CVEGAN: A perceptually-inspired GAN for Compressed Video Enhancement

Di Ma, Fan Zhang, David R. Bull

Abstract: Highlights•Presenting a novel block structure, Mul2Res which is the first use of a nested residual learning structure with various kernel sizes.•Employing enhanced residual non-local blocks and enhanced convolutional block attention modules to improve the representational capability of the network.•Designing a new GAN training methodology, Relativistic SphereGAN to achieve better training performance.•Proposing a novel perceptual loss function to further optimise video quality during training.