It has been demonstrated that GAN-based algorithms can generate realistic images in single image super-resolution. However, these methods usually generate undesired artifacts in the accompanied images. We proposes a new GAN-based super resolution method to further improve the performance of super-resolved results. In this fashion, we introduce dense compression unit as our basic unit. Then, we use an additional noise into the generator to enhance the quality of generator network. To enhance the supervision of texture recovery, we use a novel quality aware function that is inspired by the SSIM index as excellent regularizer for GAN objective functions. Finally, we demonstrate our method in extensive experiments that the generated images has more realistic textures and it has a great potential in remote sensing tiny-object detection.
Image Super-Resolution Using Quality Aware Generative Adversarial Networks
Lect. Notes Electrical Eng.
2021-10-30
11 pages
Article/Chapter (Book)
Electronic Resource
English
Image Super-Resolution Using Quality Aware Generative Adversarial Networks
British Library Conference Proceedings | 2022
|