基于PRNU的自然图像和计算机生成图像来源取证

发布时间：2018-01-10 20:25

本文关键词：基于PRNU的自然图像和计算机生成图像来源取证　出处：《湖南大学》2012年硕士论文　论文类型：学位论文

【摘要】：随着数码设备和图像处理软件技术的快速发展,人们已可以很轻松地获取和修改数字图像。但先进的技术在给人们的生活带来便利的同时,也暴露出了很多安全问题。若不法分子将伪造的数字图像用于新闻、证物和科学研究等正式场合,将会混淆视听,对事件的真实性和社会的稳定性产生严重的影响。因此,确保数字图像真实性和完整性的数字图像取证技术已受到广大研究者的关注。本文主要研究了自然图像和计算机生成图像的来源鉴别。首先,对数字图像取证技术的研究背景、意义及国内外研究现状进行了阐述,对数字图像取证技术的研究内容及研究成果进行了综述。其次,对本文算法所涉及的相关理论知识进行了介绍。最后,针对自然图像和计算机生成图像来源的识别问题,提出了两种鉴别算法： 1.提出了一种基于复合特征的两类图像来源鉴别方法。该方法基于自然图像和计算机生成图像在统计、纹理和噪声特性上的不同,首先提取灰度图像直方图在空域和小波域的均值、方差、峰值、偏度和中位数作为统计特征；然后提取灰度图像及其小波域子带的分形维数作为纹理特征；最后针对基于小波滤波提取的光照响应不一致性噪声(Photo-Response Non-Uniformity Noise, PRNU)的不足,先将图像经过高斯高通滤波预处理,再提取PRNU的统计和纹理特征,作为噪声特征,共48维特征。采用支持向量机(Support Vector Machine, SVM)进行分类,平均鉴别率为94.29%,其中对计算机生成图像鉴别率为97.30%,自然图像鉴别率为91.28%,表明该方法适合两类图像的来源鉴别,而且鉴别效果比已有方法在性能上有所改善。 2.提出了一种基于PRNU与彩色滤镜阵列(Color Filter Array, CFA)插值特性的两类图像来源鉴别方法。该方法利用CFA插值是自然图像的特有操作和PRNU作为相机的“数字指纹”的特性,首先分析了CFA插值对PRNU的影响在两类图像中的差异,然后利用PRNU邻域方差直方图来表达此不同,并分别从RGB三颜色通道中提取PRNU邻域方差累加和及其方差直方图的最大值、加权平均和方差,共12维特征,最后采用SVM进行分类,平均鉴别率达到96.55%,为自然图像和计算机生成图像的鉴别提供一种新的有效方法。本文提出的两个来源鉴别算法,能将自然图像和计算机生成图像进行有效地准确地分类。
[Abstract]:With the rapid development of digital equipment and image processing software technology, people can easily obtain and modify digital images. It also exposes a lot of security problems. If criminals use fake digital images for official occasions such as news, evidence and scientific research, it will confuse the public. It has a serious impact on the authenticity of events and the stability of society. Therefore, digital image forensics, which ensures the authenticity and integrity of digital images, has attracted the attention of many researchers. This paper mainly studies the source identification of natural images and computer-generated images. Firstly, the research background, significance and research status of digital image forensics technology are described. The research content and research results of digital image forensics technology are summarized. Secondly, the related theoretical knowledge of this algorithm is introduced. Finally. In order to identify the source of natural image and computer generated image, two identification algorithms are proposed. 1. Two kinds of image source identification methods based on compound features are proposed, which are based on the differences of statistical, texture and noise characteristics between natural images and computer-generated images. Firstly, the mean, variance, peak value, deviation and median of gray image histogram in spatial domain and wavelet domain are extracted as statistical features. Then the fractal dimension of gray image and its sub-band in wavelet domain is extracted as texture feature. Finally, the deficiency of Photo-Response Non-Uniformity Noise (PRNU) based on the inconsistent noise of illumination response extracted by wavelet filter is discussed. Firstly, the image is pre-processed by Gao Si high-pass filter, and then the statistical and texture features of PRNU are extracted as noise features. The support vector machine support Vector Machine (SVMs) was used to classify the 48 dimensional features. The average discriminant rate was 94.29%. The computer generated image identification rate is 97.30 and the natural image identification rate is 91.28, which indicates that this method is suitable for the source identification of two kinds of images. Moreover, the discrimination effect is better than that of the existing methods. 2. A color Filter Array based on PRNU and color filter array is proposed. This method uses CFA interpolation as the special operation of natural image and PRNU as the "digital fingerprint" of camera. The difference of the influence of CFA interpolation on PRNU in the two kinds of images is analyzed firstly, and then the difference is expressed by using the PRNU neighborhood variance histogram. The maximum, weighted average and variance of the PRNU neighborhood variance cumulative sum and its variance histogram were extracted from the RGB three-color channel, respectively, and the 12 dimensional features were obtained. Finally, SVM was used to classify. The average discriminant rate is 96.55, which provides a new and effective method for the identification of natural images and computer-generated images. The two source identification algorithms proposed in this paper can classify natural images and computer generated images effectively and accurately.
【学位授予单位】：湖南大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41;D918.2

【参考文献】