Satellite Image Super-Resolution by 2D RRDB and Edge-Enhanced Generative Adversarial Network

Abstract

With the gradually increasing demand for high-resolution images, image super-resolution (SR) technology has become more and more important in our daily life. In general, high resolution is often accomplished by increasing the accuracy and density of the sensor. However, such an approach is too expensive on the design and equipment. Particularly, increasing the sensor density of satellites incurs great risks. Inspired by EEGAN, some parts of networks: Ultra-Dense Subnet (UDSN) and Edge-Enhanced Subnet (EESN) are modified. The UDSN is used to extract features and obtain high-resolution images which look clear but are deteriorated by artifacts in the intermediate stage, while the EESN is used to purify, enhance and extract the image contours and uses mask processing to eliminate the image corrupted by noise. Then, the restored intermediate image and the enhanced edge are combined to become a high-resolution image with clear contents and high authenticity. Finally, we use Kaggle open source, AID, WHU-RS19, and SpaceWill datasets to perform the test and compare the SR results among different models. It shows that our proposed approach outperforms other state-of-the-art SR models on satellite images.

Keywords

• edge enhancement
• generative adversarial network (GAN)
• residual in residual dense block (RRDB)
• satellite images
• super-resolution (SR)