See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/301743192

Efficient Markov feature extraction method for image splicing detection using maximization and threshold expansion

Article  in  Journal of Electronic Imaging · April 2016

DOI: 10.1117/1.JEI.25.2.023031

CITATIONS 50

READS 530

4 authors, including:

Il Kyu Eom

Pusan National University

107 PUBLICATIONS   1,181 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Il Kyu Eom on 25 June 2021.

The user has requested enhancement of the downloaded file.

Efficient Markov feature extraction method for image splicing detection using maximization and threshold expansion

Jong Goo Han Tae Hee Park Yong Ho Moon Il Kyu Eom

Jong Goo Han, Tae Hee Park, Yong Ho Moon, Il Kyu Eom, “Efficient Markov feature extraction method for image splicing detection using maximization and threshold expansion,” J. Electron. Imaging 25(2), 023031 (2016), doi: 10.1117/1.JEI.25.2.023031.

Efficient Markov feature extraction method for image splicing detection using maximization and threshold expansion

Jong Goo Han,^a Tae Hee Park,^b Yong Ho Moon,^c and Il Kyu Eom^a,* aPusan National University, Department of Electronics Engineering, 2, BusandaehakRo 63 BeonGil, GeumjeongGu, Busan 46241, Republic of Korea bTongMyong University, Department of Mechatronics Engineering, 428, SinseonRo, NamGu, Busan 48520, Republic of Korea cGyeongsang National University, Department of Aerospace and Software Engineering, 501 Jinjudaro, Jinju 52828, Republic of Korea

Abstract. We propose an efficient Markov feature extraction method for color image splicing detection. The maximum value among the various directional difference values in the discrete cosine transform domain of three color channels is used to choose the Markov features. We show that the discriminability for slicing detec tion is increased through the maximization process from the point of view of the Kullback–Leibler divergence. In addition, we present a threshold expansion and Markov state decomposition algorithm. Threshold expansion reduces the information loss caused by the coefficient thresholding that is used to restrict the number of Markov features. To compensate the increased number of features due to the threshold expansion, we propose an even–odd Markov state decomposition algorithm. A fixed number of features, regardless of the difference direc tions, color channels and test datasets, are used in the proposed algorithm. We introduce three kinds of Markov feature vectors. The number of Markov features for splicing detection used in this paper is relatively small com pared to the conventional methods, and our method does not require additional feature reduction algorithms. Through experimental simulations, we demonstrate that the proposed method achieves high performance in splicing detection. © 2016 SPIE and IS&T [DOI: 10.1117/1.JEI.25.2.023031]

Keywords: splicing forgery; splicing detection; Markov probability matrix; discrete cosine transform; maximization; threshold expan sion; Markov state decomposition; support vector machine.

Paper 15904 received Dec. 15, 2015; accepted for publication Apr. 11, 2016; published online Apr. 29, 2016.

1 Introduction A significant amount of information has been digitized because of the advancements in computer hardware and soft ware. Digital multimedia information, such as voice, image, and video, can be reproduced rapidly and with ease. Because digitized images are easily replicated or manipulated, image forgery techniques such as image splicing and copymove are rendered possible with minimal expertise.¹ Furthermore, it is difficult to verify the authenticity of images. Thus, there are abundant research efforts to detect image splicing and copymove forgeries using various types of image features in several domains. If we can uncover some evidence of image alterations, we can conclude that the image has been forged. Image splicing, the combining of two or more images into a new image, is one of the most common types of image tempering. It is often used as simple entertainment or the ini tial step of a photomontage, which is popular in the image editing area. However, spliced images used for malicious purposes can lead to adverse consequences for society. Therefore, developing reliable splicing detection methods to determine the authenticity of images has become an important issue. In recent years, various kinds of image splic ing detection approaches have been proposed.^2,3 The major ity of the research for splicing detection is based on the fact that the image splicing process can cause discontinuities of

edges and corners. These abnormal transitions are an impor tant clue in the verification of image authenticity. Early attempts to detect spliced images focused on changes of global statistical natures by abrupt discontinuities of the spliced images.^4–6 However, these splicing detection meth ods have a limitation in that the statistical moments for an entire image cannot efficiently reflect the local discontinu ities caused by a splicing operation, and detection rates based on these approaches are not high. Effective splicing detection techniques that exploit local transition features are Markov modelbased approaches.^7–11

Wang et al. modeled the edge of chroma component as a finitestate Markov chain and extracted low dimensional fea ture vectors from its stationary distribution for splicing detec tion. They showed effective detection results using a small number of features. However, they performed under very easy testing conditions. Detection methods using Markov features on the transform domain appear to have a superior performance compared with other statistical features. A Markov model–based splicing detection algorithm⁸ in both discrete cosine transform (DCT) and discrete wavelet transform (DWT) domains was reported in 2012. This method achieved a detection rate of 93.55% on the Colombia gray image dataset.¹² However, this scheme required up to 7290 features. Therefore, a dimension reduc tion algorithm, such as recursive feature elimination, was necessary. An enhanced Markov state selection method⁹

