| With the arrival of information age,digital pictures and videos play a significant role of our life and are admitted to be a part of judicial evidence.The ubiquitous surveillance cameras infers the importance of digital videos.However,videos can be easily forged by different kinds of video editing software.Thus more and more researchers have focus on how to ensure the authenticity and integrity.Double compression is of prime significance in video forensics,since a video must undergo the process of double compression if it is forged meaningfully.Existing methods to detect double compression have high performance when the primary quantiser scale is different from the second quantiser scale.While these methods fail to detect double MPEG compression with the same quantiser scale due to the slight difference between singly compressed and doubly compressed videos.It is a challenge to detect double MPEG compression with the same quantiser scale.In this paper,we analyze from an encoding and decoding standpoint,then some methods are proposed to detect double MPEG compression with the same quantiser scale based on the encoding features.Firstly,we propose a novel method to detect double MPEG compression based on the statistical feature of macroblock mode.Macroblock mode consists of macroblock type and motion vector and the algorithm utilizes the convergence of macroblock mode to detect double compression.It is found that in the process of repeatedly compressing a MPEG video using the same quantiser scale,the number of macroblocks with different macroblock modes decreases monotonically between two sequential compressions.The statistical feature of macroblock mode is extracted and is fed to the support vector machine to classify the singly compressed and doubly compressed videos.Experiments have demonstrated the effectiveness of the proposed method based on macroblock mode and is robust to different quantiser scales and encoders.Another experiment shows that it is also effective to detect double MPEG compression with the same bit rate.Secondly,an error-based method is proposed to detect double MPEG compression with the same quantiser scale.It can be observed that the rounding error and truncating error generated during decoding have different characteristics between the singly compressed and doubly compressed videos.The feature vector involves the number of error blocks,the max absolute value,the mean value and variance of both rounding error and truncating error.And the support vector machine is also used to detect double compression.It can be found that the accuracy is quite high when the quantiser scale is small.At last,an improved method based on the combined features consisted of macroblock mode and error feature is proposed.The error-based feature is less discriminative when the value of quantiser scale is large.In contrast,the method based on macroblock mode has higher performance when the quantiser scale is large.Thus we propose a novel method combining macroblock mode and error features at last in order to overcome the limitations of both features.Furthermore,we shrink the combined features so that the model can maintain simple and be more general.Experimental results have demonstrated that the statistical feature combined with macroblock mode and error features improve the overall accuracy and robustness to different quantiser scales and encoders. |
