Information Fusion Network
金维 1李佳田 1段烨1
作者信息
- 1. 昆明理工大学国土资源工程学院,昆明 650000
- 折叠
摘要
针对星载遥感图像分割精度低、边缘分割模糊问题,提出一种高低层特征信息融合网络模型.以U-Net网络模型为基础,在编码器中加入注意力机制模块获取图像低层特征信息,并在解码器中利用语义嵌入分支将图像的高层信息与低层信息进行融合,在编码器末端利用不同空洞扩张率的混合空洞卷积构建相应模块.为验证网络有效性,以WHDLD数据集和DeepGlobe-Road数据集作为数据源,将SAA-UNet模型与常用语义分割模型进行对比.实验结果显示,SAA-UNet模型整体分割精度优于对比模型,对小 目标地物的分割效果更好.在WHDLD数据集中,平均交并比和类别平均像素准确率分别高于次优模型0.013和0.027.此外,本文采用DeepGlobe-Road数据集进行泛化性.结果表明,本文模型可以有效提高星载遥感图像的分割精度.
Abstract
Aiming at the problems of low accuracy and blurred edge segmentation of satellite-based remote sensing image segmentation,this paper proposes a high and low level feature information fusion network model.Based on the U-Net network model,the attention mechanism module is added in the encoder to obtain the low-level feature information,and the semantic embedding branch is used in the decoder to fuse the high-level information with the low-level information,and a module is constructed at the end of the encoder by using the mixed hole convolution with different hole expansion rates.To verify the effectiveness of the network,the WHDLD dataset and DeepGlobe-Road dataset are used as data sources,and the SAA-UNet model is compared with the current commonly used semantic segmentation network models.The experimental results show that the overall segmentation accuracy of SAA-UNet model is better than that of the comparison model,and the segmentation effect on small target features is better.The mean intersection ratio and category mean pixel accuracy in the WHDLD dataset are higher than that of the suboptimal model by 0.013 and 0.027,respectively.To verify the model generalization ability,the DeepGlobe-Road dataset is used for validation.The results show that SAA-UNet can effectively improve the segmentation accuracy of satellite-based remote sensing images.
关键词
深度学习/语义分割/混合空洞卷积/注意力机制/语义嵌入分支Key words
deep learning/semantic segmentation/hybrid null convolution/attention mechanism/semantic embedding branch引用本文复制引用
基金项目
国家自然科学基金地区科学基金(41561082)
出版年
2024
NETL
NSTL
万方数据