The Research Of Passive Forensics For Semantic Object Level Image

With the popularity of powerful image editing tools,even people without any knowledge about image processing can easily fake an image with leaving any visually annoying artifacts.There are ofen tampered images spreading over televisions,magazines and Internet.This breaks our traditional concept of “seeing is believing” and also brings serious cises to public confidence.The verification of image integrity and authenticity is a hot topic in the community of image content security.Active forensics techniques need to embed digital watermark into host image or generate digital signature in advance.However,blind forensics techniques directly exploit the inherent nature of digital image to judge its authticity,which makes it have much better adaptability.Semantic object-based image manipulation is a characteric applicationof image processing.It can not only improve the visual quality after image compression,but also provide better interactive performance.However,object-based image operations are often adopted to change the semantic contents of digital images for malicious purposes.Some common object-based image operations include copy-move,seam carving and exemplar-based image inpainting.Compared with copy-move,seam carving and exemplar-based image inpainting support content-aware mechnaisms,which makes the traces left by image forgeries are much weaker.Their forensics confront more technical challenges.In this thesis,we research on the blind forgery detection against seam carving and exemplar-based image inpainting.Specifically,the main works and contributions are summarized as follows:Firstly,a blind detection approach is proposed for image seam carving by using weber local descriptor(WLD)and local binary pattern(LBP).By analyzing the process of seam carving,we know that it mainly leads to the changes of local texture.Both WLD and LBP have desirable capabilities of local texture description.LBP considers the sign of pixel difference between centre pixel and its adjacent pixels.Thus,the LBP-based texture feature is an index of discrete pattern,which can not provide intensity information about image texture.Luckily,WLD is composed of differential excitation and orientation,which can make up the disadvantages of LBP.Therefore,candidate images are firstly divided into blocks,and both WLD-based and LBP-based statistical features are extracted from each block.Then,the Kruskal-Wallis method is exploited to select a subset of more discriminative features from those features.Finally,the features after reducing dimension are input into support vector machine(SVM)as classifier to judge whether an image has been suffered from seam carving or not.The experimental results show that compared with the state-of-the-art approaches,the proposed approach achieves desirable detection accuracy.Secondly,a blind detection approach is presented for seam carved image with low scaling ratio(LSR)by exploiting multi-scale spatial and spectral entropies.From preliminary experiments,we know that when few seams are deleted from an original image,there are also obvious changes of local spatial and spectral entropies for those blocks at the right side of those carved seams.That is,local image entropy can effectively capture the position changes of adjacent pixels left by seam carving,and it can also measure the loss of local information.In this paper,the statistical features are extracted to describe the distribution of spatial and spectral entropy from three aspects,which include centralized tendency,dispersion tendency and distribution tendency.Then,the LBP-based improved energy features are combined with them to form the final features for tampering detection.Finally,support vector machine(SVM)is exploited as classifier to distinguish whether a candidate image is suffered from seam carving or not.The experimental results show that the proposed approach can achieve better accuracies,especially for those resized image with low scaling ratios.The proposed approach is also robust to post-processing such as JPEG compression and seam insertion.Thirdly,a local Tchebichef moment(LTM)based universal detection approach is proposed for image resizing/retargeting.There are many image resizing techniques,which include scaling,scale-and-stretch and seam carving.They have own advantages and are suitable for different application scenerios.Therefore,the universal detection of image resizing will be more practical.Motivated by the successful applications of LTM in texture classification,we found that no matter which content-aware image resizing technique is used,it will destroy the correlations among adjacent pixels to some extent.Therefore,a LTM-based universal detection approach is proposed for image resizing/ retargeting.The residual is obtained by image pre-processing.Then,the histogram features of LTM are extracted from the residual.Finally,an error-correcting output code strategy is adopted by the ensemble classifier,which turns the multi-classfication problem into binary classification sub-problems.The experimental results that the proposed approach can achieve acceptable detection accuracy for various content-aware image resizing techniques.Fourth,a robust forgery detection approach is proposed for object removal by exemplar-based image inpainting.Most existing forgery detection approaches for image inpainting utilize similar block pairs between inpainted area and the rest areas,but they are invalidated when those inpainted images are further subjected to some post-processing operations such as JPEG compression,adding Gaussian noise and Gaussian blurring.From some preliminary experiments,we found that the similarities of block pairs are destroyed by these post-processing operations,which disturbs the correlations among adjacent pixels to some exent simultaneously.Inspired by the strong capability of joint probability density matrix(JPDM)in characterizing such correlatioin,a hybrid blind forensics strategy is proposed.Firstly,our earlier method is exploited to detect whether a candidate image is forged or not.Secondly,for those undetected images after the first step,JPDM is computed for each difference array to model the correlations among adjacent DCT coefficients,and the average of these matrixes are computed as feature vectors to further expose tampering traces.Finally,ensemble classifier is adopted to distinguish whether those undetected images are tampered or not.Experimental results show the proposed approach can effectively detect object removal by exemplar-based image inpainting either with or without post-processing.Up to present,passive forensics against semantic object-level image tampering is still in its preliminary exploration stage.It is hoped that the present study maybe promote the development of passive image forensics.