C3. A Robust Image Watermarking Technique Based On ... - IEEE Xplore

31st National Radio Science Conference (NRSC2014) April 28 – 30, 2014, Faculty of Engineering, Ain Shams University, Egypt

C3. A Robust Image Watermarking Technique Based On Image Interlacing Mohamed M. Ibrahim1, Neamat S. Abdel Kader 2, M. Zorkany3 National Telecommunications Institute, Cairo, Egypt, [email protected] 2 Faculty of Engineering, Cairo University, Cairo, Egypt, [email protected], [email protected] 1,3

ABSTRACT As a result to ensure and facilitate data authentication, security and copyright protection of digital media, digital watermarking techniques are widely used in these domains. The non-blind watermarking systems have a main problem that for extracting the watermark image from the watermarked image, the original image is required. This means an overhead over the system resources like memory in both sender and receiver and bandwidth in communications channel. In this paper a solution and more effective non-blind image watermarking for this problem is proposed. The proposed method depends on using image interlacing as a new way in the watermarking systems. The performance of this method is compared with the ordinary non-blind watermarking technique over a wide number of images. Simulation results show the effectiveness of the proposed method and robustness against attacks.

Keywords: watermarking, non-blind, interlacing, sub-images, de-interlacing and DWT I. INTRODUCTION As the rapid development in the information and communication technology the copyright protection of the digital multimedia like image, audio and video is very necessary. Digital watermarking plays an important role in this field by hiding secret information into the original media called watermark. This watermark is then extracted out at receiver to represent the ownership and/or the identity of this digital media [1]. Digital image watermarking is categorized according to the original image domain into two types: Spatial domain and Frequency or Transform domain. In spatial domain the watermark embedding is performed by modifying the original image pixels themselves such as LSB (Least Significant Bit), SSM (Spread-Spectrum Modulation). In Frequency or Transform domain the watermark embedding is performed by modifying the frequency components of the original image such as DFT (Discrete Fourier Transform), DCT (Discrete Cosine Transform), DWT (Discrete Wavelet Transform),… The transformation from spatial domain to frequency domain and the inverse operation makes the frequency domain algorithms more complex than the spatial domain algorithms but more robust [1]. In this paper Three Level DWT is used as an example of frequency domain algorithms as shown in section II. Digital image watermarking is also categorized according to the need of the original image in the watermark recovery operation into two types: Blind and Non-blind watermarking. Non-blind watermarking needs the original image in the watermark recovery operation while blind watermarking doesn’t need it [2]. For more robustness in the both types, the watermark image is encrypted before embedding it in the original image. The encryption key may be symmetric or asymmetric [3]. This encryption key and the watermarking operations parameters formed a secret key between sender and receiver. In this paper Arnold Transform is used as one of popular Symmetric key image encryption methods as shown in section II. The availability of the original image in the watermark recovery operation gives the non-blind watermarking two advantages over the blind watermarking, but also gives one disadvantage. For the advantages: First, more simplicity in watermark embedding and recovery, Second, more robustness against attacks because the extracted watermark is more similar to the original one [2]. For the disadvantage: Double storage capacity in memory in both sender and receiver [1] and double communication traffic per time (Bandwidth) in the communication channel between them. The main goal of this paper is how to use image interlacing to solve this disadvantage of the non-blind image watermarking. More details about the definition of the image interlacing and its history in section II, and the proposed method in section III. Many tools are used to evaluate the performance of any watermarking systems. PSNR [4] (Peak Signal to Noise Ratio): used to measure imperceptibility of the watermarking system by measuring the quality of the watermarked image. NC [4] (Normalized Correlation) or the 978-1-4799-3821-6/14/$31.00 ©2014 IEEE

92


similarity factor: used to measure the robustness of the watermarking system against attacks by measuring the similarity between the original watermark image and the recovered one. The simulation results are given in section IV.

