#### DMCA

## Image Tampering Detection For Forensics Applications

Citations: | 1 - 1 self |

### Citations

11966 | Maximum likelihood from incomplete data via the em algorithm
- Dempster, Laird, et al.
- 1977
(Show Context)
Citation Context ...lution, where one set of variables are optimized in the Expectation step (E-step) through the computation of expected values and the other in the Maximization step (M-step) through maximum likelihood =-=[73]-=-. Depending on specific models at hand, one may encounter intractable forms in either the E-step or M-step. Proper approximations are therefore needed. Such inexact EM algorithms are termed Variationa... |

3787 | Normalized Cuts and Image Segmentation
- SHI, MALIK
- 2000
(Show Context)
Citation Context ...utomated process is needed. The image segmentation problem has been studied extensively and there have been numerous tools available. Among these tools, the state-of-the-art algorithm Normalized Cuts =-=[47]-=- will be incorporated as our automatic segmentation component, although other methods such as Mean Shift [48] may also be considered. Both manual and automatic segmentation schemes will be tested with... |

3483 | Conditional random fields: Probabilistic models for segmenting and labeling sequence datasets
- Lafferty, McCallum, et al.
- 2001
(Show Context)
Citation Context ...tructures that link blocks within same area . . . . . . . . . . . . . . 116 4.6 Random Field progression illustrated in 1D (a) traditional HMM (b) MEMM introduced in [68] (c) Conditional Random Field =-=[69]-=-. . . . . 118 4.7 Impact on inference accuracy of fusion and DQ only settings. (a) images 1∼45 (b) images 46∼90 (average accuracy over 90 images: parallel fusion 83.49%, cascade fusion 81.71%, individ... |

3391 | A Tutorial on Support Vector Machines for Pattern Recognition, Data Mining Knowledge Discov.
- Burges
- 1998
(Show Context)
Citation Context ... and within the boundary segment. Each image (or boundary segment) will be represented by a feature vector composed of these CRF fitting measures. Statistical Support Vector Machine (SVM) classifiers =-=[49]-=- are applied on such feature vectors to first learn authentic and spliced classes from a set of training instances and then used to determine whether an incoming test image is authentic or spliced. SV... |

2120 | Fast approximate energy minimization via graph cuts
- Boykov, Veksler, et al.
- 2001
(Show Context)
Citation Context ...orizontal or vertical neighbors (termed 2-cliques), as shown in Fig. 4.5a. As the exact MAP solutions for hidden labels y in Eqn. (4.2) is only obtainable under certain restrictions (e.g., Graph Cuts =-=[60, 61, 62, 63]-=-) and is generally intractable, there have been a significant number of approximate solutions, including the traditional simulated annealing, Mean Field (MF) and Loopy Belief Propagation (LBP) [64, 65... |

1047 | What energy functions can be minimized via graph cuts
- Kolmogorov, Zabih
- 2004
(Show Context)
Citation Context ...orizontal or vertical neighbors (termed 2-cliques), as shown in Fig. 4.5a. As the exact MAP solutions for hidden labels y in Eqn. (4.2) is only obtainable under certain restrictions (e.g., Graph Cuts =-=[60, 61, 62, 63]-=-) and is generally intractable, there have been a significant number of approximate solutions, including the traditional simulated annealing, Mean Field (MF) and Loopy Belief Propagation (LBP) [64, 65... |

1012 |
Individual comparisons by ranking method,”
- Wilcoxon
- 1945
(Show Context)
Citation Context ...icance test tool, its inherent independence and Gaussianity assumptions are not adequate for our case. To properly address the conditions mentioned above, we adopt Wilcoxon’s signed rank test instead =-=[78]-=-. This test relaxes both assumptions: each pair of measurements can be dependent on each other and the distribution of each measurement set does not have to be Gaussian. It is conducted in the two-way... |

859 | Recovering High Dynamic Range radiance maps from photographs. - PE, Malik - 1997 |

676 | Loopy belief propagation for approximate inference: an empirical study”,
- Murphy, Weiss, et al.
- 1999
(Show Context)
Citation Context ...62, 63]) and is generally intractable, there have been a significant number of approximate solutions, including the traditional simulated annealing, Mean Field (MF) and Loopy Belief Propagation (LBP) =-=[64, 65]-=-. The MRF framework has also been widely used in the image processing community, solving problems as texture analysis, image segmentation and object recognition [66, 67]. For our fusion framework, the... |

