Video matting tampering detection based on time and space domain traces

Deep learning-based videos usually leave imperceptible traces when tampered with. Tampered videos may be used for malicious video manipulation, which raises privacy and security concerns. Therefore the detection and localisation of tampered video traces is necessary. In this paper, we locate the tampering region by using the traces left behind by time and space domain information, and use VOS as a refinement network to improve the model performance. Firstly, the base network enhances the tampering traces by intra- and inter-frame residuals, and a dual-stream network is designed as an encoder to extract the special diagnosis from the frame residuals. Afterwards, a bidirectional convolutional LSTM and transposed convolution are embedded in the decoder to generate a prediction mask. Afterwards, a VOS network is used to obtain more accurate object boundaries. Extensive experimental results on public and synthetic manipulated datasets show that the proposed method can accurately locate tampered regions and outperforms and is robust to state-of-the-art methods.