基于双通道卷积神经网络的视频目标移除取证算法

针对现有数字视频目标移除取证算法的伪造帧识别准确率低的问题，本文提出了一种基于双通道卷积神经网络的视频目标移除取证算法。该算法利用双通道结构，分别提取视频绝对帧差图像的RGB特征和噪声特征，并利用双线性池化对二者进行特征融合，而后通过分类层输出视频帧的分类结果，从而有效地识别经过篡改的视频帧。其中，RGB通道能够发现绝对帧差图像中不自然的篡改边界和对比度，噪声通道能够发现原始区域和篡改区域之间噪声的不一致性。此外，算法在网络前端增加了预处理层来放大篡改视频帧的伪造痕迹。实验结果显示，所提算法有效地提高了伪造视频帧的识别准确率，且相对于传统的单通道网络结构，双通道特征融合的方式取得了更好的检测性能。 

Video Forensics for Object Removal Based on Two-Channel Convolutional Neural Network

In order to deal with the problem that the existing video object removal forensics methods had low recognition accuracy of the forged frame, a video forensics algorithm based on two-channel convolutional neural network was proposed in this paper. This method used a two-channel structure to extract the RGB feature and noise feature of the absolute frame difference image respectively, and the bilinear pooling was utilized to perform feature fusion on the two features. The classification result of the video frames was output through the classification layer, and this method could then identify the forged frame effectively. The RGB channel was able to find the unnatural tampering boundary and contrast in the absolute frame difference image, and the noise channel was able to find the inconsistency of noise between the original area and the tampered. In addition, the algorithm added a preprocessing layer to amplify the forged traces of tampered video frame. The experimental results revealed that the proposed detection algorithm effectively improves the recognition accuracy of forged video frame, and compared with the traditional single-channel network structure, the pattern of two-channel feature fusion achieved better detection performance.