Multi-scale compression artifact attention-driven compressed video quality enhancement

Abstract

Compressed video quality enhancement is crucial for mitigating artifacts introduced by video coding. Video compression often results in unevenly distributed artifacts across different regions of video frames, leading to significant quality variations. Existing algorithms treat all regions uniformly, ignoring these localized differences, limiting their ability to extract high-quality information from reference frames and accurately reconstruct residuals. Additionally, larger temporal gaps between reference and target frames can cause alignment errors, which propagate during fusion and degrade performance. To address these challenges, we propose a Multi-scale Compression Artifact Attention (MSCAA) module that captures artifact distribution, guiding the model to focus on distorted regions. We also introduce a Progressive Fusion Stage that sequentially fuses reference frames based on temporal proximity, reducing error propagation and enhancing temporal coherence. The experimental results show that the proposed method improves average ΔPSNR by 5.15% and ΔSSIM by 3.35% compared to the state-of-the-art method, demonstrating its superior performance in quality enhancement.