由于多光谱图像所含细节信息较少,导致其在各领域中应用受到限制.因此,如何提升多光谱的空间分辨率成了重中之重.多光谱图像超分辨率重建(image super reso-lution reconstruction,SR)旨在从单一多光谱图像中通过重建算法重构出超分辨率多光谱(super resolution multi-spectrum,SRMS)图像,但现有方法重构的SMSR图像中仍存在边缘细节模糊问题.提出了 一种新的多层级对比学习的多光谱图像超分辨率重建来缓解上述问题.首先,构建自重构网络提取全色(panchromatic,PAN)图像的高频特征和多光谱图像的低频特征.其次,在特征嵌入空间通过多层级对比学习引导SRMS图像学习PAN图像高频特征并远离低分辨多光谱图像的模糊属性.定性和定量评估表明,所提出的方法性能优异.
Multispectral image super-resolution reconstruction based on multilevel contrast learning
Due to the lack of detailed information contained in multispectral images,its application in various fields is limited.Therefore,how to improve the spatial resolution of multispectral becomes the top priority.Multi-spectral image super resolution reconstruction(SR)aims to reconstruct the SRMS image from a single multi-spectral image by the reconstruction algorithm,but the edge details in the reconstructed SMSR image are still fuzzy.In this paper,we propose a new multispectral image super-resolution reconstruction based on multilevel contrast learning to alleviate the above problems.Specifically,we first extract the high-frequency features of panchromatic(PAN)images and the low-frequency features of multispectral images through a self-reconstruction network.Then,the SRMS image is approached to the PAN image and pulled away from the low-resolution multispectral image by multi-level contrast learning in the feature embedding space.Qualitative and quantitative evaluation shows that the proposed method performs well.
万方数据
维普