*Address all correspondence to: Il Kyu Eom, Email: ikeom@pusan.ac.kr 10179909/2016/$25.00 © 2016 SPIE and IS&T

Journal of Electronic Imaging 0230311 Mar∕Apr 2016 ^• Vol. 25(2)

Journal of Electronic Imaging 25(2), 023031 (Mar∕Apr 2016)

to reduce the number of features was reported. This approach analyzed the distribution characteristic of transform domain coefficients and mapped the vast number of coefficients with limited states with the coefficients according to various pre supposed function models. However, this method sacrificed detection performance to reduce the number of features. Most recently, an image splicing detection technique¹⁰

using a twodimensional (2D) noncausal Markov model was introduced. In this method, a 2D Markov model was applied in the DCT domain and discrete Meyer wavelet transform domain and the crossdomain features were con sidered the final discriminative features for classification. This scheme achieved a detection rate of 93.36%; however, up to 14,240 features were required. ElAlfy and Qureshi¹¹

proposed a blind detection method of image splicing using Markov features in both spatial and DCT domains. They also used principal component analysis (PCA) to select the most relevant features. They achieved a detection rate of 98.82% with an easier testing condition. Because the abovemen tioned splicing detection approaches are applied to gray image datasets, these are hardly applied to color image data sets, such as CASIA1 and CASIA2.¹³ Moreover, detection algorithms using only luminance information demonstrate poor detection rates for color image sets. Splicing direction schemes applicable to color datasets are presented.^14–18 Muhammad et al. proposed an imposing image forgery detection method based on a steerable pyra mid transform and local binary pattern with feature reduction.¹⁴ This method demonstrated the best performance with a rate of 97.33% on the CASIA2 dataset. However, this scheme requires an enormous number of features and addi tional feature selection techniques to reduce the number of features. A splicing detection approach using the run length, run number, and kernel PCA is introduced.¹⁵ This algorithm achieves good detection accuracy with a small number of features. However, splicing detection methods for color images should test for three color channels, then select one color channel, which has the best detection accuracy. This channel selection process takes three times the training and testing time. A color channel design method¹⁶ to find the most discriminative channel, which is referred to as an opti mal chromalike channel for a given feature extraction method, is presented. Four widely used features for image splicing detection are employed to test the effectiveness of the chromalike channel. This algorithm achieves better detection accuracy than those extracted from traditional color channels. More recently, an image splicing detection method using multiscale Weber local descriptors¹⁷ is intro duced. This algorithm achieves high detection accuracies on three color image datasets for splicing forgery detection. However, this method also requires feature dimension reduc tion as well as color channel selection. In summary, local statistical featurebased splicing detec tion algorithms have two main drawbacks. First, these approaches often require additional feature reduction algo rithms because the features could include numerous redun dancies. The feature reduction algorithms may be dependent on the image datasets because the reduction process is based on the trained and tested dataset. Second, all of the approaches should perform the detection process for three color channels and select one color channel with the best detection accuracy.

In this paper, we present an efficient splicing detection scheme using maximization and threshold expansion to con struct Markov states for feature extraction. In our method, the maximum value among the various directional difference values in the DCT domain of three color channels is used to choose the Markov features. Through the maximization process, we can select a fixed number of features from the Markov feature set, regardless of the difference direc tions, color channels, and test datasets. Additionally, we present the threshold expansion algorithm, which reduces the information loss caused by the coefficient thresholding that is used to restrict the number of Markov features. The even–odd Markov state decomposition algorithm is introduced to reduce the increased number of features due to threshold expansion. In the proposed method, the number of Markov features for splicing detection is from 85 to 170, and we do not require additional feature reduction methods. Based on the simulations, we will demonstrate that our pro posed algorithm can effectively detect splicing forgery. This paper is organized as follows. Section 2 describes the conventional Markov feature extraction methods for image splicing detection. The proposed feature extraction method for detecting splicing forgery is presented in Sec. 3. Section 4 reports the experimental results obtained using the proposed approach, and Sec. 5 draws conclusions from this paper.

