数字图像及视频篡改检测关键技术研究

发布时间：2019-05-29 00:54

【摘要】：各种数字图像、视频编辑软件以及相应教程的普及,导致图像和视频可能成为作恶者的工具,人们长久以来对影像媒体的信任也发生了动摇。数字图像和视频篡改检测的相关理论和技术就是在这种背景下应运而生。数字图像和视频篡改检测属于媒体内容取证范畴,目的是鉴别影像媒体内容的真伪。数字图像和视频从拍摄到成像的过程中,场景中感知对象反射的光线需要经过光学镜头的折射,并经历光学滤波、光电转换、色彩插值、后处理等操作,这个过程中的每一个步骤可以看做是对光线的一次变换,本文将一组变换的有序组合定义为一个变换链,并将数字图像和视频建模为经历了某个变换链的若干感知对象的结构化组合,进而对图像和视频的篡改手段进行了形式化的分析,并将篡改行为在媒体中留下的痕迹归结为两种情况:一是媒体中出现异常相似的感知对象,二是某些感知对象经历了与其他感知对象不同的变换链。在此基础上,本文将图像及视频的篡改检测建模为一个“描述——发现”的过程:对于待检测的媒体,首先找到某种特征,用以描述媒体中的感知对象或感知对象所经历的变换链中的某个环节,进而通过匹配或校验的方式,去发现媒体中异常相似的感知对象,或去发现某些感知对象经历了与其他感知对象不一致的变换链。本文围绕篡改检测模型中的特征构造、匹配和校验方法等关键技术展开研究,解决了数字图像及视频篡改检测技术中存在的若干问题。本文的主要研究工作和创新点包括以下几个方面:第一,提出了一种基于有序序列聚类的特征匹配方法。特征匹配是检测媒体中异常相似对象的关键技术之一,在基于特征点的图像区域拷贝检测方法中,当特征空间中同时存在多个高度相似的特征时,现有的特征匹配方法会漏检大量实际上匹配的特征对。针对该问题,本文提出了基于有序序列聚类的特征匹配方法,并基于贝叶斯分类器实现了聚类过程中参数的自适应选择,显著地提高了合格匹配特征收集的完备性。第二,针对基于特征点的图像区域拷贝检测方法对平滑区域拷贝行为检测能力弱的问题,提出了层次化特征点检测结合特征融合的区域拷贝检测方法。该方法在不显著增加特征点总数的情况下,保证不同区域内特征点的覆盖率。对于平滑区域的特征点,本文构造了基于局部梯度和色彩的融合特征,提高了局部特征在平滑区域内的区分性。第三,提出了基于DCT系数分析的压缩历史不一致检测所应满足的边界条件,并设计了相应的参数求解方法。图像中各感知对象的压缩历史不一致即某些感知对象在压缩编码阶段所经历的变换与其他感知对象不同,这意味着图像有局部区域遭到了篡改。在基于DCT系数分析的JPEG压缩历史不一致检测方法中,通常以篡改成分和非篡改成分的DCT系数分布作为特征来反映各DCT块所经历的压缩历史。而篡改成分和非篡改成分的DCT系数分布的估计则依赖于对二者的混合模型的参数求解。现有的方法普遍以“盲”的方式进行参数估计,因而对参数的估值往往不够准确。考虑到DCT系数所应遵循的实际约束,本文向混合模型对应的似然函数中补充了必要的边界条件,并结合似然函数的平滑特性设计了粗粒度搜索结合梯度上升的参数估计方法,实现了更为准确的篡改检测和定位。第四,针对降质视频,提出了基于位置敏感哈希和帧配准的匹配方法。在降质视频中,受各种降质因素的影响,视频帧局部结构细微变化的累积将导致帧特征产生实质性的变化。因此传统方法中普遍采用的“特征提取——阈值化”的匹配方法很难实现稳定的雷同帧检测。本文基于位置敏感哈希实现视频帧序列的初步匹配,并基于配准技术完成雷同帧的校验。为了实现对视频降质的鲁棒性,本文将视频帧中各区域的稳定性信息编码到配准能量函数中,并基于概率推理的方式近似求解全局最优匹配问题,实现了更为鲁棒的降质视频帧拷贝检测。第五,提出了针对高码率视频的快速帧拷贝检测方法。在高码率视频中,当具有相同内容的帧之间并不存在显著的波动时,为了在保证检测能力的前提下降低帧拷贝检测的时间开销,本文提出了二维视频帧的三维骨架特征并设计了相应的匹配方法。本文首先基于骨架的拓扑信息实现数据筛选,进而基于几何信息进行细粒度的雷同帧判别,实现了高码率视频中的快速帧拷贝检测。第六,提出了基于码流异常突变的视频删/插帧检测方法。基于码流分析的删/插帧检测方法通常以码流中存在异常的周期效应作为码流异常的特征。然而,异常的周期效应并非总能够可靠地检测到。此外,现有的方法不能有效地定位篡改操作发生的位置。本文把码流中各P帧对应的预测残差均值和帧内预测宏块数量的同时突变作为检测删/插帧操作的依据,设计了用于度量预测残差均值和帧内预测宏块数量的变化强度的指标,进而基于这两种指标构造了融合特征,以及用于捕捉码流信息突变的校验方法。在对视频编码参数不做任何约束的情况下,本文的方法能够有效地检测和定位视频帧插入/删除操作。
[Abstract]:The popularity of various digital images, video editing software, and corresponding tutorials has led to the potential for images and videos to become the tools of the perpetrator, and the trust of the image media has been shaken for a long time. The related theories and techniques of digital image and video tamper detection are in this background. The digital image and the video tampering detection belong to the category of the media content forensic, and the purpose of the invention is to identify the authenticity of the image media content. in the process of shooting the digital image and the video from the shooting to the image, the light reflected by the sensing object in the scene needs to pass through the refraction of the optical lens and undergo the operations of optical filtering, photoelectric conversion, color interpolation, post-processing and the like, each step in this process can be seen as a transformation of light, which is defined herein as a transform chain and the digital image and the video are modeled as a structured combination of several perceptual objects that have undergone a transformation chain, In this paper, a formal analysis of the tampering of the image and the video is carried out, and the trace of the alteration behavior in the media is summed up as two cases: one is a perception object in the media which is similar to the other, and the other is that some of the perceptive objects experience different transformation chains than the other perceptual objects. On the basis of this, the process of modeling the tampering detection of the image and the video is a "Description _ Discovery" process: for the media to be detected, a certain feature is first found to describe a link in the transformation chain experienced by the perceived object or the perceived object in the media, Furthermore, by matching or checking, the perceptually similar perceived objects in the media are found, or some sense objects are found to experience a transformation chain that is not consistent with other perceived objects. In this paper, the key technologies such as feature structure, matching and checking method in the tamper detection model are studied, and some problems in the detection technology of digital image and video tampering are solved. The main research work and innovation points of this paper include the following aspects: first, a feature matching method based on ordered sequence clustering is proposed. The feature matching is one of the key technologies for detecting the abnormal similar objects in the media. In the method of image area copy detection based on the feature point, when a plurality of highly similar features exist in the feature space, the existing feature matching method can miss a large number of actually matching feature pairs. In order to solve this problem, the feature matching method based on the ordered sequence clustering is proposed, and the self-adaptive selection of the parameters in the clustering process is realized based on the Bayesian classifier, and the completeness of the collection of the qualified matching features is remarkably improved. Secondly, aiming at the problem that the image area copy detection method based on the feature point is weak in the detection