595 |
Illumination for computer generated pictures. In:
- Phong
- 1975
(Show Context)
Citation Context ...cident lights. One typical light model, the Phong Illumination Model, contains three components,14 L N R V Figure 2.2: Phong illumination model. diffusive, specular and ambient, as shown in Fig. 2.2 =-=[23]-=-: E = kdL T N + ks(R T V) α + kaA (2.1) where kd is the diffusive constant of the surface, ks the specular constant, and ka the ambient constant. They characterize how the surface responds to differen... |

561 | Maximum entropy Markov models for information extraction and segmentation.
- McCallum, Freitag, et al.
- 2000
(Show Context)
Citation Context ...egmentation boundary (c) relaxed structures that link blocks within same area . . . . . . . . . . . . . . 116 4.6 Random Field progression illustrated in 1D (a) traditional HMM (b) MEMM introduced in =-=[68]-=- (c) Conditional Random Field [69]. . . . . 118 4.7 Impact on inference accuracy of fusion and DQ only settings. (a) images 1∼45 (b) images 46∼90 (average accuracy over 90 images: parallel fusion 83.4... |

531 | Image inpainting
- Bertalmio, Sapiro, et al.
- 2000
(Show Context)
Citation Context ...form of copy-andpaste (splicing), edge smoothing/matting after splicing (using either 2D filtering or alpha blending), color adjustment, deletion and duplication in scientific images [10], inpainting =-=[11]-=-.... etc.7 One typical tampering operation that has been studied by many researchers is splicing, for which many solutions have been proposed [1, 8, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21]. For image... |

519 | On discriminative vs. generative classifiers: A comparison of logistic regression and naive Bayes
- Ng, Jordan
- 2002
(Show Context)
Citation Context ...oyed according to the applications in question. Also, they do not entirely conflict each other. In fact, there has proven to be shared links under certain circumstances. Further study can be found in =-=[70]-=-.119 4.2.5 Discriminative Random Field The migration from traditional MRF to CondRF has offered a suitable solution for our fusion task. The CondRF removes the emission probability p(xi|yi). It also ... |

516 | Markov Random Field Modeling in Computer Vision
- Li
- 1995
(Show Context)
Citation Context ...rger portion of this 8x8 block. 4.2.2 Markov Random Field Markov Random Field (MRF) offers well established theories for solving labeling problems and can be viewed as the 2D version of Markov Chains =-=[59]-=-. The most common form of MRF is a generative formulation, characterizing the observations based on hidden class labels. The observation on each node xi is influenced by its hidden label yi (usually t... |

429 |
Exact maximum a posteriori estimation for binary images.
- Greig, Porteous, et al.
- 1989
(Show Context)
Citation Context ...orizontal or vertical neighbors (termed 2-cliques), as shown in Fig. 4.5a. As the exact MAP solutions for hidden labels y in Eqn. (4.2) is only obtainable under certain restrictions (e.g., Graph Cuts =-=[60, 61, 62, 63]-=-) and is generally intractable, there have been a significant number of approximate solutions, including the traditional simulated annealing, Mean Field (MF) and Loopy Belief Propagation (LBP) [64, 65... |

360 | An optimal graph theoretic approach to data clustering: Theory and its application to image segmentation
- ZhenyuWu, Leahy
- 1993
(Show Context)
Citation Context ... output with foreground, background regions and region boundary explicitly indicated.63 bustness against over-segmentation, demonstrating improvement over the standard minimum cut method proposed in =-=[50]-=-. It treats pixels of the image as vertices in a graph and considers dissimilarity measures between pixels. The results are multiple subgraphs that exhibit high similarity within each subgraph and min... |

286 |
Lazy snapping.
- Li, Sun, et al.
- 2004
(Show Context)
Citation Context ...ample images are shown in Fig. 3.10. Another set, the Advanced data set, contains 21 authentic images and 38 highquality spliced images with heavy post processing developed in Microsoft Research Asia =-=[51, 52, 53, 54, 55, 56, 57]-=-. This is a much more realistic and challenging set81 (a) (b) (c) (d) Figure 3.11: Example images from the Advanced data set (a)(b) authentic (c)(d) spliced. since these images are typically JPEG com... |