2 Conventional Markov Feature Extraction Method Typically, the Markov featurebase splicing detection algo rithm consists of three procedures: coefficient difference calculation, thresholding, and Markov probability matrix extraction. To detect splicing forgery, the Markov transition probabilities are extracted in various domains including spa tial,¹¹ DCT,^8–9,11 and DWT^8–10 domains. In the difference cal culation procedure, the difference values of DCT coefficients for several directions are most frequently used in detection splicing forgery. For a given image with size of N × M, let Bðx; yÞ be a b × b image block with a spatial location in the block ðx; yÞð1 ≤x; y ≤bÞ. Bðx; yÞ can be transformed as fol lows:

EQTARGET;temp:intralink;e001;326;300Rðu; vÞ ¼ DCTðBðx; yÞÞ; (1)

where Rðu; vÞ is a transformed block for Bðx; yÞ, ðu; vÞð1 ≤ u; v ≤bÞ is a location in the transform domain, and DCTfzg is the DCT on z. In general, the truncated difference value in the transform domain is used to calculate a local statistical feature for image splicing detection. Let D^d w^{ðu; vÞ be an} interblock or intrablock difference value with direction d and w ∈finter; intrag. d is denoted by a superscript d ¼ f→ ; ←; ↑; ↓; ↖; ↘; ↙; ↗g indicating the direction of the dif ference calculation. For example, for a horizontal direction left to right, the intrablock difference is represented as D^→ intra^{ðu; vÞ ¼ Rðu; vÞ −Rðu; v þ 1Þ, and for a vertical direc} tion top to bottom, the interblock difference is represented as D^↓ inter^{ðu; vÞ ¼ Rðu; vÞ −Rðu þ b; vÞ. All possible intrablock} or interblock differences can be easily obtained. Examples of constructing the difference of interblock and intrablock are shown in Fig. 1. Because the difference values in the transform domain have a wide dynamic range, the number of the local features becomes extremely large. Therefore, the truncation process

Journal of Electronic Imaging 0230312 Mar∕Apr 2016 ^• Vol. 25(2)

Han et al.: Efficient Markov feature extraction method for image splicing detection. . .

for the difference values is necessarily involved. The trun cated difference value F^d w^{ðu; vÞ is defined by}

EQTARGET;temp:intralink;e002;63;479F^d w^{ðu; vÞ ¼ h}T^f½Dd w^{ðu; vÞg;} (2)

where [] is a round operator and

EQTARGET;temp:intralink;e003;63;436hTfzg ¼
z; jzj ≤T signðzÞT; otherwise ^: (3)

In Eq. (3), signðzÞ indicates the sign of integer z, and T is an integer threshold to restrict the range of the difference values. The threshold determines the size of the Markov feature vec tor. If the value of T is small, the number of features is cor respondingly small. However, this necessarily produces information loss; therefore, the transition probability matri ces may be insufficient to distinguish authentic and forged images. A large value can reduce information loss; however, the number of features becomes correspondingly high. Therefore, determining the value of T is a tradeoff between detection performance and computational cost. In general, T ¼ 3 is often used in Markovmodel based splicing detec tion approaches. The features using F^d w^{ðu; vÞ are extracted in various ways.} The most commonly exploited local statistical feature for image splicing detection is the firstorder Markov transition probability. This is defined as follows:

EQTARGET;temp:intralink;e004;63;210M^d w^{ðs; tÞ ¼ PrðFd} w^{ðu; vÞ ¼ sjFd} w^{ðu 0; v 0Þ ¼ tÞ;} (4)

where M^d w^{ðs; tÞ is the Markov transition probability with} the difference direction d, ðu ⁰; v ⁰Þ is a transform domain index according to d and w, and ðs; tÞ ∈fðn1; n2Þjn1; n2 ¼ −T; · · · ; −1;0; 1; ; · · · ; Tg. The number of features for each direction is ð2T þ 1Þ². In Ref. 8, four intrablock Markov fea ture matrices are used to detect spliced images. Four inter block Markov features are exploited to detect forged images in a similar manner. Consequently, 8ð2T þ 1Þ² Markov fea tures in the DCT domain are used. ElAlfy and Qureshi¹¹

exploit M^→ w ^{ðs; tÞ, M↓} w^{ðs; tÞ, M↘} w ^{ðs; tÞ, and M↖} w ^{ðs; tÞ as}

Markov features in both the DCT and spatial domain. A 2D noncausal Markov model¹⁰ can also be described by the combination of Eq. (4). The majority of the Markov fea ture selection methods reported in the literature can be obtained by using the various combinations of M^d w^{ðs; tÞ.}

3 Proposed Method

3.1 Feature Reduction Using Maximum Operation The image splicing process can cause discontinuities of edges and corners. Because humans are more sensitive to luminance than to chroma, unnatural discontinuities will be left in the chroma channel. In addition, the discontinuities are commonly captured by the difference. Therefore, many splicing detection algorithms exploit the chroma space based on color differences such as the YCrCb space for splicing detection. In addition, the majority of splicing detection approaches use various difference directions. Therefore, the number of Markov features can be significantly increased. Let D^d;c w ^{ðu; vÞ be a block coefficient difference} for color channel c ∈fY; Cb; Crg with d and w. Let jdj be the number of the difference direction. Because a color image has three color channels, there are 3jdj difference blocks. Thus, we have 3jdj × ð2T þ 1Þ² Markov transition probability matrices for detecting spliced images. These fea tures are doubled by using both interblock and intrablock differences. For this reason, many splicing detection schemes exploit feature reduction algorithms, such as recursive fea ture elimination⁸ and PCA.¹⁵

The image splicing process causes discontinuities of edges and corners. These discontinuities appear as the mag nitude of the DCT coefficient. Therefore, we assume that the maximum difference value has an important role in con structing the Markov transition probability matrix. Based on this fact, we present a new Markov feature selection method by using the maximum value of the DCT coefficient differences. Let D^max w ^{ðu; vÞ be the maximum value of} D^d;c w ^{ðu; vÞ. The maximum operations are performed twice} as follows:

Fig. 1 Examples of constructing the difference values in the transform domain. (a) Intrablock difference and (b) interblock difference.

