基于能量特性与目标跟踪的数字视频对象篡改被动取证方法研究

With the development of multimedia technology, people can easily tamper with video, making it more urgent to determine the source and authenticity of video. Existing object forgery detection algorithms generally suffer from inefficiency and poor robustness, which are difficult to meet the needs of practical application. Accordingly, focusing on video object forgery, the following three aspects research would be carried out: (1) in view of video object deletion forgery, a passive forensics method based on energy factor and adaptive foreground extraction is proposed; (2) aiming at video object duplication forgery, a passive forensics method based on local coherency error and fast object tracking is proposed; (3) in allusion to video object splicing, a passive forensics method based on deep learning is studied. The main innovations include: (1) a fast method based on energy factor to locating suspicious sequences is proposed, which can be applied to the detection of multiple types of tampering; (2) constructed local coherency error is used in feature matching, which improves the matching speed and has strong universality; (3) object tracking and foreground extraction are applied to video passive forensics, which improves the detection efficiency and robustness of the algorithms. The research results of this project will effectively reduce detection time and improve the accuracy and robustness, thus playing an important role in information security, judicial evidence collection, social stability and other fields.

随着多媒体技术的发展，人们可以方便地对视频进行篡改，使得对于数字视频来源和真实性判定显得更为迫切。已有的视频对象篡改检测方法仍存在时间效率低、鲁棒性弱等问题，难以满足实际应用需求。本项目对视频对象篡改的被动取证开展三方面研究：1）针对视频对象删除篡改，研究基于能量因子与自适应前景提取的被动取证方法；2）针对视频对象复制粘贴篡改，研究基于局部一致性误差和快速目标跟踪的被动取证方法；3）针对视频对象拼接篡改，研究基于深度学习的被动取证方法。主要创新点包括：1）构造局部一致性误差应用于特征匹配，提高了匹配速度，并具有较强的通用性；2）提出基于能量因子的可疑序列定位方法，该方法速度快、可应用于多种篡改方式；3）将目标跟踪、前景提取应用于视频篡改取证，提高了算法的检测效率和鲁棒性。本项目的研究成果将有效减少视频篡改检测时间，提高准确性和鲁棒性，从而在信息安全、司法取证、社会维稳等领域发挥重要作用。

随着多媒体技术的发展，人们可以方便地对视频进行篡改，使得对于数字视频来源和真实性判定显得更为迫切。已有的检测方法仍存在时间效率低、鲁棒性弱等问题，难以满足实际应用需求。针对上述问题，本项目对视频对象篡改开展研究，具体取得的研究成果如下：1）提出了基于自适应特征融合和通道组损失函数的深度伪造篡改检测方法，从不同的角度探讨了三种不同的特征，并设计了一种合适的方法来结合人为设计的特征和深度学习，极大的提升了时间效率；2）提出了基于自适应注意力和约束差异一致性的深度伪造篡改检测方法，该方法能够自适应不同大小的生成人脸，拥有更好的泛化性；3）提出了基于双重监督的篡改被动取证方法，加强对高级特征的提取，捕获更完善的更深层次的语义信息，从而实现像素级别上的精确定位；4）提出了基于边界引导的多特征多尺度监督的拼接检测方法，该方法结合了多尺度特征信息和边缘特征信息，能够在预测篡改主体区域和边缘区域上同时获得高准确性；5）提出了基于特征融合的篡改被动取证方法，利用不同级别特征的性质充分融合特征，以充分捕获篡改特征。本项目的上述成果将在信息安全、司法取证、社会维稳等领域发挥重要作用。