A fast forgery frame detection method for video copy-move inter/intra-frame identification

Digital video is critical visual evidence in various fields and is easily manipulated under different techniques such as the popular video copy-move forgery. In the past decades, although machine intelligence has been widely adopted to detect the forgery in digital images automatically, It still remains a very challenging detection task for carefully-crafted copy-move forgery in digital video for three reasons: (i) A video of medium length containing hundreds of frames already incurs a prohibitive computational cost; (ii) Similar backgrounds in contiguous frames are easily mistakenly detected as copy-move forgery regions, resulting to a large number of false alarms; (iii) Most state-of-the-art methods cannot detect video copy-move inter-frame or intra-frame forgeries; To effectively address these issues, a fast forgery frame detection method for video copy-move inter/intra-frame identification is proposed: (i) The sparse feature extraction and matching speed-up the algorithm processing and reduce the time cost greatly (Defect (i)); (ii) The adaptive two-pass filtering and copy-move frame-pair matching can address the similarity problem (Defect (ii)) to locate truly forgery frame-pairs (FFP); (iii) Based on the results of these FFP, the type of video copy-move forgery detection can be identified (Defect (iii)). Furthermore, the copy-move frame-pair matching algorithm locates truly FFP, thus further reducing the computation cost and false alarm for detecting the inter/intra-frame forgery efficiently and effectively (Defect (i)). Finally, based on the truly FFP, the video can be checked for forgery or original. If there is no truly FFP, the video is considered as the original one. Otherwise, the video is checked if the forgery is inter-frame (i.e., truly FFP frames are two different frames) or intra-frame (the same frame). The experimental results show that our proposed algorithm achieves higher detection accuracy and higher robustness (false alarm = 2 and F1 = 0.90) in the whole GRIP dataset than the existing state-of-the-art methods under various adverse conditions.