Compressed Video Quality Enhancement Based on Optimal Reference Selection

Abstract

Compressed video enhancement (CVE) leverages both inter-frame information and codec bitstream cues, but doing so effectively requires choosing references that align with the prediction structure. In contrast, most existing methods simply stack neighboring frames, which often yields redundant references and makes it difficult to control error propagation along prediction chains. We address this gap with Optimal Reference Selection (ORS), which selects a single reference for each target frame while explicitly accounting for the codec prediction structure via prediction-path independence, low temporal-layer preference, and temporal proximity. Based on ORS, we present ORS-CVE, which combines the selected reference with bitstream priors for enhancement. Our model performs robust long-range alignment through motion-vector guided hierarchical optical flow refinement, and it restores textures and fine details using low-QP residual cues from the lowest temporal layer. Experiments demonstrate consistent improvements in both quantitative metrics and visual quality over prior CVE approaches. Source code available at https://github.com/RenGM123/ORSCVE