Research On Key Technology Of Image Region Duplication Forgery Detection For Digital Image Forensic Applications

“Seeing is believing” is an old proverb which means that the visible facts cannot be denied and should be accepted,even if these facts are presented in the form of digital images.Digital images are generally accepted as credible news in the journalism and media.But,with the rapid growth of highly sophisticated digital cameras and powerful image editing softwares(such as Adobe Photo Shop,Corel DRAW,Gimp,and etc.),it is now possible for common users to handle and process the digital contents.Although,most of the forged images are made for entertainment purposes,they have also shown up in the courtrooms as critical evidence for conviction or innocence.Consequently,the above proverb is no longer holds.To address this issue,“Digital Image Forgery Detection” area has emerged and developed by providing some trust to digital images.Copy-move forgery is one of the most commonly used methods for digital image forgery,where a region of the image is copied and then pasted into the same image in order to hide or replicate some details.The counterfeiters may perform some post-processing operations(e.g.,JPEG compression,noise addition,geometrical transformations,blurring,and etc.)on the forged regions after copy-move operation in order to camouflage the scene,which make the task of detecting the forgery significantly harder.Therefore,copy-move forgery detection algorithms aim at detecting the same or mostly similar regions in the forged images even under postprocessing attacks,by exploiting the similarity between features in these regions.Several methods have been proposed to detect and locate the tampered regions,while many methods fail.This has led us to the proposal of different passive forgery detection schemes to detect “Copy-move” or “Region Duplication” forgery.In this thesis,we first present a block-based forgery detection algorithm for detecting copy-move areas in a digital image.We specifically study the rotational-invariant features using Polar Complex Exponential Transform(PCET).A forged image is first divided into small overlapping circular blocks,and PCET is employed to each block to extract the invariant features,thus,the PCET kernels represent each block.Second,the Approximate Nearest Neighbors(ANN)Searching Problem is used for identifying and collecting the potential similar blocks by means of locality sensitive hashing(LSH).In order to make the algorithm more robust,morphological operations are applied to remove the wrong similar blocks(false alarms).Experimental results show that our proposed technique is robust to geometric transformations with low computational complexity.As for the second focus of this thesis,we move beyond keypoint-based forgery detection algorithms and propose a framework to detect only the points with high entropy(keypoints)instead of blocks.First,the feature points are detected from the forged image by using the F ¨orstner Operator.Second,the algorithm extracts the features from each keypoint by using Multi-support Region Order-based Gradient Histogram(MROGH)descriptor,and then matching the extracted features.Finally,the affine transformation parameters can be estimated using the RANdom SAmple Consensus(RANSAC)algorithm.Experimental results are presented to confirm that the proposed framework is effective to locate the altered regions with geometric transformation(rotation and scaling).The usability and efficacy of the proposed method is verified by comparing with state-of-the-art methods.Finally,detecting region duplication forgeries from the smooth regions is one of the main crucial problems.Most of the existing forgery detection methods fail to detect the forgery in the smooth(non-textured)regions due to the poor performance of the keypoint detectors in these regions,in addition to its sensitivity to geometric changes in the duplicated regions.To solve these problems and detect keypoints which can cover all the regions in the image,two steps for keypoint detection is proposed.First,the Scale-invariant Feature Operator(SFOP)is used to detect the spatially distributed keypoints from the textured regions.After that,the Harris corners detector is used to detect the keypoints from the missing regions,including the smooth regions.To improve the matching performance,the local feature points are described by using the MROGH descriptor.Finally,the RANSAC algorithm is used to estimate the affine transformations applied.Tests on a common benchmark data set reach the performance levels of 98.32% precision,showing promising potential for real-world scenarios compared with the state-of-the-art methods.