Deep Feature Fusion Network for Compressed Video Super-Resolution

The majority of conventional video super-resolution algorithms aim at reconstructing low-resolution videos after down-sampling. However, numerous low-resolution videos will be further compressed to adapt to the limited storage size and transmission bandwidth, leading to further video quality degradation. Significantly, the noise brought by compression often has a strong correlation with the content of the video frame itself. If we super-resolve compressed video frames directly, the noise may be amplified, leading to loss of important information or lower super-resolution performance. To ease those problems, we present an end-to-end deep feature fusion network with ordinary differential equation and dual attention mechanism for joint video compression artifacts reduction and super-resolution. The proposed network commendably enhances the spatial-temporal features fusion of different depths, improves the acquisition of meaningful information ability, and perfects reconstruction quality. In addition, we leverage several skip connections to fuse the captured in-depth feature information and the shallow to prevent information loss. The experimental results show that our proposed method is competent to reduce bit-rate and have excellent quality improvement effectively.