Journal of Electronic Imaging 0230313 Mar∕Apr 2016 ^• Vol. 25(2)

Han et al.: Efficient Markov feature extraction method for image splicing detection. . .

EQTARGET;temp:intralink;e005;63;752D^max w ^{ðu; vÞ ¼ max} c ^fmax d ^½Dd;c w ^{ðu; vÞg:} (5)

At first, the maximum value for three colors is taken, then the maximum value of the jdj directional difference values is obtained. The proposed maximization method generates only one set of coefficient difference values for each DCT block. From Eq. (5), we obtain the Markov states using the maximum operation as follows:

EQTARGET;temp:intralink;e006;63;658F^max w ^{ðu; vÞ ¼ h}T^f½Dmax w ^{ðu; vÞg;} (6)

where F^max w ^{ðu; vÞ is the truncated difference value obtained} by threshold T. The firstorder Markov transition probability is defined by using F^max w ^{ðu; vÞ as follows:}

EQTARGET;temp:intralink;e007;63;594M^max w ^{ðs; tÞ ¼ Pr½Fmax} w ^{ðu; vÞ ¼ sjFmax} w ^{ðu 0; v 0Þ ¼ t;} (7)

where M^max w ^{ðs; tÞ is the Markov transition probability} obtained by F^max w ^{ðu; vÞ with ðs; tÞ ∈fðn}1^{; n}2^Þjn1^{; n}2 ^¼ −T; · · · ; −1;0; 1; ; · · · ; Tg. In this case, only ð2T þ 1Þ² Markov features are chosen regardless of the difference directions and color channels. To verify the effect of the proposed maximization oper ation for the difference value, we calculate the Kullback– Leibler divergence (KLD)¹⁸ for the probability distribution function of the DCT coefficients before and after maximiza tion. The KLD is a measure of the difference between two probability distributions. For the probability distribution functions P1 and P2, the KLD is defined as

EQTARGET;temp:intralink;e008;63;431DKLðP2kP1Þ ¼ X

i P2ðiÞ ln ^P2ðiÞ P1ðiÞ ^; (8)

where DKLðP2kP1Þ is the KLD of P2 from P1. Because the KLD is a nonsymmetric measure, we use DKLðP2kP1Þ þ DKLðP1kP2Þ to calculate the distance measure between the authentic images and spliced images in this paper. To calcu late the KLD values, we use P1 as the probability density function of D^d;c w ^{ðu; vÞ or Dmax} w ^{ðu; vÞ for all authentic images,} and we use P2 as the probability density function for all spliced images. Table 1 shows the KLD values for CASIA1 image dataset. As shown in Table 1, the KLD val ues are increased when the maximization process is exploited. The largest KLD value is achieved when D^max w ^{ðu; vÞ is used. Based on this result, we expect that} the proposed maximization scheme can improve the splicing detection accuracy.

3.2 Threshold Expansion and State Decomposition The feature reduction algorithm using the maximization operation casts away the color and directional information. To compensate dropped information, we introduce the threshold expansion and the Markov state decomposition algorithm in this paper. First, we investigate the distribution of the difference values in the DCT domain in the view of T. Table 2 presents the percentages of the intrablock coefficient difference values greater than T for three difference direc tions. The percentages are obtained using the CASIA1 image dataset¹¹ for the splicing detection test. As show in Table 2, ∼45% of the difference values can be lost in the process of conventional thresholding (T ¼ 3). These results suggest that the conventional Markovbased splicing detec tion methods do not use sufficient information to construct Markov states. The conventional Markov states used in various splicing detection methods are composed of integers with unit inter vals. In our proposed method, we obtain the Markov states using the expanded threshold as follows:

EQTARGET;temp:intralink;e009;326;529eF^max w ^{ðu; vÞ ¼ h}2T^f½Dmax w ^{ðu; vÞg;} (9)

where eF^max w ^{ðu; vÞ is the truncated difference value obtained} by threshold 2T. The Markov states calculated by Eq. (6) range from −2T to þ2T with unit intervals. In this case, the number of features is ð4T þ 1Þ². To prevent an increment in the number of features, the state decomposition algorithm is proposed in this paper. Markov states are decomposed into even and odd numbers as follows:

