Image Splicing Detection Algorithm Based On Merging Steganalysis Feature In DCT Domain With Run-Length Feature

With the development of information technology, people's lives are surrounded by digital images which are a sort of important visual message carrier. Meanwhile, all kinds of image edition software get rapid development, making the modification of images easier and easier for some purposes. However, the authenticity of some images is very important such as court evidence, news report, etc. Forgers are very likely to do some post processing to cover the traces of modification, which makes it difficult to tell visually. Therefore it's important to distinguish the authentic images from the forged ones.Researches show that steganalysis features can also apply into image splicing detection. A steganalysis method is used in image splicing detection in this paper. The steganalysis feature consists of many submodels which describe different relationships of neighboring pixels through residual and co-occurrence matrix computing. The experiment results show that the spatial steganalysis feature can't balance computational efficiency and accuracy; considering DCT has the advantage of decorrelation and energy compaction, a method applying the steganalysis features into frequency domain was proposed, and the suitability of it is analyzed. Experiment results show that the frequency feature is far superior to spatial feature. Furthermore, the experiment of merging some submodels in spatial and DCT domain was made. The result shows that merged features in either domain or in both domains make very slow promotion on accuracy.To further improve the recognition rate, a method that combines a sort of run length feature and second order submodel spam12 hv with higher accuracy is proposed. The run length feature is extracted along directions of 0o, 45 o, 90 o and 135 o in the decorrelated pixel difference 2-D array. Both the steganalysis and run length features are extracted from chroma space. Support vector machine is chosen as the classifier. The highest accuracy over dataset CASIA v1.0 can achieve 98.20% and over the more complicated dataset CASIA v2.0 can achieve 97.37%. Experiment results show that the proposed method has significant advantages compared with some other existing algorithms.