Publication

Abstract : In this paper, researchers present a generative adversarial network (GAN) technique for improving the perceived quality of compressed videos. In a single model, our approach takes into account and adapts to various quantization parameters (QPs). The attention module makes use of global receptive fields, which are capable of capturing and aligning long-distance correlations between successive frames and can improve the perceived quality of films. The proposed deep Kalman model allows for the recovery of decoded frames, and video restoration is particularly specified as a Kalman filtering process. The less noisy preceding restored frame is employed recursively instead of the noisy previous decoded frames, resulting in high quality restored frames. A deep Kalman filtering network is developed within the suggested technique by merging many deep neural networks to estimate the related states in the Kalman filter. The upgraded frame and its surrounding frames are sent into the deep network, and in the first step, features at various depths are retrieved. Following a succession of upsampling and convolution layers, retrieved features are input into attention blocks to study global temporal correlations. The QP-conditional adaptation module, which utilises the related QP information, processes the output features at the end. With similar performance, a single model may be utilised to improve adaptively to different QPs without the need for separate models for each different QP value. Experimental findings show that the suggested ENet performs better than cutting-edge compressed video quality improvement methods.