EQTARGET;temp:intralink;e010;326;421ðse; teÞ ∈fðn1; n2Þjn1; n2 ¼ −2T; · · · ; −2;0; 2; · · · ; 2Tg; (10)

EQTARGET;temp:intralink;e011;326;378ðso;toÞ ∈fðn1;n2Þjn1;n2 ¼ −2T þ1; ···;−1;1;; ···;2T −1g; (11)

where ðse; teÞ and ðso; toÞ are the Markov state for the even part and the odd part, respectively. We construct two Markov transition probabilities as follows:

EQTARGET;temp:intralink;e012;326;306eM^max w ^ðse^{; t}e^{Þ ¼ Pr½eFmax} w ^{ðu; vÞ ¼ s}e^jeFmax w ^{ðu 0; v 0Þ ¼ t}e^; (12)

EQTARGET;temp:intralink;e013;326;263eM^max w ^ðso^{; t}o^{Þ ¼ Pr½eFmax} w ^{ðu; vÞ ¼ s}o^jeFmax w ^{ðu 0; v 0Þ ¼ t}o^; (13)

where eM^max w ^ðse^{; t}e^{Þ and eMmax} w ^ðso^{; t}o^{Þ are the Markov prob} abilities constructed by ðse; teÞ and ðso; toÞ, respectively. The

Table 1 KLD values of before and after maximization operation when T ¼ 3.

KLD value for

Y Cb Cr

→ ↓ ↘ → ↓ ↘ → ↓ ↘

D^d;c w ^{ðu; vÞ} 0.1082 0.0916 0.1005 0.0344 0.0306 0.0313 0.0387 0.0355 0.0378

maxd½D^d;c w ^{ðu; vÞ} 0.1222 0.0624 0.0712

D^max w ðu; vÞ 0.1399

Journal of Electronic Imaging 0230314 Mar∕Apr 2016 ^• Vol. 25(2)

Han et al.: Efficient Markov feature extraction method for image splicing detection. . .

number of features for each difference direction is ð2T þ 1Þ² þ ð2TÞ². By using the threshold expansion and state decomposition method, we can compensate the loss of information by using the maximum operation. When T ¼ 3, we can additionally use 5% of the coefficient differ ence values because the threshold exploited in our method is 2T ¼ 6, as shown in Table 2. Figure 2 shows the concept of the proposed threshold expansion and state decomposition method.

3.3 Markov Feature Selection In this paper, we introduce four Markov transition probabil ity matrices as feature vectors for splicing detection. The first feature vector, f1, is constructed by M^max w ^{ðs; tÞ as follows:}

EQTARGET;temp:intralink;e014;63;404f1 ¼ fðs; tÞjM^max intra^{ðs; tÞ; Mmax} inter^{ðs; tÞg:} (14)

The dimension of this feature vector is 2ð2T þ 1Þ². When T ¼ 3, the dimension of f1 is 98. To compensate the loss caused by dropping the directional and color information, we proposed the threshold expansion and state decomposi tion algorithm. The second feature vector, f2, obtained by using eM^max w ^ðse^{; t}e^{Þ and eMmax} w ^ðso^{; t}o^{Þ is defined as}

EQTARGET;temp:intralink;e015;326;752 f2 ¼ fðse; teÞjeM^max intra^{ðse; teÞ; eMmax} inter^{ðse; teÞg}

∪fðso; toÞjeM^max intra^{ðso; toÞ; eMmax} inter^{ðso; toÞg:} (15)

The dimension of this feature vector is 2ð2T þ 1Þ² þ 2ð2TÞ². When T ¼ 3, the dimension of f2 is 170. We further reduce the number of features by summing two feature vectors as follows:

EQTARGET;temp:intralink;e016;326;667 f3 ¼ fðse; teÞjeM^max intra^{ðse; teÞ þ eMmax} inter^{ðse; teÞg}

∪fðso; toÞjeM^max intra^{ðso; toÞ þ eMmax} inter^{ðso; toÞg;} (16)

where f3 is the Markov feature vector obtained by summing decomposed two feature vectors. When T ¼ 3, the dimen sion of f3 is reduced to 85.

3.4 Summary of Proposed Markov Feature Extraction Method Figure 3 depicts the proposed feature vector extraction method. In this figure, the arrows in the boxes indicate the direction of calculating the difference values and Markov chain constructions. The solid line represents the intrablock operation and the dashed line expresses the inter block operation. The majority of the splicing detection algo rithms for color images should select the color channel with the maximum detection rate. These algorithms require three trainingtesting processes. However, we require only one trainingtesting process to determine the detection rate, as shown in Fig. 3. In conclusion, the proposed algorithm can be used regardless of the color channels and difference directions.

