Multiple scaling factors based Semi-Blind watermarking ... - IEEE Xplore

42 downloads 0 Views 2MB Size Report
Blind watermarking scheme for grayscale image watermarking using Online Sequential Extreme Learning Machine (OS-ELM) is proposed. Four-level DWT is ...
Multiple scaling factors based Semi-Blind Watermarking of Grayscale Images using OS-ELM Neural Network Ankit Rajpal

Anurag Mishra

Rajni Bala

Department of Computer Science Deendayal Upadhyay College, University of Delhi, New Delhi, India [email protected]

Department of Electronics Deendayal Upadhyay College, University of Delhi, New Delhi, India [email protected]

Department of Computer Science Deendayal Upadhyay College, University of Delhi, New Delhi, India [email protected]

Abstract— In this paper, a multiple scaling factor based SemiBlind watermarking scheme for grayscale image watermarking using Online Sequential Extreme Learning Machine (OS-ELM) is proposed. Four-level DWT is applied on three standard test images of size 512  512 . LL4 sub-band coefficients are chosen for watermark embedding. OS-ELM is initially tuned with a fixed number of training data used in its initial phase and size of data block learned in each iteration. The training set is developed by combining the quantized and desired LL4 sub-band coefficients. This training set is fed to OS-ELM for training. The output of OS-ELM is a sequence of predicted coefficients which is sorted and divided into three equal parts. Multiple Scaling Factor (MSF) scheme is used for embedding a binary watermark in a Semi-Blind manner. Results show good visual quality of signed images and good similarity between the extracted and original watermarks from signed and attacked images. The PSNR and BER are found to be well optimized both in case of signed and attacked images. The time for embedding and extraction is in milliseconds, which makes the proposed technique effective for developing real time watermarking applications.

Index Terms— Semi-Blind Watermarking, Neural Networks, Extreme Learning Machine, OS-ELM, PSNR, BER

I. INTRODUCTION The enormous transmission of multimedia content over the internet requires protection mechanism for avoiding piracy. Digital watermarking is used effectively to provide the realtime protection of high-quality audio, still images, and video sequences [1]. Many research groups have proposed Semi-Blind watermarking techniques using single scaling factor (SSF) for embedding and extraction [2, 3]. Cox et al. [4] suggest the use of multiple scaling factors (MSF) instead of SSF. They are of the opinion that a single scaling factor is not sufficient to effectively modulate the coefficients of all regions in the given test image. According to them, different regional coefficients exhibit variable luminance and contrast levels within the

978-1-5090-2708-8/16/$31.00 ©2016 IEEE

image. Hence, a single scaling factor is unable to create necessary changes in the coefficients to carry out effective watermarking. Instead, they proposed the use of multiple scaling factors for watermarking the image. Over many years, the watermarking is being perceived as an optimization problem. It requires optimization of the twin requirements – visual quality and robustness. This problem is now converged to obtain optimal or suboptimal solutions by minimizing the trade-off between these two parameters. Traditionally, metaheuristic techniques are used to search for optimal or suboptimal solutions in case of different optimization problems. Therefore, various meta-heuristic techniques have been widely reported in the literature for image watermarking. These include Genetic Algorithms (GA) [5], Particle Swarm Optimization (PSO) [6], Ant Colony Optimization (ACO) [7], Firefly Algorithm (FA) [8] and Cuckoo Search Technique (CS) [9]. All these algorithms use both SSF and lately, the MSF to carry out watermarking embedding and extraction from images. The results obtained in case of MSF based metaheuristic techniques clearly outperform those obtained in case of SSF based meta-heuristic techniques. Having even settled that, the use of MSF for watermark embedding has not been extended using any other technique except those which use meta-heuristic concepts. In this paper, we present a novel algorithm for grayscale image watermarking using MSF. This algorithm makes use of a training data set constructed using LL4 sub-band coefficients obtained after computing 4-level DWT of the host image. This dataset is fed to a newly developed fast Single layer feedforward neural network, commonly known as Online Sequential Extreme Learning Machine (OS-ELM). The OSELM learns the training data one-by-one or chunk-by-chunk (with fixed or varying size) and discards the training data as long as the training procedure for those data is completed [10]. This neural network is initially tuned with a fixed number of training data used in its initial phase and size of data block learned in each iteration. The training set is developed by combining the quantized and desired LL4 sub-band coefficients. The output of the OS-ELM is a sequence of