281 |
Color imaging array
- Bayer
(Show Context)
Citation Context ...image, with different sites recording colors of different wavelengths. There are multiple widely used color filter array configurations, among which the Bayer Pattern is the most popular (Fig. 2.12b) =-=[30]-=-. Note in Bayer Pattern the green channel is sampled twice as much as the red and blue channels. To generate a multispectral image from the single channel CCD output, interpolation from neighboring si... |

250 | Radiometric self calibration.
- MITSUNAGA, NAYAR
- 1999
(Show Context)
Citation Context .... . . . . . . . . . . 39 2.19 CRF estimation based on color space colinearity (a)(b) test images, with edge patches highlighted in green (c)(d) CRF estimation results (blue: first benchmark algorithm =-=[37]-=-, black: second benchmark algorithm [38], red: proposed in [39], green: ground truth from Macbeth Chart) [39]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.20 Equalized greysca... |

210 | Markov random fields with efficient approximations,”
- Boykov, Veksler, et al.
- 1998
(Show Context)
Citation Context |

167 | Radiometric ccd camera calibration and noise estimation - Healey, Kondepudy - 1994 |

163 |
The probable error of a mean.
- Student
- 1908
(Show Context)
Citation Context ...est does not assume any form of distribution, e.g., Gaussian distribution of the measurements, so that it is generally applicable regardless of the behavior of our data. Although the Student’s t-test =-=[77]-=- has been a widely used significance test tool, its inherent independence and Gaussianity assumptions are not adequate for our case. To properly address the conditions mentioned above, we adopt Wilcox... |

154 | Improved fast Gauss transform and efficient kernel density estimation.
- Yang, Duraiswami, et al.
- 2003
(Show Context)
Citation Context ...umerous tools available. Among these tools, the state-of-the-art algorithm Normalized Cuts [47] will be incorporated as our automatic segmentation component, although other methods such as Mean Shift =-=[48]-=- may also be considered. Both manual and automatic segmentation schemes will be tested within the proposed consistency checking algorithm. Their respective merits and drawbacks will also be discussed ... |

149 | Exposing digital forgeries in color filter array interpolated images.
- Popescu, Farid
- 2005
(Show Context)
Citation Context ... suitable camera signature for source identification and forgery detection purposes. 2.2.2.1 Demosaicking Estimation Using EM Algorithm An expectation-maximization (EM) based algorithm is proposed in =-=[16]-=- for demosaicking estimation since the interpolation sources (i.e., at site (i, j), which channel it records for itself and which channels are interpolated from its neighbors) and the interpolation co... |

146 | On advances in statistical modeling of natural images,”
- Srivastava, Lee, et al.
- 2003
(Show Context)
Citation Context ...orth noting that the definition of image authenticity depends on actual application scenarios. In this thesis, we use the terms authentic images and natural5 images interchangeably. For instance, in =-=[5]-=-, natural images refer to photographic images of natural scenes and in [6], natural images are defined as those distinct from range images. In this thesis, an authentic image is defined as an image ca... |

141 | Accelerated training of conditional random fields with stochastic gradient methods.
- Vishwanathan, Schraudolph, et al.
- 2006
(Show Context)
Citation Context ...es. Among these two options, LBP achieves higher inference accuracy and better convergence behavior, therefore we report results based on LBP only. We use the open source CRF2D toolbox implemented by =-=[76]-=-. The learning procedure is outlined as follows: 1. Randomly initialize parameters w and v 2. Based on current parameters ˆw, ˆv, infer labels ˆy123 3. Based on current labels ˆy, update parameters ˆ... |

133 | Image completion with structure propagation.
- Sun, Yuan, et al.
- 2005
(Show Context)
Citation Context ...ample images are shown in Fig. 3.10. Another set, the Advanced data set, contains 21 authentic images and 38 highquality spliced images with heavy post processing developed in Microsoft Research Asia =-=[51, 52, 53, 54, 55, 56, 57]-=-. This is a much more realistic and challenging set81 (a) (b) (c) (d) Figure 3.11: Example images from the Advanced data set (a)(b) authentic (c)(d) spliced. since these images are typically JPEG com... |

130 | Video object cut and paste.
- LI, SUN, et al.
- 2005
(Show Context)
Citation Context ...ample images are shown in Fig. 3.10. Another set, the Advanced data set, contains 21 authentic images and 38 highquality spliced images with heavy post processing developed in Microsoft Research Asia =-=[51, 52, 53, 54, 55, 56, 57]-=-. This is a much more realistic and challenging set81 (a) (b) (c) (d) Figure 3.11: Example images from the Advanced data set (a)(b) authentic (c)(d) spliced. since these images are typically JPEG com... |

