||Compressed Video Restoration Using a Generative Adversarial Network for Subjective Quality Enhancement
||Dayoung Chun;Tae Sung Kim;Kyujoong Lee;Hyuk-Jae Lee
|| Deep neural network; CNN; GAN; Compressed video restoration; High efficiency video coding
||High Efficiency Video Coding (HEVC) is a widely used video compression standard that minimizes the sacrifice in visual quality. Convolutional neural networks (CNNs) are being used as a post-processing tool for video restoration degraded by compression. Improving on CNN-based video restoration, this paper attempts a new generative adversarial network (GAN)-based video restoration called a compressed video restoration generative adversarial network (CVRGAN). Although a GAN is widely used for perceptual image enhancement in super-resolution and noise reduction, it is not yet used for compressed video restoration. The proposed CVRGAN is the first attempt to utilize a GAN to create the texture of a degraded image, and consequently, to generate detailed textures that were lost due to compression. In order to avoid the side effect of a GAN boosting the blocking and ringing artifacts incurred by compression, the CVRGAN employs a new content loss that is a combination of VGG feature difference, which represents a perceptual loss, and an objective loss, such as mean squared error (MSE) or mean absolute error (MAE). The new loss function is effective in the enhancement of subjective image quality while suppressing artifact boosting. An extensive mean opinion score (MOS) test shows that the CVRGAN achieves an improvement in perceptual quality over previous CNN-based video restoration.