A New Framework Based on Spatio-Temporal Information for Enhancing Compressed Video

Artifacts are generated during video compression; many methods have been proposed to solve this problem. Most of the methods to get better temporal information, using optical flow to estimate temporal motion compensation. However, optical flow could be inaccurate because compressed video might be distorted by various compression artifacts. The optical flow also introduces additional model complexity and computational cost; it is also susceptible to error propagation in challenging scenarios with significant motion and heavy occlusion. To address these problems, we propose an effectively multi-frame with spatio-temporal information-guided quality enhancement network. Our algorithm employs spatio-temporal deformable convolution to aggregate temporal information. We define inputs as a target frame and its neighboring reference frames that can jointly predict offset fields to deform the spatio-temporal sampling positions. Specifically, we enhance the quality of reconstruction compressed video by devising an efficient deformable alignment module with a receptive field block to handle bad cases in Compression Artifact Removal.