大気歪み画像モデルを組み込んだ深層学習によるリモートセンシング画像の画質改善

本研究的目的是使用深度3D卷积神经网络（CNN）来提高遥感图像的质量。尽管提高遥感图像的质量是一项令人满意的任务，但在参数少、处理时间短的情况下提高基于CNN的方法的性能是非常困难的。因此，我提出了一个新的2D CNN网络，使用并行连接的骨干，其架构包括剩余连接和通道注意机制。该工作已被ACCV 2020年视觉语义分析机器学习和计算研讨会接受。此外，我提出了一种新的多光谱图像融合方法，该方法使用所提出的轻量级3D VolumeNet模型（已被IEEE Transactions on Image Processing，2021接受）和使用其他模态高分辨率图像的纹理转移方法相结合。实验结果表明，该方法在客观准确性评价、效率和视觉主观评价方面均优于现有方法。因此，我计划将这项工作提交给IEEE Transactions on Geoscience and Remote Sensing，总体而言，研究进展与原计划基本一致。我研究并参考了各种最先进的方法，然后建立了我的原始模型。值得注意的是，本文提出的方法不仅在精度上优于现有方法，而且具有更快的处理速度和更低的保存模型的硬件要求，适合于实际应用。

The purpose of this research is to enhance the quality of remote sensing images using deep 3D convolutional neural networks (CNNs).Improving the performance of CNNs-based methods with a few parameters and short processing time is very difficult, although it is a desirable task to improve the quality of remote sensing images. Thus, I proposed a new 2D CNN network using a parallel-connected backbone, the architecture of which consists residual connections and channel-attention mechanism. This work has been accepted by ACCV Workshop on Machine Learning and Computing for Visual Semantic Analysis, 2020.In addition, I proposed a new multi-spectral image fusion method using a combination of the proposed lightweight 3D VolumeNet model (which has been accepted by IEEE Transactions on Image Processing, 2021) and the texture transfer method using other modality high-resolution images. The experimental results show that the proposed method outperforms the existing methods in terms of objective accuracy assessment, efficiency and visual subjective evaluation. Consequently, I plan to submit this work to the IEEE Transactions on Geoscience and Remote Sensing.Overall, the progress of the research is basically in line with the original plan. I studied and referred to various state-of-the-art methods and then built my original models. It is worth noting that the proposed methods not only can exceed the existing methods in accuracy, but also has a faster processing speed and lower hardware requirements for saving the model, so they are suitable for practical applications.