ICSPCC2016

predicted coefficients which are subsequently used in the formulation of watermarking. The basis of selecting the multiple scaling factors (MSF) is described in detail in Section III. The visual quality of the signed and attacked images is evaluated by determining PSNR parameter. Bit Error Rate (BER) of the extracted watermarks from the attacked images is used to measure the degree of similarity between embedded and recovered watermarks [11]. High PSNR values and very low BER values indicate that the proposed watermarking algorithm is robust enough against selected image processing attacks. In addition, the MSF based watermarking technique outperforms SSF based technique using the same embedding and extraction scheme. To the best of our knowledge, the use of MSF to carry out grayscale image watermarking using a neural network is done for the first time. This research paper is organized as follows. Section II gives the review of Online Sequential Extreme Learning Machine (OS-ELM). Section III gives the experimental details of proposed watermarking scheme. Section IV discusses the observed results and their analysis. Finally, the paper is concluded in Section V. II. OS-ELM FORMULATION A novel multiple scaling factor based grayscale image watermarking technique using Online Sequential Extreme Learning Machine (OS-ELM) is implemented. This is done with a view to achieve fast computation, good generalization capability and accuracy. More so, it offers a solution in the form of a system of linear equation H  Y . The SVD method is used to calculate the Moore-Penrose generalized inverse [12] of H according to the Equation (1).

ˆ  H T H  H T Y 1

(1)

where H is the hidden layer output matrix of the neural network, ˆ is the estimate of the output weights and Y is the expected output. The sigmoid activation function in Equation (2) is used in training the OS-ELM.

g ( x)  where



1 1  e  x

is the gain parameter.

(2)

The OS-ELM consists of two main phases. The first phase Boosting phase is to train the SLFNs using the primitive ELM method with some batch of training data in the initialization stage and these boosting training data will be discarded as soon as boosting phase is completed. The required batch of training data is very small, which can be equal to the number of hidden neurons. In this work, the number of hidden neurons has been set to 20, the number of initial training data to 50 and size of block of data learned by OS-ELM in each step to 10. In the second phase - Sequential Learning Phase, the OSELM will learn the train data one-by-one or chunk-by-chunk and all the training data will be discarded once the learning procedure on these data is completed. III. EXPERIMENTAL DETAILS In this work, a multiple scaling factor based Semi-Blind grayscale image watermarking scheme using the OS-ELM in DWT domain is implemented. For this purpose, LL4 sub-band coefficients are used to carry out embedding and extraction processes. The OS-ELM is supplied with the training set of size 1024  2 . This training set is constructed by augmenting a reshaped vector of size 1024  1 containing desired LL4 subband matrix of size 32  32 with the quantized sub-band coefficients of size 1024  1 . Three standard grayscale images of size 512  512 - Baboon, Goldhill and Peppers are used to embed binary watermark after training the OS-ELM using the quantized values of the LL4 sub-band coefficients. The size of the input dataset is 1024 2 while it produces an output of size 1024  1 containing the predicted LL4 sub-band coefficients. These predicted coefficients are first sorted in ascending order and then divided into almost three equal parts. One 32  32 sized binary watermark is embedded by applying three different values of embedding strength (  ). These values are -  1  0.03,  2  0.08, and  3  0.13 . These values are average values obtained in three different regions within the plot of BER vs Scaling Factor. This plot is given and discussed in Section IV. The signed images are tested for visual quality by computing the PSNR. The signed images are attacked with selected image processing attacks to verify robustness of the embedding scheme. These attacks are: - (a) JPEG (QF=75 and QF=90) (b) Rotation (at angle 900 clockwise) and (c) Gaussian Noise (10%) and Salt and Pepper Noise (0.1 % and 0.5%). Semi-Blind extraction of the watermarks from the signed images is carried out before and after executing image processing attacks. The watermark extraction is carried out by using the same OS-ELM model as that obtained while embedding. Thus, only the signed or the attacked image is required to recover the watermark (SemiBlind extraction) as the case maybe. A comprehensive analysis of the results obtained in this simulation is given in Section IV. A. Watermark Embedding Algorithm The watermark embedding algorithm is given as follows:

Fig. 1. Architecture of OS-ELM

ICSPCC2016

1) Apply 4-level DWT on host image of size 512  512 . Obtain LL4 ( 32  32 matrix) coefficients and reshape it to a vector Ci of size 1024  1 . 2) Construct a training set of size 1024 2 by augmenting Ci vector with its quantized vector of size 1024  1 . The neural network model is constructed by OS-ELM and a 1024  1 sized output is produced by this network. This is mathematically represented by Equation 3.  Pi '  OSELM  Round 

 Ci   Q

   

(3)

Fig. 3. Extraction Procedure

C. Performance Evaluation The performance of watermarking algorithm is evaluated by the following parameters. 1) Peak Signal to Noise Ratio (PSNR) The signed images are examined for visual quality by PSNR is used to examine the visual quality of the signed image. This is a full reference metric which requires both original and signed images to assess the quality. The mathematical formulation of PSNR is given by Equation (7).

