融合深度卷积神经网络和Swin Transformer的露天矿遥感图像超分辨率重建
Super-resolution Reconstruction for Remote Sensing Images of Open-pit Coal Mines Based on CNN and Swin Transformer
聂雅琳 1王海军 2石念峰 1刘保罗1
作者信息
- 1. 洛阳理工学院计算机与信息工程学院,河南 洛阳 471023
- 2. 河南科技大学数学与统计学院,河南 洛阳 471023
- 折叠
摘要
针对现有露天矿遥感图像超分辨率重建模型提取特征能力弱、特征利用不充分的问题,提出了一种融合深度卷积神经网络和Swin Transformer网络的露天矿遥感图像超分辨率重建方法.首先,利用卷积神经网络和Swin Transformer网络将露天矿遥感图像映射到全局和局部特征空间,充分提取遥感图像的深层特征;然后,构造了一种基于注意力机制的多尺度特征融合网络,实现遥感图像局部和全局特征的深度融合,强化有效特征表达的区分能力;最后,将深度融合特征作为超分辨率解码模块的输入,重建出高分辨率的露天矿遥感图像.通过在自建露天矿区图像数据集和开源数据集上进行测试,试验结果表明:与当前主流的图像超分辨率重建算法相比,所提方法重构出的超分辨率图像具有较好的视觉感知,在均方根误差方面也低于其他对比方法.
Abstract
Aiming at the problems of weak feature extraction ability and insufficient feature utilization of existing super-resolution reconstruction models of remote sensing images of open-pit mines,a new super-resolution reconstruction method of open-pit remote sensing images based on deep convolutional neural network and Swin Transformer is proposed.Firstly,convolu-tional neural network and Swin Transformer network are used to map the remote sensing images of open-pit mine to the global and local feature spaces,and fully extract the deep features of the remote sensing images.Then,a multi-scale feature fusion net-work based on attention mechanism is constructed to realize the deep fusion of local and global features of remote sensing ima-ges and strengthen the distinguishing ability of effective feature expression.Finally,the deep fusion features are used as the in-put of the super-resolution decoding module to reconstruct high-resolution remote sensing images of open-pit mines.Through testing on the self-built open-pit mine image dataset and open source data set,the experimental results show that compared with the current mainstream image super-resolution reconstruction algorithms,the proposed method has better visual perception and lower RMSE than other comparison methods.
关键词
露天矿/超分辨率重建/深度卷积神经网络/Swin/TransformerKey words
open-pit mine/super-resolution reconstruction/deep convolutional neural networks/Swin Transformer引用本文复制引用
出版年
2024
NETL
NSTL
万方数据