Visual Perception Enhancement for HEVC Compressed Video Using a Generative Adversarial Network

The emergence of generative adversarial network (GAN) promotes the great progress of deep learning generation model. In this paper, generative adversarial network is used to remove the visual artifact of compressed video, and a visual perception enhancement algorithm for HEVC compressed video is proposed. Specifically, after HEVC compression, the reconstructed image is output by GAN generator. The output image can effectively guide the discriminator of GAN to approximate the mapping between the encoded frame and the original frame. The adversarial loss of the generator to keep learning this mapping, which not only improves the visual perception quality of compressed video, but also removes the artifact. Experimental results demonstrate the superiority of our GAN network over other methods in terms of both Perceptual Index and visual quality.