116 | The Nonlinear Statistics of High-Contrast patches
- Lee, Pedersen, et al.
- 2003
(Show Context)
Citation Context ...plication scenarios. In this thesis, we use the terms authentic images and natural5 images interchangeably. For instance, in [5], natural images refer to photographic images of natural scenes and in =-=[6]-=-, natural images are defined as those distinct from range images. In this thesis, an authentic image is defined as an image captured by a single camera in a single process. Following this definition, ... |

115 |
Exposing Digital Forgeries by Detecting Inconsistencies in Lighting.
- Johnson, Farid
- 2005
(Show Context)
Citation Context ...2.3 Practicality of Tampering Detection Systems . . . . . . . . . . 151 References 154 iiiList of Figures 1.1 Examples of doctored photographs (a) celebrities Cher and Brad Pitt spliced side-by-side =-=[1]-=- (b) ex-U.S. presidential election candidate John Kerry spliced side-by-side with actress Jane Fonda [2] (c) doctored image of British soldier pointing machine gun at Iraqi people [3] (d) O. J. Simpso... |

115 | Poisson matting
- Sun, Jia, et al.
- 2004
(Show Context)
Citation Context |

102 | Statistical tools for digital forensics,” in
- Popescu, Farid
- 2004
(Show Context)
Citation Context ...used on compression related artifacts. As the lossy JPEG compression is commonly used in practice, there has been extensive study on double quantization effects in JPEG images for tampering detection =-=[21, 42, 43, 44]-=-. Among these, the double JPEG quantization detection algorithm in [21] will be summarized below since it is more relevant to the focus of this thesis: forensics of digital images. We will also includ... |

98 |
Discriminative random fields,”
- Kumar, Herbert
- 2006
(Show Context)
Citation Context ...uld include such ”inconsistency observations” and utilize this information to determine the optimal hidden labels. This has led us to a slightly different framework, Discriminative Random Field (DRF) =-=[71]-=-, an extension of the Conditional Random Field family. The DRF model has been used to classify the image content in fixed size blocks in an image as natural or human-made. It can be defined as ˆy = ar... |

88 | Tutorial on variational approximation methods
- Jaakkola
- 2001
(Show Context)
Citation Context ... intractable forms in either the E-step or M-step. Proper approximations are therefore needed. Such inexact EM algorithms are termed Variational EM. A summary of approximation methods can be found in =-=[74]-=-. The learning process can also be divided into parameter estimation and inference steps according to the optimal variables in question. The parameter estimation is concerned with model parameters and... |

80 | Higher-order wavelet statistics and their application to digital forensics
- Farid, Lyu
- 2003
(Show Context)
Citation Context ...ibutions of wavelet coefficients [5, 6], produces a sound foundation for many works in this direction. Some commonly used features include correlations between wavelet coefficients in different bands =-=[7]-=- and higher order statistics such as bicoherence [8]. In [9], physics based geometry related features are proposed to distinguish natural images from computer graphics rendered images. The bottom-up a... |

76 | Comparametric Equations with Practical Applications - Mann |

71 | Estimation of primary quantization matrix in double compressed JPEG images,” in - Lukas, Fridrich - 2003 |

70 | Background cut,
- Sun, Zhang, et al.
- 2006
(Show Context)
Citation Context |

70 | Why the logistic function? a tutorial discussion on probabilities and neural networks.
- Jordan
- 1995
(Show Context)
Citation Context ...zij ) −1 (4.4) (4.5) This choice has been theoretically justified in that the logistic model is a natural form of posterior probabilities if the emission probability belongs to the exponential family =-=[72]-=-. As most real world data roughly follows exponential family distributions, it is a sensible choice to use logistic models for posteriors. In this work, since we are using only one detector for authen... |

66 | Radiometric calibration from a single image,” in
- Lin, Gu, et al.
- 2004
(Show Context)
Citation Context ...ck: Sony P72 correctly detected, grey: misclassified) [17]. . . . . . . . . . 36 2.17 Illustration of Camera Response Function (CRF). . . . . . . . . . . . 38 2.18 Color space colinearity along edges =-=[39]-=-. . . . . . . . . . . . . . . . . . 39 2.19 CRF estimation based on color space colinearity (a)(b) test images, with edge patches highlighted in green (c)(d) CRF estimation results (blue: first benchm... |

