| Digital video forensics plays an important role in judicial forensics,media reports,e-commerce,finance and public security.Although domestic and foreign research teams have proposed a large number of methods,there is little research on forensics for video with complex scenes.At the same time,due to the unique attributes of complex scene video,the recommended requirement for traditional video forensics is difficult to satisfy by the existing algorithms in most real applications.For example,there could be severe motion,illumination noise,and jitter noise in complex scene video.For such videos,only using traditional video forensic methods could easily cause a large number of false detections.Therefore,this project studies in depth the inter-frame tamper detection of complex scenes video,and proposes the following two methods:1.We propose a video tampering detection algorithm that combines spatial constraint and gradient structure information.On the one hand,for video with severe motion,to avoid false detection caused by forensics using only the similarity between adjacent frames,the proposed method introduces the spatial constraint rule to obtain the regions with the rich quantitative correlation that is beneficial to tampering detection.On the other hand,for the feature extraction of blurred video,this project improves the original methods for feature extraction and description to extract the optimal similarity feature of the video on the quantized correlation rich area.Experiments show that compared with the existing algorithms,for the video with severe motion,the proposed algorithm has a lower false detection rate,higher accuracy,and stronger robustness.Furthermore,the algorithm can accurately locate the video tampering point and detect subtle inter-frame tamper.2.We propose a detection method that adapts to brightness and jitter for video inter-frame forgery.On the one hand,classical optical flow approaches are usually invalid for video of changing illumination.Therefore,this project introduces brightness-relaxing factors and intensity normalization to deal with the issue.On the other hand,to solve the problem that most traditional video detection methods could cause false detection on jitter video,the proposed algorithm combines motion entropy and texture smoothness to remove the interference of video jitter,so as to improve the robustness of the algorithm.Extensive experimental results in diverse and realistic forensic set-ups prove that this method is not only suitable for video with dynamic changes in brightness but also jittery video and large-motion video.Compared with traditional video forgery detection methods,the proposed algorithm improves the robustness and the accuracy of video tamper detection in complex scenes. |