4 Simulation Results

4.1 Datasets and Classifier To verify the performance of the proposed splicing detection method, we selected three datasets: Columbia color DVMM,¹² CASIA1, and CASIA2.¹³ The Columbia color image dataset consists of 183 authentic and 180 spliced images in TIFF format. The image size is 1152 × 768,

Fig. 2 Examples of constructing Markov probability transition matrices by using the conventional method and the proposed method by threshold expansion and state decomposition.

Table 2 Percentages of the intrablock coefficient difference values greater than T.

Threshold T

Percentage of difference values larger than T

→ ↓ ↘ average

3 43.5 43.1 47.8 44.8

4 41.7 41.3 46.1 43.0

5 40.0 39.7 44.5 41.4

6 38.3 38.1 42.8 39.7

7 36.8 36.6 41.3 38.2

Journal of Electronic Imaging 0230315 Mar∕Apr 2016 ^• Vol. 25(2)

Han et al.: Efficient Markov feature extraction method for image splicing detection. . .

and no postprocessing was applied to the forged images. All the forged images are spliced images. The CASIA1 dataset contains 800 authentic and 921 forged color images. Different geometric transforms such as scaling and rotation have been applied on the forged images. All the images have a size of 384 × 256 pixels in JPEG format. The CASIA2 dataset is an extension of the CASIA1 dataset. This dataset consists of 7491 authentic and 5123 forged color images in JPEG, BMP, and TIFF formats, where image sizes vary from 240 × 160 to 900 × 600 pixels. To detect splicing forgery, a support vector machine (SVM) classifier with the radial basis function (RBF) kernel¹⁹ was employed in our work. The important parameters of the RBF kernel SVM are

“complexity” and “shape.” These parameters were set by a grid search processing. We used sixfold crossvalidation to evaluate the SVM model parameters. In sixfold crossval idation, the authentic images and the spliced images were randomly divided into six equal groups each. In each itera tion, five groups each from the authentic images and the forged images were used for training, while the remaining images were used for testing. Therefore, at the end of six iterations, all six groups were tested. There was no overlap between the training set and the testing set in an iteration. As with other studies, we constructed independent experiments 30 times and used the average results to reduce the stochastic impact.

Fig. 3 Block diagram of the proposed feature extraction method for splicing detection.

Table 3 Detection performances of the proposed two features when T ¼ 3 (unit: %).

Feature vector # of features TPR TNR ACC AUC

DVMM Columbia color dataset

f1 98 91.35  1.05 92.66  0.49 91.98  0.66 91.92  0.53

f2 170 91.28  1.12 94.58  0.85 92.89  0.61 92.91  0.65

f3 85 87.94  0.89 92.76  0.86 91.83  1.28 90.32  0.70

CASIA1 dataset

f1 98 95.79  0.52 98.08  0.36 97.01  0.27 96.93  0.27

f2 170 98.26  0.30 98.74  0.15 98.50  0.13 98.48  0.17

f3 85 97.62  0.45 98.07  0.28 97.86  0.22 97.84  0.23

CASIA2 dataset

f1 98 94.88  0.04 94.11  0.23 94.58  0.11 94.53  0.13

f2 170 95.29  0.19 94.26  0.31 94.87  0.19 94.77  0.19

f3 85 94.14  0.18 92.57  0.17 93.47  0.18 93.32  0.15

Journal of Electronic Imaging 0230316 Mar∕Apr 2016 ^• Vol. 25(2)

Han et al.: Efficient Markov feature extraction method for image splicing detection. . .

4.2 Splicing Detection Results of Proposed Method To evaluate the performance, we calculate the true positive rate (TPR), the true negative rate (TNR), the accuracy (ACC), and the area under curve (AUC). The TPR is the

rate of accurately identified authentic images, and the TNR is the rate of correctly identified spliced images. The ACC represents the detection rate, which is the average of the TPR and TNR values. The AUC is calculated as the area under the receiver operating characteristic curve. A per fect model will score an AUC of 1, whereas random guessing will score an AUC of around 0.5. Our algorithm is imple mented in MATLAB R2013a. Tests are performed on a desk top running 64bit Windows 7 with 16.0 GB RAM and Intel (R) Core(TM) i53570 3.40 GHz CPU. The detection time for an image is composed of two parts: feature extraction time and verification time. We select a 900 × 600 image ran domly of CASIA2 dataset and perform the detection process 10 times. The average feature extraction times for f1, f2, and f3 are 1.66, 1.71, and 1.72 s, respectively. The average veri fication time for three features is ∼0.02 s. In total, the maxi mum splicing detection time is ∼1.74 s. Table 3 illustrates the detection performance of the pro posed method on the three tested datasets. In Table 3, the values are the 30 averaged detection rates and their standard deviations (average  standard deviation). The detection accuracies are 91.98% (for f1), 92.89% (for f2), and 91.83% (for f3) for the Columbia DVMM color database, as shown in Table 3(a). The detection accuracy for the CASIA1 image set achieves 97.01% when the maximum operation is exploited (feature vector, f1), as shown in Table 3(b). When the feature vector is f3, the detection accu racy is 97.86%. The best performance, 98.50%, is achieved when the feature vector f2 is used. This is because the dimen sion of the feature vector, f2 is the largest, that is, 170. As shown in Table 3(c), the detection accuracies of the CASIA2 image set are 94.59%, 94.87%, and 93.47% for the feature vectors f1, f2, and f3, respectively. The best detection accu racy is achieved when f2 is used because it has the largest number of features. The feature vectors f1 and f3, which exploit the maximization operation, state decomposition, and summation, also achieve proper splicing detection per formance despite a relatively small number of features. From these results, we can verify that the four proposed Markov features can generate reasonable splicing detection performance. In the CASIA1 database, two types of forgeries, such as copymove and splicing, can be found. The CASIA2 dataset

