基于同幅图像的复制粘贴篡改的盲检测

发布时间：2018-02-13 20:21

本文关键词： 区域复制粘贴篡改块匹配 Hu矩 Zernike矩特征点匹配 SIFT算法高斯几何不变矩　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着数码相机、高性能智能手机的普及和高性能摄像头的不断革新,数字图像作为日常生活中的信息载体越来越得到普遍应用。随着数字图像处理编辑软件也大量出现,虽然对图像处理带来了很大的益处,但任何事情都是"双刃剑",这也给那些心怀叵测之人带来了危害社会的更加容易的手段。目前,网络上和社会上充满着大量伪造图片在混淆视听,已经对人们的切身利益造成了伤害。是以,针对伪造图像的识别与研究具备重大价值,也应该提到探索日程上来。当前图像篡改检测算法琳琅满目,我国关于这方面的研究虽然取得了非常大的成就,但目前还处于初步阶段,也存在许多不足之处,算法性能比较单一。本文中主要针对数字图像区域复制粘贴篡改手段盲检测研究,区域复制粘贴篡改属于局部篡改手段类别。目前,针对图像区域复制粘贴篡改盲检测技术主要分为基于图像块检测算法和基于特征点检测算法,本文对前人算法不足进行分析,并提出相应的改进算法。论文主要工作:1、论文首先具体讲述了关于数字图像篡改认证技术研究背景与意义、国内外研究现状,系统的分析了图像篡改的方式,着重介绍了同幅图像copy-move篡改手段,详细分析了图像copy-move篡改模型以及概述现存对copy-move篡改盲取证技术认证算法,详细介绍了现存较为经典检测方法。2、本文第三章针对图像盲检测技术基于图像块检测算法,分析现存图像块算法的优缺点基础上,提出了基于改进Hu矩和Zernike矩结合的图像块匹配法。该算法基于改进Hu矩和Zernke矩来表征特征向量,该算法首先对图像进行滑动分块,提取图像块特征向量,利用图像子块特征向量相关性来识别图像篡改且定位其位置,该算法实时性得到提高,且对平移、旋转鲁棒性较好。实验结果表明该算法可以有效抵制对篡改区域进行旋转、平移操作处理。3、本文第四章针对图像盲检测算法基于特征点检测算法,分析传统SJIFT算法缺陷,提出了融合高斯几何不变矩的改进SIFT的特征点图像篡改取证算法。该算法中先用改进SIFT算法提取图像关键点,为特征点分配主方向,然后提取关键点邻域窗口的高斯几何不变矩作为关键点的特征描述子,最后进行特征描述子的匹配,该算法采用欧氏距离进行特征点匹配,并利用自适应欧氏距离阈值与RANSAC结合算法剔除误匹配对,实现篡改区域的识别与定位。实验结果表明该算法可以基本保持图像提取的特征点数,甚至会减少提取的特征点数,但可以增加特征点匹配点数,减少误匹配点,并且可以提高特征点提取时间,因为特征描述子维数的减少,对于匹配效率也有一定的提高,实验效果图可以看出,该算法对于复制区域的平移、尺度缩放与旋转操作都具有非常好的鲁棒性,检测精度也较高。
[Abstract]:With the popularization of digital camera, high-performance smart phone and the innovation of high performance camera, digital image is becoming more and more popular as information carrier in daily life. While there are great benefits to image processing, everything is a "double-edged sword," which also gives those who have evil intentions an easier means of harming society. There are a lot of fake images on the Internet and in society, which are confusing and harmful to people's vital interests. Therefore, it is of great value to identify and study fake images. We should also mention the agenda of exploration. At present, there are a great variety of algorithms for image tampering detection. Although great achievements have been made in this area in our country, it is still in its preliminary stage, and there are also many deficiencies. The performance of the algorithm is relatively simple. This paper mainly focuses on the blind detection of digital image region copy-paste tamper, which belongs to the category of local tamper. The blind detection technology of image region copy and paste tamper is mainly divided into image block detection algorithm and feature point detection algorithm. In this paper, the shortcomings of previous algorithms are analyzed. The thesis mainly focuses on the research background and significance of digital image tampering authentication technology, the current research situation at home and abroad, and the systematic analysis of image tampering methods. This paper mainly introduces the copy-move tampering method of the same image, analyzes the image copy-move tamper model in detail, and summarizes the existing authentication algorithms for copy-move tampering blind forensics technology. This paper introduces the existing classical detection method. 2. The third chapter analyzes the advantages and disadvantages of the existing image block detection algorithm based on the image block detection algorithm. An image block matching method based on improved Hu moments and Zernike moments is proposed, which is based on improved Hu moments and Zernke moments to represent feature vectors. By using image subblock feature vector correlation to identify image tampering and locate its position, the real-time performance of the algorithm is improved, and the robustness to translation and rotation is good. Experimental results show that the algorithm can effectively resist the rotation of tampered regions. Translation operation processing. 3. In Chapter 4th, aiming at blind image detection algorithm based on feature point detection algorithm, the defects of traditional SJIFT algorithm are analyzed. An improved feature point image tampering and forensics algorithm based on Gao Si's geometric invariant moment is proposed in this paper. The improved SIFT algorithm is first used to extract the key points of the image and assign the main direction to the feature points. Then the Gao Si geometric invariant moment of the neighborhood window of the key points is extracted as the feature descriptor of the key points. Finally, the feature descriptors are matched, and the Euclidean distance is used to match the feature points. The adaptive Euclidean distance threshold and the RANSAC algorithm are used to eliminate the mismatch pairs to realize the tamper recognition and localization. The experimental results show that the proposed algorithm can basically keep the feature points extracted from the image and even reduce the extracted feature points. But it can increase the matching points of feature points, reduce the mismatch points, and improve the extraction time of feature points, because the reduction of subdimension of feature description can also improve the matching efficiency. The algorithm is robust to the translation, scaling and rotation operations of the replication region, and the detection accuracy is also high.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】