The Research Of Digital Media Passive Forensic Based On Object Edge Analysis

With the rapid development of image/video editing tools, even ordinary users canproduce forge or tamper digital media in an easy way and spread them over internetmaliciously. To overcome the strict crisis of public trustworthiness, it is in urgent need todevelop effective detection techniques to verify the originality and integrity of digital video,and find out the possible tampering operations. Image and video passive forensics makesjudgments about whether digital media has been modified or inpainted by itself. Since passiveforensics does not need any prior information such as digital watermark, it has gained wideresearch attentions.For digital images or videos, object-based manipulations such as adding, deleting orchanging video object are usually malicious forgery operations because they directlyinfluence the viewer’s understanding towards the content of digital media. The copy-pastetampering is easily to be realized in an image or different images. For digital video, theresearch on passive forensics for object-based tampering is more meaningful than that fordouble MPEG compression, or frame-based tampering. In this thesis, we start fromresearching on the statistical features near the object edge, and then apply them to digitalimage and video forensics. The main works are summarized as follows：Firstly, for object-based forgery operations, a passive forensics scheme is proposedbased on the statistical characteristics of object contour. It extracts suspicious video objects bybackground subtraction, and analyzes adjustable width object boundary(AWOB) of videoobjects using digital morphological technology, then digs the trace of forgery in small-scaleby analyzing statistical features such as moment features of detailed wavelet coefficients andgradient strength of each color channel etc., of which feature vectors are obtained andcombined as the input of support vector machine (SVM), thus natural objects and forged oneswill be successfully classified. Experimental results show that the proposed approach willachieve a high accuracy of correct detection of above90%.Secondly, a localization of abnormal object contour is proposed based oncontourlet transform and ringing artifacts. The object-based operations usually leadsubtle clues in local areas near object, including blocking artifacts and ringingartifacts. For the local region near image/video object, Non-subsampled ContourletTransform is utilized to analyze the edge and its adjustable width region. It initially judges abnormal edge pixels, and then combines with ringing artifact of single pixel band to furtherlocate abnormal edge points.At present, most forensics schemes are utilizing the statistical features for theclassification of natural and tampered images and videos, while not be able to directlylocate the abnormal object block. The works in this thesis tries to do passive forensicsfrom the local object point of view, and preliminary results are obtained.