摘要
现有的高光谱图像超分辨率算法通常将高光谱图像裁剪为多个小尺度图像块,这将破坏图像整体性。并且,由于高光谱图像具有大量的光谱信息,导致现有的超分辨率算法很难充分挖掘高光谱图像的空间和光谱信息。为了解决上述问题,文章提出了一种基于空谱邻域和多流融合的超分辨率网络(Spatial-Spectral Adjacent Domain Fusion Network,SSADFN)。首先,提出了一种空谱邻域输入模块,该模块充分利用图像空间块和光谱邻域信息,可以有效地捕捉图像内部结构的整体信息。其次,设计了双向循环网络结构。该结构包括正向单元和反向单元两部分,并对两种单元设置不同的多流融合残差块,充分挖掘图像的空间细节信息和光谱细节信息。在两个高光谱遥感图像数据集Chikusei和Pavia Centre上的实验结果表明,提出的方法相较于前沿高光谱超分辨率算法具有更好的结果和更少的参数。
Existing hyperspectral image super-resolution algorithms usually crop the hyperspectral image into multiple small-scale image patches,which destroys the image integrity.Moreover,the hyper-spectral image has a large amount of spectral information,which makes it difficult for existing su-per-resolution algorithms to fully mine the spatial and spectral information of the hyperspectral image.This paper proposed a super-resolution network based on spatial-spectral adjacent domain network and multi-stream fusion to solve the above problems(SSADFN).First,a spatial-spectral adjacent domain input module is proposed,which makes full use of the image space patches and spectral adjacent information and can effectively capture the overall information of the internal structure of the image.Second,a bi-directional recurrent network structure is designed.The struc-ture consists of two parts:the forward unit and the backward unit.Different multi-stream fusion residual blocks are set for the two kinds of units to fully exploit the spatial and spectral detail infor-mation of the image.Experimental results on two hyperspectral remote sensing image datasets,Chikusei and Pavia Centre,show that the proposed method has better results and fewer parameters compared to the state-of-the-art hyperspectral super-resolution algorithms.
作者
王清
Qing Wang(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai)
出处
《建模与仿真》
2025年第5期92-104,共13页
Modeling and Simulation
关键词
超分辨率
循环神经网络
光谱相关性
注意力机制
Super-Resolution
Recurrent Neural Network
Spectral Correlation
Attention Mechanism
