Research On The Detection Of Copy-move Digital Image Tampering

Nowadays,with the rapid development of digitalization,we can find the digital image everywhere,no matter from our phone or all kinds of network media.At the same time,digital image processing technology is gradually mature and popular,a variety of image processing and editing software in our field of vision,which means that the ordinary users,can edit,modify and store the operation of digital photos.Not all of the image tampering is simply to enhance the visual effect of the image,for the purpose of some people,they falsification and tampering of digital image,causing adverse effects to other people and the society.Therefore,it is of great significance to carry out the research on image tamper detection technology.Image copy-move operation is a common operation of image tampering.The corresponding copy-move passive blind detection is also an important part of image tampering detection.In this thesis,image copy-move tamper detection technology is analyzed and studied.The blind detection method of image copy-move is mainly composed of two kinds,one is block matching based and another is the feature point matching based.The block matching method based on sliding window is the first image block,feature extraction for each block,by searching for similar image blocks to locate the tampered area;Based on the feature point matching method is mainly through extracting the key points of a certain number of high degree of identification,and at these points around the local feature extraction to construct feature vector.Compared with the method based on image block,the number of key points extracted by this method is much less than the number of overlapped image blocks,so it has better real-time performance.However,the number of feature points extracted from the single texture region is relatively small,so the detection effect is not ideal for the single region of the texture,such as the sky,the earth,the lake and so on.In this thesis,these two kinds of detection methods are studied respectively.The image copy-move tampering models,the common tampering methods of copy-move,and the existing authentication algorithms for copy paste and tamper with the blind forensics technology are analyzed in detail.(1)HU invariant moment algorithm based on proportion invariance is proposed.First,through theoretical analysis,it is found that in the discrete case,the Hu invariant moment does not have a proportional invariance.Therefore,we improved the Hu moment invariants and constructed 6 new invariant moments.Theoretical deduction and experiments proved that the modified Hu moment invariants satisfy proportional invariance in discrete cases.(2)For the replication of single texture region,we use the image block matching algorithm after analyzed the advantages and disadvantages of the existing image block algorithm.Based on the improved Hu moments and Lab,a blind detection algorithm of image copy-move tamper is proposed.Copy-move the texture of single tampered regions has no meaning in itself.But by copying the textured area paste to a target image to hide the target,the tampering means widely applied in real life,it is very meaningful to detect such tampering.Combining the improved Hu invariant moments and the three channel features of Lab color space,the image blocks are jointly represented,and image block feature vector correlation is used to identify the image tampering and locate its location.The experimental results show that the algorithm of replica location textured area accurately,and can also accurately locate the tampered region after scaling,rotation and other geometric transformations,for brightness change,blur,noise,JPEG compression postprocessing operation has good robustness.(3)For non-textured area replication,in order to further improve the efficiency of image copy paste tampering blind detection,detection algorithm for feature point matching based on the analysis of the traditional SIFT algorithm,proposed an improved SIFT algorithm,the clustering algorithm of SIFT feature vector column vector matrix using K-means algorithm,by selecting and the overall difference of the smallest column vector as the initial cluster centers,finally extract the feature vector of the initial clustering center of the same class,greatly reduces the dimension of feature vector.This algorithm first extracted with improved image key point SIFT algorithm,generating feature descriptors,finally matching feature descriptor,the algorithm uses the Euclidean distance of feature point matching,RANSAC algorithm is used to eliminate the false matches,to achieve recognition and tamper location test results showed that: compared with the traditional SIFT algorithm and PCA-SIFT algorithm in this thesis,the algorithm has higher detection efficiency,robustness and high for translation,replication region and rotating operation zoom scale.(4)Compared the results of the two algorithms proposed in this thesis,we draw the following conclusions: when you need to detect single texture region,the improved Hu moment and the Lab algorithm has more property.And when detecting non-single texture region,the SIFT and K-means algorithm has a higher detection efficiency.