View article

Video matting is developed to replace the foreground object in video frames by forecasting their alpha matte, which was originally used in film special effects production, advertisement production, and live streaming. However, it can be used as a malicious tampering operation, and hence its passive forensics is especially important as the key foundation of judicial forensics. A key challenge in the current detection of video matting tampering is how to locate the complete tampered region without generating false alarms, not only that tampering detection of video needs to take into account the temporal dimension, but successive frames need to be analyzed to detect forged video regions. In this paper, a dual-branch network is constructed by integrating the object contour information to capture rich tampering traces and contour features, which are introduced by focusing on the information of the entire video sequence. In addition, a manipulation contour detection module and a feature enhancement module are integrated to identify tampering regions better. Extensive experimental results on publicly manipulated and synthetic manipulated datasets show that the proposed method can accurately locate tampered regions and outperform state-of-the-art video matting forensic methods.