See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/338865502

Image splicing detection using maskRCNN Image splicing detection using maskRCNN

Article  in  Signal Image and Video Processing · July 2020

DOI: 10.1007/s11760020016360

CITATIONS 89

READS 2,562

4 authors, including:

Belal Ahmed

Alexandria University

3 PUBLICATIONS   97 CITATIONS   

SEE PROFILE

T. Aaron Gulliver

University of Victoria

972 PUBLICATIONS   9,847 CITATIONS   

SEE PROFILE

Saif Alzahir

University of Alaska Anchorage

86 PUBLICATIONS   877 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Saif Alzahir on 28 January 2020.

The user has requested enhancement of the downloaded file.

1 23

Signal, Image and Video Processing ISSN 18631703 SIViP DOI 10.1007/s11760020016360

Image splicing detection using maskRCNN

Belal Ahmed, T. Aaron Gulliver & Saif alZahir

1 23

Your article is protected by copyright and all rights are held exclusively by SpringerVerlag London Ltd., part of Springer Nature. This e

offprint is for personal use only and shall not be selfarchived in electronic repositories. If you wish to selfarchive your article, please

use the accepted manuscript version for posting on your own website. You may further deposit the accepted manuscript

version in any repository, provided it is only made publicly available 12 months after

official publication or later and provided

acknowledgement is given to the original

source of publication and a link is inserted

to the published article on Springer's website. The link must be accompanied by the following text: "The final publication is

available at link.springer.com”.

Signal, Image and Video Processing https://doi.org/10.1007/s11760020016360

ORIGINAL PAPER

Image splicing detection using maskRCNN

Belal Ahmed¹ · T. Aaron Gulliver¹ · Saif alZahir²

Received: 13 June 2019 / Revised: 28 November 2019 / Accepted: 6 January 2020 © SpringerVerlag London Ltd., part of Springer Nature 2020

Abstract Digital images have become a dominant source of information and means of communication in our society. However, they can easily be altered using readily available image editing tools. In this paper, we propose a new blind image forgery detection technique which employs a new backbone architecture for deep learning which is called ResNetconv. ResNetconv is obtained by replacing the feature pyramid network in ResNetFPN with a set of convolutional layers. This new backbone is used to generate the initial feature map which is then to train the MaskRCNN to generate masks for spliced regions in forged images. The proposed network is specifically designed to learn discriminative artifacts from tampered regions. Two different ResNet architectures are considered, namely ResNet50 and ResNet101. The ImageNet, He_normal, and Xavier_normal initialization techniques are employed and compared based on convergence. To train a robust model for this architecture, several postprocessing techniques are applied to the input images. The proposed network is trained and evaluated using a computergenerated image splicing dataset and found to be more efficient than other techniques.

Keywords Image forgery · Image splicing · MaskRCNN · Imagenet

1 Introduction

The advent of the internet along with the plethora of social media and other applications has made digital images a dom inant source of information. They are used to document evidence for legal purposes as well as in medical imaging for diagnostic purposes, sports, and many other fields [1– 3]. While digital imaging provides numerous possibilities for creation, it can also be used to produce forged docu ments. Image forgery is almost as old as photography itself and started as early as 1865 when photographer Mathew Brady added General Francis P. Blair to an original pho tograph to make it appear that he was present as shown in Fig. 1.

Image editing tools have become commonplace, and even

sophisticated software is available free of charge. This allows anyone with a computer to easily manipulate an image. As a consequence, forged images can be found everywhere, on the covers of magazines, in courtrooms, and on the inter

B Belal Ahmed

bahmed@uvic.ca

1 Department of Electrical and Computer Engineering, University of Victoria, Victoria, Canada

2 Department of Computer Science and Engineering, University of Alaska Anchorage, Anchorage, USA

net. Therefore, a robust and effective method for image forgery detection is of great importance in digital image forensics.

Image forgery detection techniques can be classified into

two categories: active and passive. In active methods, cer tain information is embedded into the image during creation such as a watermark or signature. With these methods, image tampering is detected by analyzing the watermark or sig nature. Although watermarking can protect an image from theft, its application is limited because human intervention is required to recover the original watermarkfree image. Conversely, passive techniques do not require manual pro cessing [4]. Forgery changes the image feature statistics and introduces artifacts resulting in inconsistencies. Most pas sive techniques use these inconsistencies to identify forged images.

