Research On Image Splicing Detection By Statistical Feature Analysis

With the development of electronic information technology, the popularity of image acquisition devices and with the help of Internet/Mobile Internet, digital images have gained great popularity in people’s daily lives. At the same time, editing and modifying images is getting easier and easier with the development of image processing software. Digital images are easy to use and to spread, hence they bring great conveniences in people’s lives. However, every coin has its two sides, digital images also introduce many problems in network-social-management and judicial forensics.Compared with the traditional ?lm photos, digital images are much easier to be tampered, which would bring great challenges in audio-visual materials authentication in judicial forensics, news material veri?cation and fake image information authentication in network public opinion analysis. Due to this situation, digital image tamper detection has gained extensive attentions in recent years, and it becomes research hotspot in image forensics.Digital image tampering detection can be divided into two categories, i.e. active detection and passive detection. Active detection technologies authenticate digital images by digital watermarking or ?ngerprint which is inserted into the image in the imaging process. Passive detection technologies detect image forgeries without the help of any prior knowledge, instead they could authenticate an unknown image by verifying the consistency of imaging principle(e.g. CFA), physical properties(e.g.camera pattern noise, light direction, etc.) or image statistical features. In this thesis,based on the research of mainstream passive image tampering detection technologies,we propose a series of key technologies to detect image splicing which is the most fundamental and popular way of image tamper. The details and novelties are listed as follows:1. When making a composite image, a forger tries to cover the splicing artifacts to deceive human eyes. However, the splicing operation would inevitably alter the underlying statistical distributions, which could be treated as an evidence to authenticate the image. In chapter 3, third-order statistical features in block DCT(BDCT)domain and Run-length run-number(RLRN) features in chroma spaces are proposed as higher-order statistical features for image splicing detection. We expand the traditional ?rst-order Markov features, and extract the third-order statistical features on the BDCT coe?cients as the discriminative features for image splicing detection. We discussed the feasibility and effectiveness of the third-order statistical features for image splicing detection comprehensively. Chroma information is introduced in our research work to reduce the interference of image content, and RLRN features in the chroma channels are proposed to authenticate digital images. Experimental results show that the higher-order statistical features outperform the traditional low-order statistical features.2. The traditional statistical features based methods treat the source image as a1-D causal signal, from which features are extracted for splicing detection. However,image is 2-D signal in nature, and each node is correlated with its surrounding neighbors, which is noncausal. Hence, much information would be lost if simply treat an image as a 1-D causal signal. In chapter 4, a 2-D noncausal model is proposed for image splicing detection, that is, the input image is treated as a 2-D noncausal signal which is depicted by the prior probabilities of states, the probability density function of observations of each state and the state transition probabilities. We give the inference of the noncausal model, and the algorithm to compute the model parameter sets.The parameter sets of the noncausal model are treated as discriminative features which are integrated with machine learning method to detect image splicing. Experimental results over public image splicing evaluation dataset show that the proposed method could detect the splicing operations effectively, and the detection performance outperforms most state-of-the-art methods.3. Since image splicing traces vary with the change of color channels, for a certain feature extraction method, it could hardly gain its best performance in a ?xed color channel. In chapter 5, we propose a chroma-like channel design method for passive image splicing detection, which is integrated with support vector machine(SVM) and generalized discriminant analysis(GDA) learning algorithm to ?nd the optimal color channel in which the feature extraction method could gain its best detection performance. Six different kinds of feature extraction methods are employed in our experimental work to verify the feasibility and effectiveness of the proposed method, and the experimental work shows that the detection accuracies of all these six feature extraction methods in their optimal chroma-like channels are higher than that in commonly used channels.4. In recent years, more and more features are involved in the image splicing detection to boost detection accuracy. Although combined features with high dimensionality could capture more splicing artifacts, it inevitably increases the computational complexity of the classi?er and probably introduces excessive redundant information which would interfere with the classi?er. A distributed local Margin learning based method is proposed in chapter 6 for optimization of high-dimensional features in image splicing detection. The proposed method, ?rst, tries to optimize a non-negative matrix to make the local margin in the transformed feature space maximum, and then the non-negative matrix is employed to determine the importance of the features. Finally, we can get a new feature set by removing the unimportant(redundant) features from the original high-dimensional feature set. In order to improve the performance of the proposed feature selection method, we introduce parallel computing in the local margin learning algorithm. Experimental results show that the proposed distributed local margin learning method could effectively reduce the dimensionality of the highdimensional features, and it could also reduce the time cost of the classi?er in the parameter learning, training and prediction phase at the cost of a small ?uctuation of detection accuracy.This research work systemically studies the image splicing detection from three aspects: feature extraction, color channel selection and high-dimensional feature processing. First, the third-order statistical feature extraction method in BDCT domain and the RLRN feature extraction method in chroma spaces are introduced for image splicing detection. Then the detection methods are expanded to two dimensionality,and a 2-D noncausal Markov model is proposed to capture more splicing artifacts and to boost the detection accuracy. After that, feature extraction method is combined with machine learning scheme to ?nd its optimal color channel, which would further improve the detection accuracy. Finally, a high-dimensional feature processing scheme is proposed to improve the detection performance of the high-dimensional features in image splicing detection.