JOFMCA: Joint Optical Flow and Motion Compensation Alignment for Old Film Restoration
Jiatong Han, Xueming Li · 2024
Old film restoration is a classic research task in vision, but deep-based methods often fail to achieve satisfactory results. To address challenges like mixed degradation and hidden knowledge extraction from adjacent frames, we propose a deep transformer-based framework of JOFMCA, which incorporates joint optical flow and motion compensation. Specifically, we first design an edge-aware (EA) preprocessing module for reducing noise while protecting edge information. A motion compensation module is then introduced to compensate and rectify for optical flow errors. Finally, a fusion module based on optical flow difference is constructed to handle artifacts in scenarios with motion and mixed degradation. In addition, we construct a novel dataset comprising both synthetic and real-world old film samples to evaluate our proposed model and further enrich the datasets for old film restoration. Extensive quantitative and qualitative experiments on multiple datasets show our method outperforms prior works. Our code will be made publicly available.