Image tampering can be done in several ways such as

image splicing, retouching, and copymove forgery [5]. Copymove forgery refers to copying a part of an image and pasting it into the same image to conceal or change infor mation. Several methods have been introduced to detect and localize copymove forgeries [6]. Image retouching is used to change the appearance of a subject in an image [7]. In image splicing, part of an image is copied and pasted into another image to hide or add information. Image splicing is

123

Author's personal copy

Signal, Image and Video Processing

Fig. 1 An example of image splicing

widely used to create forgeries in images. Several approaches have been proposed to detect image splicing based on the abnormal transients at splicing boundaries. A method for image splicing detection based on a natural image model was introduced in [8]. This model uses statistical features extracted from the image and 2D arrays generated by apply ing a multisize block discrete cosine transform (MBDCT) to the image. The statistical features include moments of characteristic functions of wavelet subbands and Markov transition probability matrices. In [9], the method in [8] was improved by capturing intrablock and interblock correla tions using DCT coefficients. The original Markov model obtained using a discrete wavelet transform (DWT) was used to extract additional features. Then, the crossdomain fea tures were used to train a support vector machine (SVM) classifier.

In [10], greylevel run length matrix (GLRLM) texture

features for forged and original images were determined. Then, statistical features extracted from the GLRLM were used to train an SVM for classification. In [11], an approach based on statistical features obtained from the run length and image edges was proposed. The method in [11] was improved in [12] by using a detection algorithm based on approximate run lengths. This improved the detection accuracy from 69% to 75% in less time.

Deep neural networks such as deep belief network [13],

deep autoencoder [14], and convolutional neural network (CNN) [15] have recently been used to extract useful, high level structure representations. This allows deep learning networks to generalize well across a wide variety of tasks such as image classification [16], speech recognition [17], camera model identification [18,19], and image and video manipulation detection [20–23]. A CNN model was intro duced in [20] to detect image splicing by extracting dense features from image patches. These features are concate nated, and then a pooling operation (max or mean) is applied on each patch feature. Then, an SVM classifier is trained on these features for classification. In [24], a twostage deep learning approach was introduced to detect tampering in images. First, a stacked autoencoder model is

trained on the wavelet features of the images to learn com plex features for each image patch. Then, the contextual information of each patch is integrated and used for detec tion.

A good detection method should consider lowresolution

images resulting from compression or resizing. Thus, in [25], a shallow CNN (SCNN) was trained to distinguish the boundaries of forged regions from the original edges in lowresolution images by discriminating changes in chroma and saturation. This was done by first convert ing the image from RGB space to YCrCb space. Then, only the CrCb channels were used in the convolutional layers to exclude the illumination information. A differ ent approach to copymove forgery detection (CMFD) based on a CNN was proposed in [26]. In this method, a pretrained CNN model was fine tuned using 3000 forged images from the ImageNet database. These images were generated by randomly moving a rectangular block from the upper left corner of the images to the cen ter.

In our research, we train a supervised MaskRCNN to

learn the hierarchical features resulting from forgery opera tions such as image splicing. To generate the initial feature map, a new ResNet architecture called ResNetconv is intro duced. This is obtained by replacing the feature pyramid network FPN in ResNetFPN [27] with a set of new convo lutional layers. A comparison is made between ResNetFPN and ResNetconv in terms of the convergence speed. Dif ferent ResNet architectures have been tested and compared including ResNet50 and ResNet101 [28]. For faster con vergence, a transfer learning strategy is used to initialize the proposed network. Different initialization techniques have been developed and compared such as ImageNet [29], Xavier_normal [30], and He_normal [31]. Our method has a higher efficiency than these techniques.

The remainder of this paper is organized as follows.

Section 2 briefly discusses the framework of MaskRCNN, Sect. 3 presents the algorithm used to generate the dataset, the experimental results, comparisons, and analysis. Finally, the conclusions are drawn in Sect. 4.

123

Author's personal copy

Signal, Image and Video Processing

Fig. 2 The MaskRCNN framework [32]

2 MaskRCNN

A CNN consists of several convolutional and pooling lay ers and ends with one or more fully connected layers. Each convolutional layer consists of three steps, convolution, non linear activation, and pooling. After each convolutional layer, a feature map is generated and passed to the next layer. A convolutional layer can be represented by [20]

Fⁿ(X) = pooling

!

f ^n!