Table 4 Detection accuracies according to the various forgery types for the CASIA1 and CASIA2 datasets (unit: %).

Type of forgeries Feature vector # of features CASIA1 CASIA2

Splicing

f1 98 97.61 96.03

f2 170 97.92 96.73

f3 85 96.84 95.90

Copymove

f1 98 96.97 93.63

f2 170 98.47 94.07

f3 85 96.78 92.00

Rotation

f1 98 — 94.87

f2 170 — 94.70

f3 85 — 94.17

Scaling

f1 98 — 96.17

f2 170 — 96.50

f3 85 — 95.77

Deforming

f1 98 — 97.07

f2 170 — 97.27

f3 85 — 96.50

No transformation

f1 98 — 94.20

f2 170 — 94.30

f3 85 — 92.73

Table 5 Detection results on the comparison between the proposed approach and other methods (unit: %).

Method

DVMM CASIA1 CASIA2 Dimension reduction Channel selection # of features Accuracy # of features Accuracy # of features Accuracy

f1 98 91.98 98 97.01 98 94.50 No No

f3 85 91.83 85 97.86 85 93.47 No No

Ref. 8 100 — — — 100 89.76 Yes No

Ref. 14 359 96.39 475 94.89 3584 97.33 No No

Ref. 15 — — 100 93.80 — — Yes Yes

Ref. 17 316 94.17 770 94.19 359 96.52 Yes Yes

Note: Bold values represent the maximum detection accuracy for each dataset.

Journal of Electronic Imaging 0230317 Mar∕Apr 2016 ^• Vol. 25(2)

Han et al.: Efficient Markov feature extraction method for image splicing detection. . .

has an additional four forgery types including rotation, scal ing, deformation, and no transformation, as well as splicing and copymove. Table 4 shows forgery detection accuracies according to the forgery types in the CASIA1 and CASIA2 datasets. The accuracy ranges between 96.84% and 98.47% for the CASIA1 dataset, and ranges between 92.00% and 97.27% for the CASIA2 dataset. Therefore, the proposed for gery detection algorithm works well for detecting both splic ing and copymove forgery with and without geometric transformations.

4.3 Comparison with Conventional Methods We make a comparison between the proposed method and other stateoftheart image splicing detection schemes. Table 5 presents the comparisons using the three databases used in our experiments. The results of the other methods were taken from the corresponding literature, and the best detection accuracies are presented. The detection accuracies of the proposed method demonstrate reasonable perfor mances, and our method does not require any dimension reduction algorithm and color channel selection. For the CAISA1 dataset, our approach shows the greatest detection accuracy when f1 or f3 is used. For the DVMM and CAISA2 datasets, the methods in Refs. 14 and 17 have greater accu racies than those of the proposed method. However, the num ber of features is greater than that for the proposed algorithm, and the features are varied according to the image datasets. Conversely, the proposed method has a fixed number of fea tures regardless of the image datasets and does not require dimension reduction and color channel selection. In conclu sion, the proposed splicing detection scheme demonstrated reasonable performance with a fixed and relatively small number of features regardless of image datasets and color channels.

5 Concluding Remarks This paper presented an efficient Markov feature extraction method for color image splicing detection. The maximum value among the various directional difference values in the DCT domain of three color channels was used to choose the Markov features. In addition, we introduced the threshold expansion and Markov state decomposition algorithm. The threshold expansion algorithm reduced the information loss caused by coefficient thresholding. We proposed an even– odd Markov state decomposition algorithm to compensate the increased number of features due to threshold expansion. A fixed number of features, regardless of the difference directions, color channels, and test datasets, were used in the proposed algorithm. The experimental results showed that the detection accuracy of the proposed method is 98.50% for CASIA1 and 94.87% for the CASIA2 image dataset. In conclusion, our proposed feature extraction method proved useful within the various image datasets for splicing detection.

Acknowledgments This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (Grant No.: 2012R1A1A2042034).

