Video Frame Interpolation Based on Unsupervised Deep Learning

Video Frame Interpolation (VFI) technology aims to enhance video smoothness by synthesizing intermediate frames between two consecutive frames, and it holds significant application value in areas such as industrial manufacturing and multimedia processing. However, traditional optical flow estimation-based methods show limited adaptability and struggle to maintain stable performance in complex scenes, whereas the latest deep learning-based methods face challenges such as the difficulty of acquiring labeled data and suboptimal cross-domain transfer performance. To address these issues, this paper proposes an unsupervised learning-based VFI method, which consists of two components: training data construction and model inference. Specifically, several keyframe pairs are selected from the original video sequences to construct training samples. Subsequently, using the inference framework trained in an unsupervised manner, intermediate frames can be predicted between any pair of adjacent frames to generate high-frame-rate videos. Experimental results demonstrate that the proposed method significantly outperforms traditional methods across multiple metrics, exhibiting both high effectiveness and strong generalization capability.