中国空间科学技术

中国空间科学技术 ›› 2022, Vol. 42 ›› Issue (6): 99-108.doi: 10.16708/j.cnki.1000-758X.2022.0088

• 论文 • 上一篇    下一篇

基于流的遥感图像超分辨率重建

任术波，孟倩，吴钻   

  1. 1 中国空间技术研究院，北京100094
    2 武汉大学 计算机学院，武汉430072 
  • 出版日期:2022-12-25 发布日期:2022-11-09

Flow-based super-resolution reconstruction of remote sensing images

REN Shubo，MENG Qian，WU Zuan   

  1. 1 China Academy of Space Technology，Beijing 100094，China
    2 School of Computer，Wuhan University，Wuhan 430072，China
  • Published:2022-12-25 Online:2022-11-09

PDF (PC)

74

摘要： 为了提高遥感图像超分辨率重建的质量，提出了一种基于流的遥感图像重建算法。首先，在Glow模型的基础上引入改进后的RRDB架构用于低分辨率图像特征提取，通过构建更多层和连接以提升训练的稳定性。然后，以一种纯数据驱动的流模型来训练分布的参数，通过最大化负的对数似然的方法进行优化，得到该算法的损失函数。实验证明该模型在网络训练过程中能够快速达到稳定收敛的状态，且具有很强的泛化能力。用重建出的图像质量对比SRCNN、SRGAN、ESRGAN，经过测试后发现，提出的算法远远优于SRCNN算法，与其他算法相比也有明显优势。重建出的图像不仅在指标上有所提升，例如与SRCNN相比，PSRN和SSIM分别提升了15%和40%，且人眼观察时有更高的清晰度，高频细节更为丰富。

关键词: 超分辨率重建, 流模型, Glow模型, RRDB架构, 负对数似然损失

Abstract: In order to improve the quality of super-resolution reconstruction of remote sensing images，a flow-based remote sensing image reconstruction algorithm was proposed.First，the improved RRDB architecture was introduced on the basis of the Glow model for low-resolution image feature extraction，and more layers and connections were built to improve the stability of training.Then，a pure data-driven flow model was used to train the parameters of the distribution，and the method of maximizing the negative log-likelihood was optimized to obtain the loss function of the algorithm.Experiments show that the model can quickly reach a stable convergence state during the network training process and has a strong generalization ability.The reconstructed image quality was compared with SRCNN，SRGAN，ESRGAN.After testing，it is found that the proposed algorithm is far superior to the SRCNN algorithm，and that it also has obvious advantages compared with other algorithms.The reconstructed image not only has improved indicators，for example，PSRN and SSIM increase by 15% and 40% respectively compared with SRCNN,but also has better clarity and richer high-frequency details.

Key words: super-resolution reconstruction, Flow, Glow, RRDB, negative log-likelihood