The Research Of Unsupervised Learning Algorithm Of Digital Splicing Image Forensics

Digital image forensics can be classified into two different categories, thedetection of copy-move forgery and image splicing, according to the different oftechnology using in the forgery process. We pay our attention to the detection ofsplicing in this paper. Since we can’t find more accurate information reflecting thefeatures of splicing, the existing algorithms at present have kinds of faults.While，many algorithms adopt supervised learning methods in image forensics,which need a large number of training samples, as a result, the stability and efficiencyof the algorithms relies heavily on the stability of some image characteristics. Since,the establishment of standard splicing image database is necessary. However, thiskind of database is very few, what’s more, it is not perfect, and it has been improved,this limits the application of supervised learning methods in image forensics.From the view of the formation process of digital image, we propose a morepractical framework of splicing forensic system based on unsupervised learningmethods, whose efficiency isn’t related to the training samples. In this paper, thetempered image is regarded as a collection of different data structures, so the forensicswill be seen as a classification problem. In that case, image authentic identificationdepends only on the image itself, without the need of training samples.Moreover, we give two different kinds of algorithms to our system, in order topromote the use of image forensics. The thesis completes the following works: weanalysis the block effect in JPEG image, as while, we give the selecting principle ofblocking artifact measurement, what’s more, an enhanced blocking artifactmeasurement based on this principle is presented, which is used in image splicingdetection, the outcome of experiment shows our algorithm can describe a moreaccurate temper area. We explore the kurtosis concentration phenomenon in authenticimage, which is quite different from temper one. Thus, we propose a criterion thatdoesn’t depend on standard database, as a result, the algorithm is more stable. Welocate the tampered position by classifying the local noise eigenvalues. K meanalgorithm is a good choice because of it’s simple and efficiency in all clusteringalgorithm, although, K means algorithm is sensitive to the initial cluster centers,. Weuse the concept of global eigenvalues, which can be seen as the initial cluster center, to overcome the shortage of k means algorithm. as a result,our algorithm can achievemore accurate tampered location...