Detection of Interpolated Frames Based on Multi-Scale Hierarchical Extraction and Hybrid-3D Networks

Video frame interpolation (VFI) enhances video smoothness by creating intermediate frames between existing ones. However, malicious tamperers often use this technique to tamper with video content. This poses a certain threat to information security. To this end, we suggest a deep VFI evidence collection method based on multi-scale hierarchical features. Specifically, our approach integrates a multi-scale feature extraction framework that corrects and enhances fine-scale features through a feedback mechanism. Moreover, a hybrid network architecture is proposed to incorporate 3D convolutions with parallel grouped convolutions, enabling more effective extraction of temporal features from video frames. To foster the integration of information across different dimensions, we employ an ensemble of attention networks in the final stage of our model. In addition, we create the interpolated frame detection dataset in our research. The experimental outcomes demonstrate that our approach surpasses several currently available state-of-theart techniques.