 I2  PSNR  10 log10  max  (7)  MSE  where I max the maximum possible pixel value of the image I

Fig. 2. Embedding Procedure

3) Sort Pi in ascending order and divide it into three parts. 4) Embed the watermark according to the predicted output of the OS-ELM ( Pi ) by using relevant  value as given in Equation (4). (4) C i'  Pi '    wi In Equation (4), wi is binary watermark,  is the scaling factor which is set to have three numerical values as stated above. These values are -  1  0.03,  2  0.08, and  3

 0.13 . The modified LL4 sub band coefficients

and MSE is the mean square error. 2) Bit-Error Rate or Bit-Error Ratio (BER) BER is a unit less performance measure. It is an approximated estimate of the bit error probability. The BER expression is given in Equation (8): m n

BER   t 1

wt  wt' m n

(8)

where w denotes the original watermark and w’ denotes the recovered watermark. Both w and w’ are of size m  n .

'

obtained after watermark embedding is denoted by C i . 5) Perform Inverse DWT to obtain signed image.

IV. RESULTS AND DISCUSSION Fig. 4(a-c) depicts three grayscale host images – Baboon, Goldhill and Peppers. Fig. 4(d) shows a binary watermark of size 32  32 .

B. Watermark Extraction Algorithm The watermark extraction in a Semi-Blind manner is carried out by using the already trained OS-ELM. The algorithm is given as follows: 1) Apply 4-level DWT on the signed image. Obtain LL4 " coefficients and reshape it to a vector C i of size 1024  1 . " 2) Quantize C i by Q as in Equation (5), and use the already trained OS-ELM model to predict the output:  Pi "  OSELM  Round  

 C i"   Q 

   

(5)

(a)

'

3) Perform watermark extraction ( wi ) using Equation (6). (6) w i'  ( Pi "  C i" )  (1 /  ) where  is the scaling factor.

ICSPCC2016

added to Pi coefficient to modulate it to be Ci which is inversly transformed (IDWT) to obtain the signed image.

(b)

Fig. 5. Effect of Scaling Factor on BER (w,w’) PSNR= 42 dB

(c)

(d)

(a)

Fig. 4. Host images- (a)Baboon (b)Goldhill (c)Peppers (d) Binary Watermark

The binary watermark is embedded within the host images to obtain signed images depicted in Fig. 6(a-c). Their respective PSNR values are mentioned above these images. A high PSNR value shows that the visual quality of these images is good. Figure 5 depicts a plot between BER with respect to scaling factor α. We divide the entire range of scaling factor (0.010.16) in three different sub ranges (0.01-0.06), (0.06 0.10) and (0.10-0.16) and thus take their approximate average values as  1  0.03,  2  0.08, and  3  0.13 respectively. These three average values of the scaling factor are then used for watermark embedding in three different groups of predicted spectral coefficients of the low frequency band of the image available in transform (DWT) domain ( Pi ). As the Pi coefficients produced by the OS-ELM network are sorted in ascending order and further divided into three groups from lowest to highest, the smallest average value of the scaling factor is used for embedding in all the Pi coefficients of the smallest group while the largest average value of the scaling factor is used for embedding in all the Pi coefficients of the largest group. Thus, the scaling factor is actually varied from smallest to largest to get embedded within the Pi coefficients and the effective influence of α is averaged out across all coefficients which contribute to the image formation. This is done in accordance with the formulation given in Equation 4. The influence of the scaling factor shall be present only when the watermark coefficient is 1, else it is absent. This factor is

PSNR= 43 dB

(b)

PSNR= 42 dB

(c)

Fig. 6. Signed Images (a) Baboon (b) Goldhill and, (c) Peppers.

ICSPCC2016

BER=0

BER=0

BER=0.0029

(c)

(b)

(a)

Baboon and Goldhill while for Peppers the BER=0.0039 which is close to 0. A similar result is obtained for QF=90. This indicates a high degree of similarity between the original and recovered watermarks.

