Research on atmospheric turbulence-degraded image restoration based on generative adversarial networks

Abstract

The imaging equipment working in the atmosphere will not only be limited by the performance of the imaging system, but also be affected by turbulence. In the fields of astronomical observation, ground-based remote sensing and remote monitoring, there is an urgent need for corresponding methods and technologies to eliminate the impact of atmospheric turbulence and obtain clear images. With the development of computer technology, atmospheric optics theory and image processing technology, more and more researchers hope to combine deep learning technology with atmospheric turbulence theory to reduce the impact of turbulence on imaging and obtain clear and stable images. In this paper, a turbulence image restoration technique based on Generative Adversarial Networks (GAN) is proposed, which is divided into generator network and discriminator network. The generator network is used to convert blurred images affected by turbulence into clear images. The discriminator network is used to compare the converted image with the real clear image to determine whether the image is real or generated. After the whole GAN is optimized and trained, the image transformed by the generator and the real and clear image cannot be distinguished from each other. Because the training of the GAN requires a large number of corresponding samples, it is difficult to obtain the images affected and unaffected by turbulence at the same time in real life, so this paper uses the statistical characteristics of turbulence to simulate a large number of images affected by turbulence. We used the trained GAN model to simulate turbulence image restoration and got some achievements.