基于生成对抗网络的遥感图像去云算法

计算机视觉中的许多问题可以抽象为将输入图像“转换”成对应的输出图像，图像转换算法是许多计算机视觉问题的通用解决方案，例如语义分割、风格转换等。本文将以遥感图像去云作为图像转换的特例，研究基于生成对抗网络的图像转换算法。提出基于残差模块的生成模型可以对单幅遥感图像进行厚云和薄云的去除；同时提出的多尺度判别网络以及VGG损失函数，有效地解决了复杂场景的云雾遮挡问题。实验结果表明，本文提出的图像转换算法在遥感图像薄云数据集上峰值信噪比提升了1.64 dB，在厚云数据集上峰值信噪比提升了1.92 dB，同时生成的无云遥感图像和真实的无云图像具有较高的结构相似性。

Cloud Removal Algorithm of Remote Sensing Image Based on GANs

Many problems in computer vision can be abstracted as “converting” an input image into a corresponding output image, which is a general solution to many computer vision problems, such as semantic segmentation, image style transfer, etc. In this paper, the remote sensing image cloud removal is used as the special case of image conversion, and the image conversion algorithm based on Generative Adversarial Networks (GANs) is studied. The GANs based on the residual module is proposed to remove the thick cloud and thin cloud from single remote sensing image. At the same time, the proposed multi-scale discriminator and VGG loss function can effectively deal with the cloud occlusion problem of complex scenes. The experimental results show that the proposed image conversion algorithm increases the peak signal-to-noise ratio on the remote sensing image thin cloud dataset by 1.64 dB and increases the peak signal-to-noise ratio by 1.92 dB on the thick cloud dataset. At the same time, the generated cloud-free remote sensing images have high structural similarity with the real cloud-free images.