capability of the smooth region copy behavior, a method for detecting the region copy of the hierarchical feature point detection and combining feature fusion is proposed. The method ensures the coverage rate of the characteristic points in different regions without significantly increasing the total number of the feature points. For the feature point of the smooth region, this paper constructs the fusion feature based on the local gradient and the color, and improves the distinguishing between the local feature and the smooth region. Thirdly, the boundary conditions to be satisfied by the non-uniform detection of the compression history based on the DCT coefficient analysis are put forward, and the corresponding method for solving the parameters is designed. The compression history of each of the perceptual objects in the image is not consistent, that is, the transformation experienced by some of the perceptual objects during the compression encoding stage is different from the other perceptual objects, which means that the image has a local area being tampered with. In the JPEG compression history inconsistency detection method based on the DCT coefficient analysis, the compression history experienced by each DCT block is generally reflected by the DCT coefficient distribution of the tamper component and the non-tamper component as a feature. The estimation of the DCT coefficient distribution of the tampered component and the non-tamper component is dependent on the parameter solving of the mixed model of the two. The existing methods are generally used in "blind" to estimate parameters, and therefore the valuation of the parameters is often not accurate. Considering the actual constraints to be followed by the DCT coefficients, the necessary boundary conditions are added to the likelihood function corresponding to the mixed model, and the method of parameter estimation of the coarse-size search combined with the gradient increase is designed in combination with the smoothness of the likelihood function. And the more accurate detection and positioning of the tampering is realized. Fourth, aiming at the quality reduction video, a matching method based on location sensitive hash and frame registration is proposed. In the quality reduction video, the accumulation of fine changes in the local structure of the video frame will lead to a substantial change in the frame characteristics due to the influence of various quality-reducing factors. Therefore, the conventional "feature extraction _ thresholding" matching method in the traditional method is difficult to realize the stable thunder and the frame detection. In this paper, the preliminary matching of the video frame sequence is realized based on the location-sensitive hash, and the verification of the same frame is completed based on the registration technique. In order to achieve the robustness of the video quality reduction, the stability information of each region in the video frame is encoded into the registration energy function, and the global optimal matching problem is solved based on the way of probability reasoning, and a more robust quality-reducing video frame copy detection is realized. Fifth, a fast frame copy detection method for high code rate video is proposed. In the high code rate video, when there is no significant fluctuation between the frames with the same content, in order to reduce the time cost of the frame copy detection under the premise of ensuring the detection capability, the three-dimensional skeleton feature of the two-dimensional video frame is proposed and the corresponding matching method is designed. In this paper, based on the topology information of the skeleton, the data selection is realized, and the fine-grained Lei-frame discrimination is carried out based on the geometric information, and the fast frame copy detection in the high-code-rate video is realized. And sixth, a method for detecting the video deletion/ insertion frame based on the abnormal mutation of the code stream is provided. The method of deleting/ inserting the frame based on the code stream analysis usually takes the periodic effect of the abnormality in the code stream as the characteristic of the code stream abnormality. However, the periodic effect of the abnormality is not always able to be detected reliably. In addition, the existing method cannot effectively locate the position of the tampering operation. in that invention, the mean value of the prediction residual corresponding to the P-frame in the code stream and the simultaneous mutation of the number of intra-prediction macro-block are taken as the basis for detecting the deletion/ interpolation frame operation, and the index for measuring the variation intensity of the mean value of the prediction residual and the number of the intra-prediction macro-blocks is designed, And then the fusion characteristic is constructed based on the two indexes, and the verification method for capturing the code stream information mutation is realized. In the case of no constraint to the video coding parameters, the method of the present invention can effectively detect and position the video frame insertion/ deletion operation.
【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】