Fig. 7. (a-c) Extracted Watermarks from images of Fig. 5(a-c)

Fig. 6(a-c) depicts three different signed images- Baboon, Goldhill and Peppers respectively which are obtained by embedding the binary watermark shown in Fig. 4(d) into images shown in Fig. 4(a-c) respectively. Their respective PSNR values are mentioned just above the figure. High computed value of the PSNR for images shown in Fig. 6 indicates that visual quality after watermark embedding is good. Note that PSNR is a full reference quality assessment metric which requires original / host image and the signed image to work out the quality. Fig. 7(a-c) depicts three binary watermarks recovered from images shown in Fig. 6(a-c) respectively. The visual assessment of the recovered watermarks is also satisfactory. Their respective BER values are reported to be 0, 0 and 0.0029. Note that all BER values are close to 0. It indicates good watermark recovery process. The recovery of the watermark is Semi-Blind in the present work. The processing time consumed in training the OS-ELM, embedding and extraction of the binary watermark is computed in this work. These values are compiled in Table I. TABLE I. PROCESSING TIME INTERVAL FOR OS-ELM TRAINING, EMBEDDING AND EXTRACTION (MILLISECONDS)

Image

Training Time

Embedding Time

Extraction Time

Baboon

115.00

421.87

125.00

Goldhill

121.00

437.50

125.00

Peppers

119.00

430.63

109.37

Note that in Table I, the embedding time includes training time. On the other hand, there is no use of training the network in the extraction process; hence the time consumed is far less than that of embedding. The observed processing time spans clearly indicate that the proposed watermarking scheme is quite suitable to satisfy real time watermarking constraints. This outcome makes the proposed watermarking scheme fit for developing future real time video watermarking applications. To examine the issue of robustness, several image processing attacks described in detail in Section III are applied to the signed images. These results are briefly described below. 1) Image Processing Attacks The following image processing attacks are applied to the signed images. a) JPEG Compression The JPEG Compression with Quality factor (QF) =75 and 90 is applied to the signed images. The recovered watermarks are shown in Table II. For QF=75, the BER values are 0 for

TABLE II. PSNR AND BER VALUES AFTER JPEG COMPRESSION Attack

JPEG (QF=75)

JPEG (QF=90)

Image

PSNR (dB)

BER (W,W’)

Baboon

31.02

0

Goldhill

34.96

0

Peppers

35.57

0.0039

Baboon

36.00

0

Goldhill

37.76

0

Peppers

37.62

0.0029

Extracted Watermark

Rotation The signed image is rotated by 900 clockwise. This result is given in Table II. The PSNR of the attacked image is low. However, the observed BER (W, W’) value is sufficiently low to maintain that the extracted watermark is recognized.

b)

TABLE III. PSNR AND BER VALUES AFTER ROTATION BY 900 Attack

Rotation by 900 Clockwise

Image

PSNR (dB)

BER (W,W’)

Baboon

12.85

0.2695

Goldhill

10.66

0.2695

Peppers

9.85

0.2686

Extracted Watermark

c) Image Noising The last attack onto three signed images is the Gaussian noise attack with 10% noise addition and the salt and pepper noise attack with noise=0.1 % and 0.5 %. In the case of Gaussian noise attack, the PSNR values for the attacked images are high indicating thereby that the visual quality of the attacked images is sufficiently good. The computed BER values are low which indicates a successful watermark recovery. These results are compiled in Table IV. Particularly in case of salt and pepper noise with noise=0.5%, which is quite high, the PSNR values for the attacked images is less as compared to that when noise=0.1%. Similarly, the BER values are high as compared to that when noise=0.1%. However, the extracted watermarks are clearly recognizable in case of both these attacks. It is therefore concluded that the present MSF and OS-ELM based watermarking scheme is sufficiently robust against the image noising attack.

ICSPCC2016

TABLE IV. PSNR AND BER VALUES AFTER IMAGE NOISING Noise Type

Gaussian Noise 10%

Salt & Pepper 0.1%

Salt & Pepper 0.5%

Image

PSNR (dB)

BER (W,W’)

