Exposing Photomontage Using High Order Statistics

Digital Images have been widely used in our life as carriers of information. Image forgeries created using publicly accessible image processing tools emerge in our life with a large quantity. People are in need of tools to identify whether an image is tampered or not. Scholars have put forward kinds of methods to exposing digital forgeries without any prior information. This makes it a difficult task and there is only a little achievement with non-ideal performance. So it is of much practical value of our researches.The fact that authentic natural images and their tampered versions have different intrinsic characteristics is usually used to identify them. Methods based on physical and geometrical characteristics need to obtain information of the imaging equipment and image-forming condition, which limit their practicability. We focus on the more general methods based on statistical characteristics. In this respect, however, most methods proposed by far are mainly based on low-order statistics and perform badly. So we plan to research the technology based on high-order statistics to identify natural images, also taking into account low-order statistics.In this paper, we propose a detection method that combines high-order and low-order statistics with image pre-processing before feature extraction. In this method, we first pre-process the images and extract image features of trained image dataset to form a 2-class classified model. For each image to be identified, we do the same pre-processing and feature extraction, which is the input of classifier. We focus on the selection of best combination of features and pre-processing method. In this paper, the characteristic vector of image is composed of the magnitude and phase of bispectrum and edge density, and the selected pre-processing method is image non-Gaussian de-noising. Results indicate that our method is more efficient than traditional algorithms.Besides, we pay attention to the detection of a special type of digital forgeries–copy-move and rely on some premises. We introduce a new detection method based on SIFT(Scale Invariant Feature Transform),which was used for image registration. The proposed method shows a good performance of detecting and locating copy-move image blocks, independent of post processing such as rotation, scaling and intensity adjustment.