Single Image Super-Resolution with U-Net Generative Adversarial Networks

摘要

It is very challenging to recover a High Resolution (HR) image with real texture from a single Low Resolution (LR) image. SRGAN[1] first applied GAN(Generative Adversarial Network) in the field of image Super-Resolution, and restored a relatively real texture HR. However, SRGAN's reconstructed HR often contains unreal artifacts and distortions. Subsequently ESRGAN[2] has improved this problem, but it still has the shortcomings of not sharp edges of objects and slow reconstruction speed. To further improve the perception quality and accelerate the reconstruction speed, we proposed an image super-resolution algorithm (SR) based on U-Net GAN [3]. As the basic block, we introduced the Residual-in-Residual Self-calibrated Convolution with Pixel Attention block(RRSCPA) [4]. From [5] to heuristic, we design the discriminator into a U-shaped structure, which can provide per-pixel feedback to the generator and promote the generator to generate a more realistic HR. Finally, we replaced the VGG-based perceptual loss[6] with the LPIPS perceptual loss[7] function. Our proposed U-Net SRGAN achieves consistently better visual quality with more realistic and natural textures than ESRGAN.