Trans-CNN GAN: Self-Attention Generative Adversarial Networkd for Remote Sensing Image Super-Resolution

In recent years, deep learning based super-resolution algorithms have made breakthrough progress, especially the adversarial characteristics of Generative Adversarial Network (GAN), which is conducive to learning local details and generating more realistic images. However, most algorithms tend to overlook the large size of remote sensing images. While they focus on local features, they neglect attention to global information during the reconstruction process, resulting in the loss of information and a pronounced blurring of edge details in the reconstructed images. This article proposes a Trans-CNN GAN architecture based on self-attention mechanism, combining the advantages of GAN and Transformer. The generator adopts a CNN structure. In order to adapt to the reconstruction of large-sized remote sensing images, a discriminator containing the Transformer structure is selected. The self-attention of the Transformer is used to compensate for the convolutional generator’s lack of focus on global information, and compared with discriminator based on CNN structure to explore the impact of different discriminator structures on the generator. The method proposed in this article can effectively improve the reconstruction quality of remote sensing images. The PSNR and SSIM of the reconstructed images have been improved to 0.082 and 0.038, respectively, and the visual effect has been greatly improved.