Research For Processing History Estimation Of Digital Image Forensics

With the proliferation of powerful yet easy-to-use image editing software,ordinary people can easily achieve image editing and tampering with desirable visual effects.However,they are also often used for malicious image forgeries,which might mislead public opinions and even cause serious social damages.Thus,image forensics,which can effectively detect image authenticity and integrity,has received wide research attentions.Unluckily,existing image forensics works can reveal only single image manipulation,whereas practical image forgeries are more likely to be made up of at least two or more image manipulations,which is referred to be image operation chain.Apparently,existing image forensics works can not meet the requirement of the detection of image operation chain.Image processing history estimation refers to the type identification of image manipulations involved in image operation chain,and the revealing of the orders of image manipulations.It is a deep forensics task,and is the inevitable requirement of image forensics for practical use as well.In this thesis,we research on the forensics works for two common image operation chains.Specifically,the main works and contributions are summarized as follows.First,for the image operation chain which is made up of image sharping,image scaling and median filtering,a forensics approach is proposed by exploiting spatial-domain residual features and DCT-domain enhanced Markov features.The operation chain is made up of three typical and distinct image manipulations.As we know,different image manipulations may leave distinct tampering traces.Similarly,image operation chains,which is made up of different image manipulations,will leave characteristics artifacts as well.The artifacts are actually formed by the ambiguous mixture of the traces left by various image manipulations.Inspired by the successful application of spatial rich model(SRM)in the field of image steganalysis,both spatial-domain residual features and DCT-domain enhanced Markov features are extracted,which are then input into the ensemble classifier for the classification of multiple image operation chains.The experimental results show that the proposed approach can achieves desirable detection accuracies,which can effectively discriminate multiple operation chains made up of image sharpening,scaling and median filtering in different combination orders.Moreover,the proposed approach achieves good robustness against JPEG compression.Second,for the image operation chain that is made up of seam carving and histogram equalization,a blind forensics approach is proposed by exploiting the co-occurrence matrix features based on both local binary pattern(LBP)and local phase quantization(LPQ).By analyzing image texture,we observe that this image operation chain will change local image texture,and thus distinct tampering traces or fingerprints are left.By exploiting both LBP and LPQ operators,spatial-domain and frequency-domain local texture information are obtained,respectively.Then,the distribution probability of local textures is modeled by grayscale co-occurrence matrix,which is exploited as the features for blind forensics.Finally,support vector machine(SVM)is used as the classifier to identify image operation chains.The experimental results prove that the proposed approach can effectively identify the types and orders of image operation chain which is made up of seam carving and histogram equalization,and desirable detection accuracies are achieved.