UFlow-Net: A Unified Approach for Improved Video Frame Interpolation

Video frame interpolation is a critical method in video processing, enabling the synthesis of intermediate frames to enhance video quality and frame rates. While conventional methods rely on optical flow estimation and kernel-based approaches, they often struggle with complex scenes and rapid motion. Recent advancements in deep learning have demonstrated the potential of hybrid architectures to address these challenges. This paper proposes UFlow-Net, a novel hybrid framework that integrates flow estimation and kernel-based methods within a modified U-Net architecture. The model incorporates a flow-enhanced encoder-decoder structure with advanced feature fusion via skip connections to preserve spatial-temporal information. Evaluated on the Vimeo90K, UCF101, and Middlebury datasets, UFlow-Net achieves state-of-the-art performance with peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) metrics. Results demonstrate its robustness in handling intricate backgrounds and dynamic motion, making it a significant advancement in the field of video frame interpolation.