Fⁿ⁻¹(X) ∗W ⁿ + B^n""

, (1)

where Fⁿ(X) is the feature map for layer n of the convolution with kernel (filter) and bias given by W ⁿ and Bⁿ, respectively, and ∗denotes convolution.

The mask regional convolutional neural network (Mask

RCNN) model was developed in [32] for semantic segmen tation, object localization, and object instance segmentation. MaskRCNN outperformed all existing singlemodel entries on every task in the 2016 COCO challenge including large scale object detection, segmentation, and captioning dataset [33]. MaskRCNN consists of two stages. The first stage, called region proposal network (RPN), scans the initial fea ture maps and generates region proposals or regions of interest (RoI) which is the same process employed by the fasterRCNN model [34]. In the second stage, an operation known as RoIpooling [35] is applied to each RoI to down sample the feature map by using a nearest neighbor approach. This process selects important features from the feature map. RoIpooling can result in misalignment between the RoI and the extracted features, so RoIalign is applied to each RoI to create more accurate RoIs. In RoIalign, the value of each sample point is calculated using bilinear interpolation from the nearby grid points on the feature map. In addition to predicting the class and bounding boxes for each object, MaskRCNN also generates a binary mask for each RoI using

a fully convolutional network (FCN). Figure 2 shows the framework of the MaskRCNN network.

Both stages of the MaskRCNN are connected to the back

bone structure. The backbone is another deep neural network that is used to create the initial feature map. In principle, the backbonenetworkcouldbeanyCNNpretrainedonanimage dataset such as ResNet [28]. ResNet is an artificial neural net work (ANN) that is based on residual learning. In residual learning, a network is trained by feeding the output of an initial layer to advanced layers to share the earlier residuals [28].

3 Proposed method

With the advent of new techniques, a forged image can eas ily be processed to make it more difficult for a human to detect the forgery [36]. In this section, the proposed method is presented and experiments are conducted to evaluate the effectiveness in detecting image forgery. To create a robust model, it is trained on a large dataset that contains different examples of forgery.

3.1 Image dataset

In order to generate a robust deep learning model, a suffi ciently large dataset is required that contains different kinds of forgery. Currently available image forgery datasets are not large and do not cover a wide range of forgeries. For this reason, a computergenerated dataset is used to train the proposed model.

Computergenerated forged images have been created

using the COCO dataset [33] and a set of random objects with transparent backgrounds [37]. This dataset consists of 80 classes, 80,000 training images, and 40,000 validation

123

Author's personal copy

Signal, Image and Video Processing

Fig. 3 Examples of forged image generation

images. The images have dimensions 480 × 640 pixels. To create image splicing, an object Y was chosen randomly and pasted into a random image X from the COCO dataset to create the spliced image Z which can be represented by the equation: X + Y $ ⇒Z, M. The COCO dataset was origi

nally developed for detecting different types of objects such as cars, people, and vegetables. However, the proposed net work is specifically designed to detect forged regions in images. Thus, the original COCO dataset labels are not used for training. Instead, a new binary mask M is generated for each spliced image with the forged region shown in white and the original region in black.

Figure 3 gives three examples of forged image generation

using image splicing. A subset of 21,000 images was selected randomly from the 80,000 training images to create forged images, and an equal number of original images was selected for a total of 42,000 training images. Validation images were created similarly by selecting 905 original images and an equal number for forged images from the validation images in the dataset for a total of 1810 images. Validation images were selected randomly and are different from the images chosen for training. This ensures the model is evaluated on images not used during training for accurate evaluation. To complete the dataset, binary masks were created for each image with the forged pixels labeled by ones and the others labeled by zeros.

A forged image can be postprocessed to alter any arti

facts resulting from the forgery process. To create a robust model that can detect forgeries in postprocessed images, image augmentation techniques were used on the computer generated dataset. These techniques are rotation, shift, shear, and zoom. In rotation, the image is rotated with an angle chosen randomly between 0^◦and 360^◦. In shifting, a spa tial shift is applied to the image with a width and height shift chosen randomly between 0 and 0.4. After shifting an image, it is cropped to its original dimensions. In shearing,

a random transformation intensity between 40^◦and −40^◦is used to create a shear matrix. Then, this matrix is applied to the image using an affine transformation which results in the upper part of the image shifted to the right and the lower part to the left. In zooming, a zoom is applied to the image by randomly selecting two zoom values from the range [1, 10] for the image width and height. These values are used to cre ate the zoom matrix which is applied to the image using an affine transformation. After zooming, the image is cropped to its original dimensions. Each of the four augmentation tech niques is applied to an image during the training process with probability 0.50. Thus, an image has no augmentation or all fourtechniquesappliedwithprobability0.0625.Addingaug mented images expands the image dataset which improves the generalization ability and helps prevent overfitting.

