基于多尺度特征模糊卷积神经网络的遥感图像分割
Remote sensing image segmentation based on multi-scale feature fuzzy convolutional neural network
马翔悦 1徐金东 1倪梦莹2
作者信息
- 1. 烟台大学计算机与控制工程学院,山东 烟台 264005
- 2. 烟台大学物理与电子信息学院,山东 烟台 264005
- 折叠
摘要
为解决高分辨率遥感图像"同谱异物、同物异谱"的不确定性以及大量空间信息利用率低的问题,提出一种基于多尺度特征的模糊卷积神经网络模型.该模型在长跳跃连接部分加入模糊学习模块去除噪声特征,缓解类别间的不确定性;利用多孔空间金字塔池化融合多尺度特征,提取完备的空间上下文信息,提升分割性能.试验结果表明,该模型在Potsdam数据集和Vaihingen数据集上的整体准确度分别达到 92.65%和 93.19%,明显优于现有流行的深度学习模型,能够显著提升高分辨率遥感图像的语义分割性能.
Abstract
In order to solve the uncertainty of"same spectrum of different objects,the same object with different spectrum"and the low utilization rate of large amount of spatial information in high-resolution remote sensing images,a fuzzy convolutional neural net-work model based on multi-scale features was proposed.The fuzzy learning module was added to the long jump connection to re-move noise features and ease the uncertainty between classes.The multi-scale features were fused by atrous spatial pyramid pooling,and complete spatial context information was extracted to improve segmentation performance.The experimental results showed that the overall accuracy of the model on Potsdam dataset and Vaihingen dataset reached 92.65%and 93.19%,respectively,which were significantly better than the existing popular deep learning models and could significantly improve the semantic segmentation per-formance of high-resolution remote sensing images.
关键词
模糊学习/多孔空间金字塔池化/多尺度特征/编码器-解码器/卷积神经网络Key words
fuzzy learning/atrous spatial pyramid pooling/multi-scale features/encoder-decoder/convolutional neural networks引用本文复制引用
基金项目
国家自然科学基金面上项目(62072391)
国家自然科学基金地区科学基金(62066013)
出版年
2024
NETL
NSTL
万方数据