65 | Exposing digital forgeries through chromatic aberration.
- Johnson, Farid
- 2006
(Show Context)
Citation Context ...nd duplication in scientific images [10], inpainting [11].... etc.7 One typical tampering operation that has been studied by many researchers is splicing, for which many solutions have been proposed =-=[1, 8, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21]-=-. For images of natural scenes, most works rely on the inconsistency among different parts within spliced images. In addition, in [8], splicing is detected via image level statistical analysis and in ... |

61 | Detecting digital image forgeries using sensor pattern noise.
- Lukas, Fridrich, et al.
- 2006
(Show Context)
Citation Context ...on (a)(c) images tampered by copying an area within the image and create duplicates at another location (b)(d) successful detection of forgery with the suspicious areas manually labeled (white areas) =-=[15]-=-. . . . . . . . . . . 27 2.11 Automatic forgery detection and localization using sensor noise correlation (a) sliding masks of various shapes (b) image with a tampered head on the person at the left (... |

56 | Drag-and-drop pasting,”
- Jia, Sun, et al.
- 2006
(Show Context)
Citation Context |

55 | What is the space of camera response functions
- Grossberg, Nayar
(Show Context)
Citation Context ...lection or diffusion in natural scenes, cameras possess a much narrower dynamic range to which the irradiances must adapt to. CRF mimicks traditional films to account for such dynamic range shrinkage =-=[31]-=- and is therefore often concave. Such point-wise nonlinear transform generally stays invariant across different areas of an image. Although some emerging models of cameras may add spatially varying CR... |

53 | Blind detection of photomontage using higher order statistics, in:
- Ng, Chang, et al.
- 2004
(Show Context)
Citation Context ...sound foundation for many works in this direction. Some commonly used features include correlations between wavelet coefficients in different bands [7] and higher order statistics such as bicoherence =-=[8]-=-. In [9], physics based geometry related features are proposed to distinguish natural images from computer graphics rendered images. The bottom-up approach, on the other hand, infers image-level decis... |

42 |
Combining independent tests of significance.
- Fisher
- 1948
(Show Context)
Citation Context ... second over all test images, we need to fuse these p-values. This can be done by meta-analysis, among which we adopt a commonly used method called Fisher’s combined probability test given as follows =-=[79]-=-: χ 2 2I = −2 I∑ log(pi) (4.11) i=1 where pi is the p-value from Wilcoxon’s test of the i-th image, I the total number of images (I=90 in our test). The variable χ 2 2I follows a chi-square distributi... |

41 | Nonintrusive component forensics of visual sensors using output images,
- Swaminathan, Wu, et al.
- 2007
(Show Context)
Citation Context ...nd duplication in scientific images [10], inpainting [11].... etc.7 One typical tampering operation that has been studied by many researchers is splicing, for which many solutions have been proposed =-=[1, 8, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21]-=-. For images of natural scenes, most works rely on the inconsistency among different parts within spliced images. In addition, in [8], splicing is detected via image level statistical analysis and in ... |

41 | Detecting doctored JPEG images via DCT coefficient nalysis,” in
- He, Lin, et al.
- 2006
(Show Context)
Citation Context ...und truth, green: estimated) (a) E1 (b) E4 [34]. . . . . . . . . . . . . . . . 49 2.28 Illustration of DQ effect (a) scenario (b) DCT coefficient histograms of background and spliced foreground areas =-=[21]-=-. . . . . . . . . . . . . 51 2.29 Double quantization (a)(e) spliced images (b)(f) their DQ detection outputs (c)(g) original authentic images (d)(h) their DQ detection outputs [21]. . . . . . . . . .... |

41 | Determining digital image origin using sensor imperfections,
- Lukas, Fridrich, et al.
- 2005
(Show Context)
Citation Context ...icients of FPN (a) test images from Canon G2 has the highest correlation ρ with Canon G2 reference FPN (b) test images from Nikon D100 also has the highest correlation ρ with Nikon D100 reference FPN =-=[27]-=-. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.10 Manual forgery detection using sensor noise correlation (a)(c) images tampered by copying an area within the image and create duplicat... |