3.2 Implementation

A MaskRCNN model is used with the ResNet model to extract the initial feature map. This is based on an exist ing implementation by Matterport Inc. released under MIT License [38]. Stochastic gradient descent (SGD) optimiza tionisusedtooptimizetheproposedmodelwithamomentum of 0.99 and a weight decay of 0.001. Using SGD with momentum helps accelerate the gradient vectors in the correct direction, thus leading to faster convergence [39]. Further, a small weight decay is multiplied by the weights after each update iteration to prevent the weights from grow ing too large. An initial learning rate of 0.01 is used, and this is reduced by 10% if the validation loss does not decrease for three consecutive epochs where an epoch denotes an iteration over all training examples. A reducing learning rate helps fine tuning the model to reach its local minimum. The proposed method was implemented using Keras [40] and evaluated on an NVIDIA GeForce GTX 1080 Ti GPU with a memory bandwidth of 11 Gbps.

3.2.1 Initialization

The initialization plays an important role in the conver gence speed of a neural network. A good initialization strategy can reduce the feasible parameter space and help the network learn robust features related to the tampering operations rather than complex image content. Initialization can be done using random weights. Examples of random weight initializations are Xavier_normal [30], He_normal [31], Random_normal [41], and Random_uniform [42]. In Xavier_normal, the weights of the network are initialized from a distribution with zero mean and variance

σ ² = 1/Nin, (2)

123

Author's personal copy

Signal, Image and Video Processing

where Nin is the number of incoming neurons. He_normal has a similar variance given by

σ ² = 2/Nin. (3)

Initialization can also be done using the weights of a net

work trained on a large dataset. In this case, the weights of a network are used to initialize another network to per form a different task. The ImageNet dataset has been used to pretrain networks such as ResNet [29]. This is because ImageNet contains over 14 million images which belong to more than 20,000 classes. Hence, networks pretrained on ImageNet can learn a wide variety of features and so can be a good backbone.

Figure 4 presents the network convergence with initializa

tion using a ResNet network pretrained on ImageNet [29], Xavier_normal [30], and He_normal [31]. The training and validation losses are defined as L = Lcls + Lbox + Lmask where Lcls is the classification loss, Lbox is the bounding box

(a)

(b)

Fig. 4 Convergence performance for three different initialization meth ods, a training and b validation

loss,and Lmask isthemaskloss[28].Classificationlossiscal culated using sparse categorical crossentropy, bounding box loss is calculated using smooth L1 loss, and mask loss is cal culated using binary crossentropy [28]. These results show that the ResNet network pretrained on the ImageNet dataset outperforms the other techniques with a difference of about 0.1 in both training loss and validation loss. Xavier_normal, and He_normal converge to the same minimum value which is about 0.2. The convergence speed is the same for both training and validation for all three methods which indicates that the model can achieve the same accuracy on new exam ples and overfitting did not occur.

3.2.2 Backbone

Convergence speed can also be improved based on the num ber of layers in the backbone. The number of layers depends on how many features exist in the dataset and that the net work needs to be trained on. For this reason, two different ResNet architectures were considered which are ResNet 50 and ResNet101 [28]. ResNet50 consists of five stages. Stage 1 is the input stage which consists of one convolu tional layer followed by batch normalization and activation functions to generate the initial feature map. Stages 2 to 5 have convolutional blocks and identity blocks. Both blocks consist of convolutional layers followed by batch normaliza tion and activation functions. The convolutional blocks have an extra bridge to add residuals learned in the input layer to the output layer. The ResNet101 architecture is the same as ResNet50 except that it has 22 convolutional blocks in stage 4 compared to only 5 in ResNet50. In total, ResNet50 has 50 layers, while ResNet101 has 101 layers.

