一种基于双JPEG压缩的数字图像篡改的检测方法

【作者】 于雪燕；

【导师】 胡金初；

【作者基本信息】 上海师范大学 ， 计算机应用技术， 2007， 硕士

【摘要】 随着互联网技术的快速发展,高分辨率数码相机以及功能强大的图像编辑软件的出现,图像篡改已经越来越普遍,它在一定程度上丰富了人们的日常生活,杂志封面和商业广告上精美的图片让人目不暇接。但是篡改图像也带来了许多问题。如果将篡改图像用在新闻媒体或法律上,对社会将会造成很大的影响。因此,对数字图像的真伪鉴别非常重要。而JPEG是目前主流的图像压缩标准,大多数图片是用JPEG格式存储的。在此背景下,本文提出了一种基于双JPEG压缩的数字图像篡改的检测方法。本文首先介绍了JPEG压缩标准,JPEG编解码原理及其实现过程。同时介绍了双JPEG压缩原理以及几个相关概念。着重分析了单量化、双量化、多次量化后DCT系数直方图的特性,并阐述了双量化后DCT系数直方图周期性产生的原因。其次,文章提出了一种基于双JPEG压缩的数字图像篡改的检测方法。首先描述了篡改图像、DCT块、篡改块、非篡改块等几个相关的概念,分析了JPEG图像的双量化效果,指出了被篡改的JPEG图像中篡改块与非篡改块可分的依据。并且给出了一种简单有效的估计直方图周期的算法。其次,根据贝叶斯决策理论设计了针对图像中篡改块与非篡改块的两类别分类器,用于检测篡改块。再次,建立了一个包含110幅图像的图像库,对其中一部分图像做篡改之后,用不同的压缩质量因子进行JPEG压缩,将其中60幅输入SVM用作训练数据,另外50幅用作测试数据。文中分析了两次JPEG压缩的质量因子与错误分类个数的对应关系。最后,给定一幅待检测图像,抽取一个四维特征向量输入到训练好的SVM中检测图像真伪。通过大量实验对该算法进行测试并对实验结果进行分析,总结了不同的压缩质量因子以及篡改图像的尺寸与该方法有效性的关系。实验证明,本文提出的基于双JPEG压缩的数字图像篡改的检测方法是非常有效的。更多还原

【Abstract】 With the rapid development of internet, the emergence of high resolution digital cameras, and powerful image editing software, doctored images can be found anywhere. It can greatly enrich the users lives. People could see many beautiful images on magazine covers and commercial advertisement. But doctored images may also cause some problems. It will cause great social problems if they are used on news report or court. It is very important to detect the doctored images. JPEG is the most frequently used image format, and most of the images are saved as JPEG format. This paper focus on the method of detecting doctored images based on double JPEG compression effect.At the beginning, this paper introduces the JPEG still picture compression standard, JPEG coding and decoding method, as well as double JPEG compression, and analyses the effect of DCT coefficient histogram of single quantization, double quantization and multiple quantization, and explains why the double quantization of a signal introduces periodic artifacts.Secondly, the paper proposes the method of detecting doctored double JPEG images. Several terms are described, The paper analyses the double quantization effect from double JPEG compression, and points out that it is divisible of doctored blocks and undoctored blocks. A simple and effective method used to estimate the period of histogram has been implemented. The paper designs a classifier of two classes using Bayesian decision theory to segment the doctored blocks and the undoctored blocks. A library containing 110 images has been built for training and testing the SVM. The images are doctored first, and then compressed by different compression factors.The paper analyses the relationship between the compression factors(q1 and q2) and error numbers. Finally, given a image, a four-dimensional feature vector has been extracted, and then feed it into the trained SVM to decide whether the image is doctored. Lots of images are used to test this algorithm. The paper comes to the conclusion that double compression factors and images sizes are closely associated with the effectiveness of this algorithm. The algorithm is effective with the experiments.更多还原