41 | Shum,“Detecting doctored images using camera response normality and consistency”,
- Lin, Wang, et al.
- 2005
(Show Context)
Citation Context ...rmality (a) authentic image, edge patches highlighted in red (b) doctored image, edge patches highlighted in red (c) CRFs estimated from the authentic image (d) CRFs estimated from the doctored image =-=[41]-=-. . . . . . . . 42 2.23 Most detected LPIPs fall on object edges. [34]. . . . . . . . . . . . . . 45 2.24 CRF estimation using geometry invariants [34]. . . . . . . . . . . . . 46 2.25 (Q, R) distribu... |

40 |
Exposing digital forgeris in video by detecting double
- Wang, Farid
(Show Context)
Citation Context ...ombining multiple tampering detection cues. Also, with video cameras, similar post processing cues exist and are even more prominent than digital still cameras. These cues include double MPEG effects =-=[45]-=- and video de-interlacing [46]. 2.3.1 Double JPEG Quantization The double JPEG quantization (DQ) effect resulted from image tampering has been studied in [21, 42, 43, 44]. The representative algorithm... |

39 | Physics-motivated features for distinguishing photographic images and computer graphics, in:
- Ng, Chang, et al.
- 2005
(Show Context)
Citation Context ...undation for many works in this direction. Some commonly used features include correlations between wavelet coefficients in different bands [7] and higher order statistics such as bicoherence [8]. In =-=[9]-=-, physics based geometry related features are proposed to distinguish natural images from computer graphics rendered images. The bottom-up approach, on the other hand, infers image-level decisions bas... |

38 | Determining the radiometric response function from a single grayscale image
- Lin, Zhang
(Show Context)
Citation Context ...ith reported RMSEs at the level of 0.01. Note both Fig. 2.19 and 2.21 display the inverse CRF (convex), rather than the CRF itself (concave).41 Figure 2.20: Equalized greyscale histogram along edges =-=[40]-=-. (a) (b) Figure 2.21: Greyscale CRF estimation based on color space colinearity (a) test image with edge patches highlighted, taken with Canon EOS-1D (b) estimated CRF (blue dashed line: benchmark al... |

37 | What can be known about the radiometric response from images
- Grossberg, Nayar
- 2002
(Show Context)
Citation Context ...Some popular parameterized models are listed as follows: • PCA-based empirical model of response (EMOR) [31] • Single-parameter gamma function R = f(r) = r α0 [32] • Polynomial R = f(r) = ∑ N n=0 rβn =-=[33]-=- • Generalized gamma curve model (GGCM) R = f(r) = r ∑n i=0 αiri [34, 35] Generally, more parameters lead to more accurate representations of the CRF with the drawback of increased complexity. Therefo... |

36 | Exposing digital forgeries through specular highlights on the eye,” in
- Johnson, Farid
- 2007
(Show Context)
Citation Context ... estimated lighting directions (c) successful tampering detection when applied to the famous Kerry-Fonda photograph [1]. . . . . . . . . . . . . . . . . . . . . . . . 16 2.4 Specular reflection model =-=[12]-=-. . . . . . . . . . . . . . . . . . . . . . . 17 2.5 Example spliced image (a) inconsistent specular reflections (b) estimated incident light directions, where each white blob indicates the effect of ... |

36 | Comparing the mean field method and belief propagation for approximate inference
- Weiss
- 2001
(Show Context)
Citation Context ...62, 63]) and is generally intractable, there have been a significant number of approximate solutions, including the traditional simulated annealing, Mean Field (MF) and Loopy Belief Propagation (LBP) =-=[64, 65]-=-. The MRF framework has also been widely used in the image processing community, solving problems as texture analysis, image segmentation and object recognition [66, 67]. For our fusion framework, the... |

29 | Exploiting inference for approximate parameter learning in discriminative fields: An empirical study.
- Kumar, August, et al.
- 2005
(Show Context)
Citation Context ...to numeric solutions seeking maximum likelihood parameters. In such numeric solutions, the gradient functions do not have a close form and additional approximations are needed. The empirical study in =-=[75]-=- has investigated several options to approximate the parameter estimation step (the M-step) in DRF learning, including Pseudo-Marginal Approximation (PMA), Saddle Point Approximation (SPA) and Maximum... |

28 |
Ccd fingerprint method -identification of a video camera from videotaped images.
- Kurosawa, Kuroki, et al.
- 1999
(Show Context)
Citation Context ...n in Fig. 2.11. In addition to the aforementioned two representative works on digital camera sensor noise, there are other efforts utilizing such signature for digital camcoder fingerprinting as well =-=[28, 29]-=-. 2.2.2 Demosaicking The term demosaicking is defined as opposed to mosaicking. Mosaicking refers to the subsampling from multispectral to single color signal at each CCD (or CMOS) sensor. This proces... |