To study the performance of the proposed model with

ResNet50 and ResNet101 backbones, the convergence is compared. Both models were initialized with ImageNet weights [29] and trained for an equal number of epochs. All layer weights have been used to initialize the proposed model except the output layer. This is because the ImageNet weights are for 1000 output categories, while our network output has only two categories: original and forged. Figure 5 shows the training loss and validation loss versus the number of epochs. Although ResNet101 contains twice the number of convolutional layers as ResNet50, the training loss with ResNet50 converges to a lower minimum. The difference in validation loss is even greater. This is because the number of layers in ResNet50 is sufficient to learn the features in the dataset. On the other hand, ResNet101 has extra layers that cause overfitting, resulting in an increase in the valida tion loss. Both models were tuned by reducing the learning rate by 0.005 if the validation loss did not improve for five consecutive epochs.

ResNetFPN has been shown to improve both the accu

racy and speed of some tasks [38]. A feature pyramid network

123

Author's personal copy

Signal, Image and Video Processing

(a)

(b)

Fig. 5 Convergence performance for ResNet50 and ResNet101, a training and b validation

(FPN) is a feature extractor designed in a pyramid for the pur pose of generating multiscale feature maps. FPN requires its own backbone network in order to create the feature pyramid.

To detect image forgeries, the proposed model needs to

learn features based on the artifacts resulting from forgeries. Adding more layers to the backbone network may increase the convergence speed, but can also lead to overfitting. To choose a backbone suitable for our problem, ResNetFPN is compared with a simplified version of ResNet called ResNet conv. In ResNetconv, the FPN network is replaced by four convolutional layers each with 256 filters, one layer for each of the four feature maps created by ResNet. Figure 6 shows the ResNetFPN and proposed ResNet (ResNetconv) struc tures.

Figure 7 presents the convergence performance with

ResNetFPN and ResNetconv. For a fair comparison, both networks were trained for the same number of epochs. Although ResNetFPN has lower validation and training losses than ResNetconv up to epoch 30, the losses subse

Input Image

ResNet

FPN

(a)

Input Image

ResNet

Convolutional

layers

(b)

Fig. 6 The backbone architectures for a ResNetFPN [27] and b ResNetconv

quently converge to the same minimum for both networks. Thus, adding FPN to ResNet does not improve the train ing which indicates that the feature maps created by ResNet have sufficient features to differentiate between the original and forged regions in an image. This is because the features in a forged region are related to sharp edges, color consis tency with surrounding pixels, and differences in contrast and brightness which are basic features that can be learned in backbone training. FPN is a very complex network that may be necessary for more complex problems such as instance segmentation of objects [38]. In summary, the results pre sented here show that an FPN is not required.

Figure 8 presents the receiver operating characteristic

(ROC) curve for the proposed network. This illustrates the ability of the network to discriminate between original and forged regions. It was created by plotting the true positive ratio (TPR) against the false positive ratio (FPR) at different threshold values. TPR and FPR are given by

TPR = TP

TP + FN

, (4)

FPR = FP

TN + FP

, (5)

where TP is true positive which represents the number of pixels that are correctly detected as forged, FN is false neg ative which represents the number of pixels that are falsely detected as original, and TN is true negative which represents

123

Author's personal copy

Signal, Image and Video Processing

(a)

(b)

Fig. 7 Convergence performance for ResNet50 with and without an FPN, a training and b validation

Fig. 8 The ROC curve for the proposed network

the number of pixels correctly detected as original. Thus, TPR and FPR are determined at the pixel level. They can be calculated by comparing the output binary mask with the truth binary mask. Threshold values from 0 to 1 in steps of 0.01 were used to calculate TPR and FPR for the ROC curve. The network performance can also be measured by calcu lating the area under the curve (AUC) which has a value

between 0 and 1. A high AUC means high TPR and low FPR and thus precise model predictions. The proposed network has an AUC value of 0.967, which is excellent.

4 Conclusion

In this paper, a new image forgery detection method based on a new variation of MaskRCNN network was introduced. A new backbone architecture called ResNetconv was designed to create the initial feature map to train the MaskRCNN. This new backbone is a simplified version of ResNetFPN which is obtained by replacing the FPN with convolutional layers. ResNetconv was shown to have the same conver gence speed as ResNetFPN. Two ResNet architectures were considered, ResNet50 and ResNet101. The convergence of ResNet50 was shown to be faster. This is because fea tures related to forgery are basic features that can be learned in the early layers of the network. Using additional layers does not improve the detection accuracy but rather slows convergence. Different augmentation techniques were used to create a model that is robust to several postprocessing techniques. The ImageNet, Xavier_normal and He_normal initialization techniques were considered. Initialization with ImageNet weights provided better results than the other tech niques. The proposed method was trained and evaluated using computergenerated forged images. Future work can consider generalizing the model to other kinds of image forgery such as copymove forgery or image retouching. This can be done by tuning the model using datasets containing examples of these kinds of forgery.

