Super-resolution generative adversarial network applied to images filtered by Fourier transform circle mask

Abstract

This paper uses a super-resolution generative adversarial network (SRGAN) with ×4 upscaling to recover a high-resolution colored image from its low-resolution Fourier single-pixel image down-sampled by circle mask at low sampling rates. The results showed that FSPI with SRGAN improves image quality at the same sampling rate. At 5% coverage, SRGAN can improve the perceptual image quality at the same level as that of a higher coverage mask. Increasing epoch value also showed increased and more stable average scores, which the method can further improve with training. This study opened the possibility of increasing the image quality of downsized and masked images at a low sampling rate and a low epoch value model training for colored images with ×4 upscaling factor, concluded visually and numerically.