References

1. H. Farid, “A picture tells a thousand lies,” New Sci. 2411, 38–41 (2003). 2. B. Mahdian and S. Saic, “A bibliography on blind methods for iden tifying image forgery,” Signal Process. Image Commun. 25(6), 389– 399 (2010). 3. H. Farid, “A survey of image forgery detection,” IEEE Signal Process Mag. 2(26), 16–25 (2009). 4. D. Fu, Y. Q. Shi, and W. Su, “Detection of image splicing based on Hilbert–Huang transform and moments of characteristic functions with wavelet decomposition,” in Proc. the 5th Int. Workshop on Digital Watermarking, Vol. 4283, pp. 177–187 (2006). 5. W. Chen, Y. Q. Shi, and W. Su, “Image splicing detection using 2D phase congruency and statistical moments of characteristic function,” Proc. SPIE 6505, 65050R (2007). 6. Y. Q. Shi, C. Chen, and W. Chen, “A natural image model approach to splicing detection,” in Proc. ACM Multimedia and Security, pp. 51–62 (2007). 7. W. Wang, J. Dong, and T. Tan, “Image tampering detection based on stationary distribution of Markov chain,” in Proc. IEEE Int. Conf. on Image Processing, pp. 2101–2104 (2010). 8. Z. He et al., “Digital image splicing detection based on Markov features in DCT and DWT domain,” Pattern Recognit. 45(12), 4292–4299 (2012). 9. B. Su et al., “Enhanced state selection Markov model for image splicing detection,” EURASIP J. Wireless Commun. 2014(7), 1–10 (2014). 10. X. Zhao et al., “Passive imagesplicing detection by a 2D noncausal Markov model,” IEEE Trans. Circuits Syst. Video Technol. 25(2), 185– 199 (2015). 11. M. ElAlfy and M. A. Qureshi, “Combining spatial and DCT based Markov features for enhanced blind detection of image splicing,” Pattern Anal. Appl. 18(3), 713–723 (2015). 12. T. T. Ng and S. F. Chang, “A dataset of authentic and spliced image blocks,” Technical Report 203–2004, Columbia University (2004), http://www.ee.columbia.edu/ln/dvmm/downloads/ (1732015). 13. J. Dong and W. Wang, “CASIA tampered image detection evaluation (TIDE) database,” v1.0 and v2.0 (2011), http://forensics.idealtest.org/ (342015). 14. G. Muhammad et al., “Image forgery detection using steerable pyramid transform and local binary pattern,” Mach. Vision Appl. 25(4), 985–995 (2014). 15. Z. Moghaddasi, H. A. Jalab, and R. Md Noor, “Improving RLRN image splicing detection with the use of PCA and kernel PCA,” Sci. World J. 2014, 606570 (2014). 16. X. Zhao et al., “Optimal chromalike channel design for passive image splicing detection,” EURASIP J. Adv. Signal Process 2012(240), 1–11 (2012). 17. M.Hussainetal.,“Evaluationofimage forgerydetection using multiscale Weber local descriptors,” Int. J. Artif. Intell. Tools 24(4), 1540016 (2015). 18. S. Kullback and R. A. Leibler, “On information and sufficiency,” Ann. Math. Statist. 22(1), 79–86 (1951). 19. C. C. Chang and C. J. Lin, “LIBSVM: a library for support vector machines,” ACM Trans. Intell. Syst. Technol. 2(3), 27 (2011).

Jong Goo Han received his BS degree in electronics engineering from Inje University, Republic of Korea, in 2004 and his MS degree in electronics engineering from Yokohama National University in 2007. Now he is a PhD candidate in electronics engineering at Pusan National University. His research areas of interest are image forensics, image processing, and forgery detection.

Tae Hee Park received her BS and MS degrees in electronics engi neering from Pukying National University, Republic of Korea, in 1993 and 1996, respectively. She received her PhD degree in electron ics engineering from Pusan National University, Republic of Korea, in 2011. Now she is an assistant professor in the Department of Electron ics Engineering, Tongmyung University, Republic of Korea. Her areas of interest are image forensics, image processing, and steganalysis.

Yong Ho Moon received his BS, MS, and PhD degrees in electronics engineering from Pusan National University, Republic of Korea, in 1992, 1994, and 1998, respectively. He has been on the faculty of Gyeongsang National University since 2007 and is now a professor in the Department of Aerospace and Software Engineering. His research areas of interest are image processing, SoC, embedded systems, and digital forensics.

Il Kyu Eom received his BS, MS, and PhD degrees in electronics engineering from Pusan National University, Republic of Korea, in 1990, 1992, and 1998, respectively. He has been on the faculty of Pusan National University since 1997 and is now a professor in the Department of Electronics Engineering. His research areas of interest are image forensics, image processing, and machine learning.

Journal of Electronic Imaging 0230318 Mar∕Apr 2016 ^• Vol. 25(2)

Han et al.: Efficient Markov feature extraction method for image splicing detection. . .

View publication stats