Acknowledgements The authors would like to thank Radwa Hammad for her comments and advice that greatly improve the manuscript. They would also like to thank the anonymous reviewers for their insightful suggestions and comments.

Compliance with ethical standards

Conflict of interest The authors declare that they have no conflict of interest.

References

1. Qazi, T., et al.: Survey on blind image forgery detection. IET Image

Process. 7(7), 660–670 (2013)

2. Mahdian, B., Saic, S.: Blind methods for detecting image fakery.

In: Proceedings of the IEEE International Carnahan Conference on Security Technology. Prague, Czech Republic, Oct 13–16 (2008)

3. Shivakumar, B., Baboo, L.D.S.S.: Detecting copymove forgery in

digital images: a survey and analysis of current methods. Glob. J. Comput. Sci. Technol. 10(7), 61–65 (2010)

4. Lin, C., Chen, C., Chang, Y.: An efficiency enhanced cluster

expanding block algorithm for copymove forgery detection. In: Proceedings of the IEEE International Conference on Intelligent

123

Author's personal copy

Signal, Image and Video Processing

Networking and Collaborative Systems. Taipei, Taiwan, Sep 2–4 (2015)

5. Ardizzone, E., Bruno, A., Mazzola, G.: Copymove forgery detec

tion by matching triangles of keypoints. IEEE Trans. Inf. Forensics Secur. 10(10), 2084–2094 (2015)

6. Cozzolino, D., Gragnaniello, D., Verdoliva, L.: Image forgery

detection based on the fusion of machine learning and block matching methods. arXiv preprint arXiv:1311.6934 (2013)

7. Bharati, A., et al.: Detecting facial retouching using supervised

deep learning. IEEE Trans. Inf. Forensics Secur. 11(9), 1903–1913 (2016)

8. Shi, Y.Q., Chen, C., Chen, W.: A natural image model approach to

splicing detection. In: Proceedings of the Workshop on Multimedia & Security. Dallas, TX, USA, pp. 51–62, Sep 20–21 (2007)

9. He, Z., et al.: Digital image splicing detection based on Markov

featuresinDCTandDWTdomain.PatternRecognit.45(12),4292– 4299 (2012)

10. Mushtaq, S., Mir, A.H.: Novel method for image splicing detection.

In: Proceedings of the IEEE International Conference on Advances in Computing, Commun. and Informatics. New Delhi, India, Sep 24–27 (2014)

11. Dong, J., et al.: Runlength and edge statistics based approach

for image splicing detection. In: Proceedings of the International Workshop on Digital Watermarking. Busan, Korea, Nov 10–12 (2008)

12. He,Z.,Lu,W.,Sun,W.:Improvedrunlengthbaseddetectionofdig

ital image splicing. In: Proceedings of the International Workshop on Digital Watermarking. Lecture Notes in Computer Science, vol. 7128. Springer, Berlin, Atlantic City, NJ, USA, Oct. 23–26 (2011)

13. Lee, H., Ekanadham, C., Ng, A.Y.: Sparse deep belief net model for

visual area V2. In: Proceedings of the Advances in Neural Infor mation Processing Systems Conference. Vancouver, BC, Canada, Dec 3–5 (2007)

14. Hugo, L., et al.: Exploring strategies for training deep neural net

works. J. Mach. Learn. Res. 10, 1–40 (2009)

15. LeCun, Y., et al.: Gradientbased learning applied to document

recognition. Proc. IEEE 86(11), 2278–2324 (1998)

16. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification

with deep convolutional neural networks. In: Proceedings of the Advances in Neural Information Processing Systems Conference. Lake Tahoe, NV, USA, Dec 3–8 (2012)

17. Swietojanski, P., Ghoshal, A., Renals, S.: Convolutional neural net

works for distant speech recognition. IEEE Signal Process. Lett. 21(9), 1120–1124 (2014)

18. Tuama, A., Comby, F., Chaumont, M.: Camera model identification

with the use of deep convolutional neural networks. In: Proceedings of the IEEE International Workshop on Information Forensics and Security. Abu Dhabi, UAE, Dec 4–7 (2016)