Baboon

30.30

0.0244

Goldhill

30.32

0.0313

Peppers

30.38

0.0391

Baboon

34.54

0.0039

Goldhill

34.46

0.0098

Peppers

34.55

0.0117

Baboon

28.60

0.0752

Goldhill

28.22

0.0820

Peppers

28.01

0.0889

Extracted Watermark

Overall, the results obtained in the present simulation – that is, for signed images, processing time spans computation and PSNR and BER values for attacked images clearly indicate that the required level of optimization and balancing the twin requirements of watermarking are successfully achieved. The visual quality of the signed and attacked images is found to be sufficiently good. In case of few attacks, the observed low BER values for extracted watermarks indicate that the embedding scheme is robust against these attacks. It shall be an interesting avenue to extend this work beyond these preliminary results and to analyze them in view of the mathematical formulation of OS-ELM algorithm in detail. But, it is sure that OS-ELM can be successfully modeled to carry out grayscale image watermarking using the concept of Multiple Scaling Factor (MSF) which is exclusively considered a domain of metaheuristic techniques. The OS-ELM algorithm is capable to minimize the tradeoff between robustness and visual quality at a fast processing speed. This makes it particularly suitable for extending this work to develop video watermarking applications. V. CONCLUSIONS Multiple scaling factor (MSF) for image watermarking is generally considered within the domain of meta-heuristic techniques used for this purpose. In the present work, we successfully carryout grayscale image watermarking using a fast single layer feed-forward neural network, commonly known as OS-ELM. For this purpose, the entire range of the scaling factor is varied in three parts in accordance with the strength of spectral coefficients obtained by computing LL4 sub-band using DWT of the host images. A training dataset is prepared and is fed to OS-ELM. The output of OS-ELM is a sequence of predicted coefficients which is sorted and divided

into three equal parts. Multiple Scaling Factor (MSF) scheme is used for embedding a binary watermark in a Semi-Blind manner. The experimental observations indicate that watermark embedding and extraction has been quite successful. Three image processing attacks are carried out over signed images and binary watermark was extracted out of the attacked images. All the phases of present scheme - training, embedding and extraction are successfully completed within millisecond time interval which makes this scheme suitable for developing real time watermarking applications – both for images and video. REFERENCES [1] Rani, B. Usha, B. Praveena, and K. Ramanjaneyulu, "Literature review on digital image Watermarking," ACM, 2015, p. 43. [2] A. Rajpal, A. Mishra, and R. Bala, "Robust blind watermarking technique for color images using online Sequential extreme learning machine," 2015 International Conference on Computing, Communication and Security (ICCCS), Dec. 2015, pp. 1-7. [3] C.-R. Piao, S. Beack, D.-M. Woo, and S.-S. Han, "A blind Watermarking algorithm based on HVS and RBF neural network for digital image," in Lecture Notes in Computer Science. Springer Science + Business Media, 2006, pp. 493– 496. [4] I. J. Cox, J. Kilian, F. T. Leighton, and T. Shamoon, "Secure spread spectrum watermarking for multimedia," IEEE Transactions on Image Processing, vol. 6, no. 12, pp. 1673– 1687, 1997. [5] C. Agarwal, A. Mishra, and A. Sharma, "Gray-scale image watermarking using GA-BPN hybrid network," Journal of Visual Communication and Image Representation, vol. 24, no. 7, pp. 1135–1146, Oct. 2013. [6] K. Kuppusamy and K. Thamodaran, "Optimized image Watermarking scheme based on PSO," Procedia Engineering, vol. 38, pp. 493–503, 2012. [7] K. Loukhaoukha, "Image Watermarking algorithm based on Multiobjective ant colony optimization and singular value decomposition in Wavelet domain," Journal of Optimization, vol. 2013, pp. 1–10, 2013. [8] A. Mishra, C. Agarwal, A. Sharma, and P. Bedi, "Optimized gray-scale image watermarking using DWT–SVD and Firefly algorithm," Expert Systems with Applications, vol. 41, no. 17, pp. 7858–7867, Dec. 2014. [9] N. Dey, S. Samanta, X. S. Yang, A. Das, and S. S. Chaudhuri, "Optimisation of scaling factors in electrocardiogram signal watermarking using cuckoo search," International Journal of Bio-Inspired Computation, vol. 5, no. 5, p. 315, 2013. [10] V. Fung, T. S. Rappaport, and B. Thoma, "Bit error simulation for pi /4 DQPSK mobile radio communications using two-ray and measurement-based impulse response models," IEEE Journal on Selected Areas in Communications, vol. 11, no. 3, pp. 393–405, Apr. 1993. [11] N.-Y. Liang, G.-B. Huang, P. Saratchandran, and N. Sundararajan, "A fast and accurate online Sequential learning algorithm for Feedforward networks," IEEE Transactions on Neural Networks, vol. 17, no. 6, pp. 1411–1423, Nov. 2006. [12] D. Serre, Matrices. Springer Science & Business Media, 2002.

ICSPCC2016