Video Frame Interpolation Based on Visibility Guidance and Texture Consistency Constraints

Existing deep learning-based video frame interpolation methods often suffer from structural misalignment and detail loss in occluded regions or areas with complex textures. To address these issues, this paper proposes a novel video frame interpolation method, termed VGTC (Video Frame Interpolation based on Visibility Guidance and Texture Consistency Constraints), which enhances the accuracy of interpolated frames in occluded areas and improves global texture continuity. The visibility guidance module estimates pixel-wise visibility weights at the intermediate timestamp based on bidirectional motion cues, enabling adaptive fusion of temporal information and effectively reducing occlusion-induced artifacts. To further improve perceptual quality, a texture consistency constraint is introduced via a pseudo-supervised mechanism that enforces local structural similarity between the interpolated frame and its adjacent frames, preserving texture coherence and structural fidelity. Extensive experiments demonstrate the efficiency and effectiveness of the proposed method. Compared with DAIN, VGTC achieves nearly 10 times faster inference speed and improves PSNR by 0.38 dB on the Vimeo-Triplets-Test dataset, while producing more visually pleasing results, especially in regions with complex motion and high-frequency details.