基于数字图像统计特征的盲取证技术研究

发布时间：2018-02-03 14:18

本文关键词： 数字图像盲取证离散余弦变换系数 Benford定律无序度　出处：《南京邮电大学》2016年硕士论文　论文类型：学位论文

【摘要】：人们可以通过听觉、视觉、味觉等多种途径获得外界信息,其中来自于视觉的信息大约占80%。数字图像作为视觉信息最为常见的载体,它传递出来的信息比通过其它途径获取的信息更为准确和真实。然而随着数码相机、个人电脑的普及,人们看到的图片可能是经过图像编辑软件恶意伪造和篡改后的图片,因此人们获取可靠和真实信息的重要途径受到严重威胁。数字图像盲取证是一种在待测图像未嵌入数字水印、数字签名、数字指纹等先验知识的情形下对其真实性以及来源进行鉴别的技术。本文研究了数字图像统计特征对数字图像盲取证的影响以及提出了一种新的自然图像和计算机生成图像的盲取证算法。首先,通过Benford定律研究了离散余弦变换(Discrete Cosine Transform,DCT)系数的选择对自然图像和计算机生成图像鉴别的影响,仿真结果表明:在使用相同判据和最佳阈值的情况下,自然图像和计算机生成图像的分类准确率随着选择的图像DCT高频系数的增多而先提高,然后达到最大值,之后随着DCT低频系数逐渐增多,分类准确率下降,因此体现图像细节的DCT高频系数更有益于图像分类准确率的提高;在使用最佳判据和最佳阈值的情况下,利用灰度图对自然图像和计算机生成图像进行分类,其准确率最高,图像的绿色(green,G)通道次之,红色(red,R)通道和蓝色(blue,B)通道都较差。然后,研究了DCT系数的选择对拼接图像鉴别的影响,仿真结果表明:从最高频往低频选择DCT系数,随着低频分量的增加,两类图像的分类准确率变化不大;从最低频往高频选择DCT系数,随着能够捕捉拼接图像锐利边缘的高频细节分量增加,两类图像的分类准确率有较大的改善;当选择相同数量的DCT系数时,从最高频向低频选择DCT系数方案的分类准确率普遍优于从最低频向高频选择DCT系数方案的分类准确率。最后,提出了一种基于图像分层模型无序度的自然图像和计算机生成图像的盲取证算法。通过提取数字图像的每个二进制图层的无序度特征,用这些特征训练支持向量机并对待测数字图像的类别进行预测,仿真结果表明:通过将二进制图层划分为合适长度的子序列计算图层的无序度,上述两类图像的分类准确率可达82.80%,受试者工作特征曲线(receiver operating characteristic curve,ROC)下面积(area under the ROC curve,AUC)值为0.8982,分类器性能属于良好,逼近优秀,对自然图像和计算机生成图像具有较好的分类效果。
[Abstract]:People can obtain external information through auditory, visual, taste and other channels, among which the information from vision accounts for about 80%. Digital image as the most common carrier of visual information. It conveys information that is more accurate and truthful than it is otherwise. However, with the advent of digital cameras, personal computers have become more popular. People may see images that have been maliciously forged and tampered with by image editing software. Therefore, the important way to obtain reliable and true information is seriously threatened. Blind digital image forensics is a kind of digital watermarking and digital signature that is not embedded in the image under test. In this paper, the influence of statistical features of digital image on blind forensics of digital image is studied, and a new natural image and calculation is proposed. Computer generated image blind forensics algorithm. First. Discrete Cosine Transform is studied by means of Benford's law. The influence of the selection of DCT coefficients on the identification of natural images and computer-generated images is discussed. The simulation results show that the same criteria and optimal thresholds are used. The classification accuracy of natural image and computer-generated image increases with the increase of the selected image DCT high frequency coefficient, then reaches the maximum, and then increases with the DCT low frequency coefficient. The classification accuracy is decreased, so the DCT high frequency coefficient, which reflects the image details, is more beneficial to the improvement of the image classification accuracy. In the case of using the best criterion and the best threshold, the grayscale image is used to classify the natural image and the computer-generated image, and the accuracy is the highest, followed by the green green GG channel of the image. The red squid R) channel and the blue blue blue B) channel are poor. Then, the influence of the selection of DCT coefficients on the stitching image identification is studied. The simulation results show that the classification accuracy of the two kinds of images varies little with the increase of the low-frequency components when the DCT coefficients are selected from the highest frequency to the low frequency. Selecting the DCT coefficient from the lowest frequency to the high frequency, the classification accuracy of the two kinds of images is improved greatly with the increase of the high frequency detail component which can capture the sharp edge of the stitched image. When the same number of DCT coefficients are selected, the classification accuracy of selecting DCT coefficients from the most high frequency to the low frequency is generally better than that of the DCT coefficients from the lowest frequency to the high frequency. Finally. This paper presents a blind forensics algorithm for natural images and computer-generated images based on the disorder degree of image stratification model. The feature of disorder degree of each binary layer of digital image is extracted by extracting the disordered degree features of each binary layer of digital image. These features are used to train support vector machines and predict the types of digital images. The simulation results show that the disorder degree of the layer is calculated by dividing the binary layer into sub-sequences of appropriate length. The classification accuracy of the two kinds of images can reach 82.80%. Receiver operating characteristic curve. The area area under the ROC curveAUC value is 0.8982, the performance of classifier is good and the approximation is excellent. It has good classification effect on natural image and computer generated image.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;D918.2

【相似文献】