HEVC Compression Artifact Reduction with Generative Adversarial Networks

Video compression technology is significant to video transmission and storage. However, compression artifacts arise in videos. Specifically, coarse quantization eliminates video details and degrades visual quality. Most of artifact reduction methods use filter processing or Mean Square Error (MSE)loss that leads to over-smoothing results. Moreover, most of the methods target to reduce single image compression artifact instead of video artifact. In this paper, we present an adversarial learning method with recurrent framework called Video Artifact Reduction Generative Adversarial Network (VRGAN). Our network contains a generator with recurrent framework that improves video consistency, a dense block that enhances receptive field for large transform unit, and a relativistic discriminator that evaluates the relationship between the generated frames and the original high-quality frames. Our VRGAN is able to generate more realistic videos. The effectiveness in reducing video compression artifacts of the method has been demonstrated qualitatively and quantitatively. The performance comparison with previous works shows the superiority of the proposed method.