Efficient Networks for Video Quality Enhancement

Abstract

Video restoration has become increasingly important with the growing demand for high-quality video content across various applications. This thesis addresses two fundamental challenges in video restoration: space-time video super-resolution and video deblurring. For space-time video super-resolution, we propose a novel deformable attention network (DANet) that effectively handles both spatial and temporal super-resolution in a unified framework. The network features a deformable interpolation block for accurate frame synthesis and a temporal fusion module for efficient multi-frame information utilization. For video deblurring, we develop a wavelet-based blur-aware decoupled network (WBDNet) that innovatively decomposes the deblurring task into structure recovery and detail enhancement through wavelet transform. The network employs a multi-scale progressive fusion module for structural reconstruction and a blur-aware detail enhancement module that leverages sharpness priors for refined detail restoration. Extensive experiments on multiple benchmark datasets demonstrate that our proposed methods achieve superior performance compared to state-of-the-art approaches in terms of both objective metrics and visual quality, while maintaining reasonable computational efficiency. The methods developed in this thesis advance the field of video restoration and show strong practical value for applications ranging from multimedia entertainment to surveillance systems.

-