Generative Adversarial Network in Wavelet Domain for Single Image Super-resolution

Generative Adversarial Network in Wavelet Domain for Single Image Super-resolution

Nowadays, generative adversarial network for single image super-resolution has achieved superior performance. However, there are few studies on large scaling factor, and the reconstruction performance is relatively poor. Therefore, this paper tackles the above problem by proposing a generative adversarial network in wavelet domain to reconstruct high-resolution image. Specifically, the generator takes low-resolution image as input, and generates wavelet packet decomposition coefficients corresponding to the reconstructed high-resolution image. Then, the discriminator constructs adversarial loss to constrain the generation process of the generator in the wavelet domain. Finally, the high-resolution image is reconstructed based on the inverse wavelet packet transform. The experimental results on CelebA dataset show that our proposed method can achieve better performance than that of comparison methods. The minimum improvements of PSNR and SSIM are 1.783dB and 0.013 on scaling factor ×4, and 0.685dB and 0.016 on scaling factor ×8 when compared with the comparison methods.