计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (18): 260-267.DOI: 10.3778/j.issn.1002-8331.2201-0173

• 图形图像处理 • 上一篇    下一篇

基于分流的高光谱遥感图像超分辨重建

徐英豪,吕玉超,刘斯凡,朱习军   

  1. 青岛科技大学 信息科学技术学院,山东 青岛 266061
  • 出版日期:2022-09-15 发布日期:2022-09-15

Super Resolution Reconstruction of Hyperspectral Remote Sensing Image Based on Shunt

XU Yinghao, LYU Yuchao, LIU Sifan, ZHU Xijun   

  1. School of Information Science and Technology, Qingdao University of Science and Technology, Qingdao, Shadong 266061, China
  • Online:2022-09-15 Published:2022-09-15

摘要: 高光谱图像(hyperspectral image,HSI)每个像素包含大量光谱带,而HSI的空间分辨率较低。高光谱图像超分辨率技术可以有效提高空间分辨率。为了解决高光谱遥感图像的空间信息分布不均匀,超分辨率重建时占据相同计算量导致重建工作不够细化的问题,提出了一种多层级分流和细节增强的高光谱遥感图像超分辨率重建框架。即通过子图分流网络对高光谱遥感图像进行预先分流,改进带细节增强的多尺度Retinex算法对图像的高低频信息进行分离,再使用不同复杂程度的分支网络分别进行重建,重建工作更加细化具体,提高重建效果和性能。实验证明该方法可以在视觉质量、指标测量和分类应用方面优于传统的基于CNN的方法,在SRE和MPSNR指标方面分别提高了4.18%和9.35%。

关键词: 高光谱, 遥感图像, 超分辨率, 卷积神经网络

Abstract: Hyperspectral image(HSI) contains a large number of spectral bands per pixel, while the spatial resolution of HSI is low. The spatial resolution is effectively improved by hyperspectral image super-resolution technology. In order to solve the problem that the spatial information of hyperspectral remote sensing image is unevenly distributed and occupies the same amount of computation during super-resolution reconstruction, which leads to the lack of refinement of reconstruction work, a multi-level shunt and detail enhancement framework for hyperspectral remote sensing image super-resolution reconstruction is proposed in this study. That is, the hyperspectral remote sensing image is pre shunted through the sub image shunting network. The maximum number of large squares that allow exchange with details to display high and low frequency details and use networks from different sources to change locations. The reconstruction work is more detailed and specific, and the reconstruction effect and performance are improved. Experiments show that this method is superior to the traditional CNN based method in visual quality, index measurement and classification application, and the SRE and PSNR indexes are improved by 4.18% and 9.35% respectively.

Key words: hyperspectral, remote sensing images, super resolution, convolutional neural network