28 |
inverse gamma correction
- Blind
(Show Context)
Citation Context ... and R are assumed to be between [0, 1]. Some popular parameterized models are listed as follows: • PCA-based empirical model of response (EMOR) [31] • Single-parameter gamma function R = f(r) = r α0 =-=[32]-=- • Polynomial R = f(r) = ∑ N n=0 rβn [33] • Generalized gamma curve model (GGCM) R = f(r) = r ∑n i=0 αiri [34, 35] Generally, more parameters lead to more accurate representations of the CRF with the ... |

27 | S.F.: Detecting image splicing using geometry invariants and camera characteristics consistency
- Hsu, Chang
- 2006
(Show Context)
Citation Context ...nd duplication in scientific images [10], inpainting [11].... etc.7 One typical tampering operation that has been studied by many researchers is splicing, for which many solutions have been proposed =-=[1, 8, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21]-=-. For images of natural scenes, most works rely on the inconsistency among different parts within spliced images. In addition, in [8], splicing is detected via image level statistical analysis and in ... |

27 | Automatic estimation of the projected light source direction.
- Nillius, Eklundh
- 2001
(Show Context)
Citation Context ... known. Recent advances of computer vision research attempt to resolve such limitations16 and recover L and N’s from a single image. One representative work is the use of object occlusion boundaries =-=[24]-=-. Leveraging the fact that points along the occlusion boundaries have a surface normal with z component equal to zero and assuming a parameterized ellipsoid model in a localized neighborhood, the auth... |

27 |
Detection of doublecompression in JPEG images for applications in steganography,”
- Pevný, Fridrich
- 2008
(Show Context)
Citation Context ...used on compression related artifacts. As the lossy JPEG compression is commonly used in practice, there has been extensive study on double quantization effects in JPEG images for tampering detection =-=[21, 42, 43, 44]-=-. Among these, the double JPEG quantization detection algorithm in [21] will be summarized below since it is more relevant to the focus of this thesis: forensics of digital images. We will also includ... |

26 | Image Splicing Detection Using Camera Response Function Consistency and Automatic Segmentation”, ICME,
- Hsu, Chanhge
- 2007
(Show Context)
Citation Context ...s/images. The segment level classification on the Basic set is unfortunately only slightly better than random guess (25% precision at 70% recall, as shown in the Precision Recall curve in Fig. 3.15a) =-=[20]-=-. One operating point with precision 29% and recall 44% is shown in the confusion matrix in Table 3.3. When generalized to the unseen Advanced set with post processing, we observe performance decrease... |

21 | K.: Component forensics of digital cameras: A nonintrusive approach - Swaminathan, Wu, et al. - 2006 |

19 | Flash matting
- Sun, Li, et al.
- 2006
(Show Context)
Citation Context |

18 | Using geometry invariants for camera response function estimation
- Ng, Chang, et al.
- 2007
(Show Context)
Citation Context ...red image, edge patches highlighted in red (c) CRFs estimated from the authentic image (d) CRFs estimated from the doctored image [41]. . . . . . . . 42 2.23 Most detected LPIPs fall on object edges. =-=[34]-=-. . . . . . . . . . . . . . 45 2.24 CRF estimation using geometry invariants [34]. . . . . . . . . . . . . 46 2.25 (Q, R) distributions from simulated images with gamma model for the CRF, R = r α0 (a)... |

13 |
Exposing digital forgeries in scientific images, in:
- Farid
- 2006
(Show Context)
Citation Context ...ude the simplest form of copy-andpaste (splicing), edge smoothing/matting after splicing (using either 2D filtering or alpha blending), color adjustment, deletion and duplication in scientific images =-=[10]-=-, inpainting [11].... etc.7 One typical tampering operation that has been studied by many researchers is splicing, for which many solutions have been proposed [1, 8, 12, 13, 14, 15, 16, 17, 18, 19, 2... |

13 |
FRAME: filters, random fields, and minimax entropytowards a unified theory for texture modeling
- Zhu, Wu, et al.
- 1998
(Show Context)
Citation Context ...Loopy Belief Propagation (LBP) [64, 65]. The MRF framework has also been widely used in the image processing community, solving problems as texture analysis, image segmentation and object recognition =-=[66, 67]-=-. For our fusion framework, the traditional MRF is not directly applicable and several revisions need to be made. The seemingly elegant edge structure has to be relaxed in order to incorporate the nat... |

