Enhancing Image Quality via Style Transfer for Single Image Super-Resolution
Xin Deng · IEEE Signal Processing Letters · 2018
Recently, by feat of the Generative Adversarial Network (GAN), single image super-resolution (SISR) has achieved great breakthroughs in enhancing the perceptual image quality. However, since the network is trained by minimizing the perceptual loss, the GAN based SISR method (SRGAN) [1] results in images with very low objective quality, i.e., peak signal-to-noise ratio (PSNR). In this letter, we aim to solve this problem in an image style transfer way, to generate an image with similar perceptual quality as SRGAN, but with much higher objective quality. Moreover, we propose a threshold-based method to automatically alter the objective and perceptual quality of the reconstructed image through adjusting only one parameter. Experimental results show that our method can achieve more than 1.6 dB PSNR improvement over SRGAN with similar Mean Opinion Score value. Also, with the same objective quality, our method can provide significantly better perceptual results than other state-of-the-art SISR methods.