19. Baroffio, L., et al.: Camera identification with deep convolutional

networks (2016)

20. Rao,Y.,Ni,J.:Adeeplearningapproachtodetectionofsplicingand

copymove forgeries in images. In: Proceedings of the IEEE Inter national Workshop on Information Forensics and Security. Abu Dhabi, UAE, Dec 4–7 (2016)

21. Rota, P., et al.: Bad teacher or unruly student: can deep learning

say something in image forensics analysis? In: Proceedings of the IEEE International Conference on Pattern Recognition. Cancun, Mexico, Dec 4–8 (2016)

22. Bayar, B., Stamm, M.C.: A deep learning approach to universal

image manipulation detection using a new convolutional layer. In: Proceedings of the ACM Workshop on Information Hiding and Multimedia Security. Galicia, Spain, Jun 20–22 (2016)

23. Wang, Q., Zhang, R.: Double JPEG compression forensics based

on a convolutional neural network. EURASIP J. Inf. Secur. 2016, 23 (2016)

24. Ying, Z., et al.: Image region forgery detection: a deep learning

approach. In: Proceedings of the Singapore CyberSecurity Con ference. Singapore, Jan 14–15 (2016)

25. Zhang, Z., et al.: Boundarybased image forgery detection by fast

shallow CNN. In: Proceedings of the IEEE International Confer ence on Pattern Recognition. Beijing, China, Aug 20–24 (2018)

26. Ouyang, J., Liu, Y., Liao, M.: Copymove forgery detection

based on deep learning. In: Proceedings of the IEEE International Congress on Image and Signal Processing, BioMedical Engineer ing and Informatics. Shanghai, China, Oct 14–16 (2017)

27. Lin, T.Y., et al.: Feature pyramid networks for object detection.

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA, Jul 22–25 (2017)

28. He, K., et al.: Deep residual learning for image recognition. In: Pro

ceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV, USA, Jun 26–Jul 1 (2016)

29. Deng, J., et al.: Imagenet: a largescale hierarchical image database.

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Miami, FL, USA, Jun 20–25 (2009)

30. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep

feedforward neural networks. In: Proceedings of the International Conference on Artificial Intelligence and Statistics. Sardinia, Italy, May 13–15 (2010)

31. He, K., et al.: Delving deep into rectifiers: surpassing human

level performance on ImageNet classification. In: Proceedings of the IEEE International Conference on Computer Vision. Santiago, Chile, Dec 13–16 (2015)

32. He, K., et al.: Mask RCNN. In: Proceedings of the IEEE Inter

national Conference on Computer Vision. Honolulu, HI, USA, Jul 22–25 (2017)

33. Lin, T.Y., et al.: Microsoft COCO: common objects in context.

In: Proceedings of the European Conference on Computer Vision. Zurich, Switzerland, Sep 6–12 (2014)

34. Ren, S., et al.: Faster RCNN: towards realtime object detection

with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137–1149 (2016)

35. Girshick, R.: Fast RCNN. In: Proceedings of the IEEE Interna

tional Conference on Computer Vision. Santiago, Chile, Dec 13–16 (2015)

36. Meena, K., Tyagi, V.: Image forgery detection: survey and future

directions. In: Shukla, R., Agrawal, J., Sharma, S., Singh Tomer, G. (eds.) Data, engineering and applications, pp. 163–194. Springer, Singapore (2019)

37. Braxmeier, J.: Stunning free images & royalty free stock. https://

pixabay.com. Cited 09Jun2019 (2018)

38. Matterport Inc.: MaskRCNN. https://github.com/matterport/

Mask_RCNN. Cited 09Jun2019 (2017)

39. Qian, N.: On the momentum term in gradient descent learning

algorithm. Neural Netw. 12(1), 145–151 (1999)

40. Kerasteam: Keras. https://github.com/fchollet/keras. Cited 09

Jun2019 (2015)

41. Casella, G., Berger, R.L.: Statistical Inference, 2nd edn. Duxbury

Press, Pacific Grove (2002)

42. Sakamoto, H.: On the distributions of the product and the quotient

of the independent and uniformly distributed random variables. Tohoku Math. J., First Stage 49, 243–260 (1943)

Publisher’s Note Springer Nature remains neutral with regard to juris dictional claims in published maps and institutional affiliations.

123

Author's personal copy

View publication stats