11 | Exposing digital forgeries in interlaced and deinterlaced video
- Wang, Farid
- 2007
(Show Context)
Citation Context ...tection cues. Also, with video cameras, similar post processing cues exist and are even more prominent than digital still cameras. These cues include double MPEG effects [45] and video de-interlacing =-=[46]-=-. 2.3.1 Double JPEG Quantization The double JPEG quantization (DQ) effect resulted from image tampering has been studied in [21, 42, 43, 44]. The representative algorithm from [21] will be summarized ... |

10 | Internet Image Archaeology: Automatically Tracing the Manipulation History - Kennedy, Chang - 2008 |

8 | Statistical Fusion of Multiple Cues for Image Tampering Detection,”
- Hsu, Chang
- 2008
(Show Context)
Citation Context ...The detector outputs are treated as observations and used to recover hidden labels indicating whether each block in the test image belongs to the foreground spliced object or the authentic background =-=[58]-=-. 4.2.1 Fusion as a Labeling Problem In a typical labeling problem, each node i is associated with a binary label yi which takes on values {−1, +1}. These labels are usually hidden and unobserved. Wha... |

6 | A theory of spherical harmonic identities for brdf/lighting transfer and image consistency
- Mahajan, Ramamoorthi, et al.
- 2006
(Show Context)
Citation Context ...directional Reflectance Distribution Function (BRDF). From the signal processing perspective, it relates the surface geometry (surface normals) and the lighting through a continuous space convolution =-=[13]-=-, or equivalently a frequency domain multiplication. It follows that if two parts of an image possess the same multiplication constant, then this image is authentic. Otherwise there is an inconsistenc... |

5 |
CCD or CMOS Camera Noise Characterisation.
- Reibel
- 2003
(Show Context)
Citation Context ... tampering (red line) [13]. . . . 19 2.7 Typical camera imaging pipeline. . . . . . . . . . . . . . . . . . . . . 20 2.8 Contribution of noise components with respect to the incoming irradiance level =-=[25]-=-. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.9 Correlation coefficients of FPN (a) test images from Canon G2 has the highest correlation ρ with Canon G2 reference FPN (b) test im... |

4 |
Vision Models and Applications to Image and Video Processing
- Lambrecht
- 2001
(Show Context)
Citation Context ...Loopy Belief Propagation (LBP) [64, 65]. The MRF framework has also been widely used in the image processing community, solving problems as texture analysis, image segmentation and object recognition =-=[66, 67]-=-. For our fusion framework, the traditional MRF is not directly applicable and several revisions need to be made. The seemingly elegant edge structure has to be relaxed in order to incorporate the nat... |

3 | Statistical and geometric methods for passive-blind image forensics
- Ng
(Show Context)
Citation Context ...do not comply to all three criteria mentioned above. More results can be found in [41].43 2.2.3.3 Single Image CRF Estimation Using Geometry Invariants The CRF estimation method proposed in [34] and =-=[35]-=- takes another route concerning the under-constrained nature of CRF and irradiance signal estimation problem: while it is in general impossible to recover the CRF f and irradiance r from brightness (i... |

1 |
photographer fired over altered image
- Times
- 2003
(Show Context)
Citation Context ...ced side-by-side [1] (b) ex-U.S. presidential election candidate John Kerry spliced side-by-side with actress Jane Fonda [2] (c) doctored image of British soldier pointing machine gun at Iraqi people =-=[3]-=- (d) O. J. Simpson’s photograph with skin color darkened [4]. . . . 2 1.2 Technical problems in image forensics. . . . . . . . . . . . . . . . . . 4 2.1 Natural image generation process. . . . . . . .... |

1 |
CCD fingerprint method for digital still cameras
- Saitoh, Kurosawa, et al.
(Show Context)
Citation Context ...n in Fig. 2.11. In addition to the aforementioned two representative works on digital camera sensor noise, there are other efforts utilizing such signature for digital camcoder fingerprinting as well =-=[28, 29]-=-. 2.2.2 Demosaicking The term demosaicking is defined as opposed to mosaicking. Mosaicking refers to the subsampling from multispectral to single color signal at each CCD (or CMOS) sensor. This proces... |