II. TECHNIQUES THAT THE PROPOSED METHOD DEPEND ON A. Arnold Transform The idea of the Arnold Transform [5] as an encryption method for any image of size N x N is the changing of the location of each pixel according to the following equation: (1) Where (x, y) is the current location and (x’, y’) is the new location, after a certain number of iterations called Arnold Period the image is returned to its original state, this period is changing according to the image size, bigger image size bigger Arnold Period (for image of size 32 x 32 the Arnold Period = 48, for image of size 64 x 64 the Arnold Period = 96 and so on). To use Arnold Transform as an encryption method, a number of iterations less than the Arnold period is required, this number of iterations is used as a key of type symmetric which means that the same number of iterations is used in decryption operation but using the inverse of the encryption formula.

B. Three Level DWT Watermarking In DWT the image is passed through a series of low pass and high pass filters which decompose the image into sub-bands of different resolutions [4]. These decompositions can be done at different levels, at the first level, the image is decomposed into four non-overlapping multi-resolution sub-bands: LL1 (Approximate), HL1 (Horizontal), LH1 (Vertical) and HH1 (Diagonal). Here, LL1 is the low frequency component where HL1, LH1 and HH1 are the high frequency (detail) components. At the higher levels, the sub band LL (n) is decomposed into LL (n+1), LH (n+1), HL (n+1) and HH (n+1), where n is equal to 1, 2, 3, ... The low frequency components like LL3 sub-band contains the most significant portions of the image in which any alterations like watermark embedding will make a distortion in the original image. The high frequency components like HH1 sub-band contains the least significant portions of the image that image processing operations like compression will eliminate them. The best areas for watermark embedding are the mid frequency components like HL3, LH3 and HH3 sub-bands. But HH3 sub-band contains edges and textures of the original image; hence it will be excluded. The rest choices are HL3 and LH3 sub-bands. In the Human Visual System (HVS) model the human eye is less sensitive to the horizontal than the vertical components, which means that the watermark embedding in the HL3sub-band is more imperceptible than the LH3 sub-band, hence the target sub-band is HL3 [4].

C. Image Interlacing Image interlacing (also known as interleaving) is an operation in which the original image is divided into subimages. Each sub-image has the same number of unrepeated pixels. The interlacing algorithm indicates the contents of each sub-image. The reverse operation called de-interlacing, in which these sub-images are combined together to generate back the original image [6]. The first use of the image interlacing is how to solve the problem of communicating the internet over a slow communications link. Before image interlacing, the image pixels are loaded in order left-to-right and top-to-bottom which means that the complete loading of the image will take a long time. By image interlacing, the original image is divided into a number of sub-images and these sub-images are loaded in sequence. As a result, the viewer can partially see a degraded copy of the image until it will be a perfectly clear copy. This operation helps the viewer to decide more quickly whether to abort or continue the transmission. Lots of image formats support interlacing, including GIF (Graphics Interchange Format), JPEG (Joint Photographic Experts Group), and PNG (Portable Network Graphics). Each of them has its own interlacing algorithm [7], [8]. Another way to use the image interlacing is presented in paper [9], where the original image is interlaced by rows (One level interlacing) into two sub-images (even rows and odd rows), and two different watermark images are embedded, one in each sub-image (Multiple Watermarking System). It is noticed that the two sub-images are very similar and from this similarity the idea of this paper is gotten.

93


III. PROPOSED METHOD In the non-blind watermarking two identical copies of the original image I and I*are used. At the sender, one of them (I) is watermarked by a watermark image W, scaled by a factor α and then sent to the receiver to be I'. The other one I* is sent as it. During the communications channel, the watermarked image I' is attacked to be I''. At the receiver, the watermark image is recovered to be W'. The following equations present the watermark embedding and extracting operations in order [10]:

I’ = I + αW W’ = (I’’ - I*)/α

(2) (3)

The proposed method is a new way to use the image interlacing in digital watermarking, but in this case in single non-blind watermarking. In this method the original image is interlaced into sub-images, then by calculating the similarity factor (NC) between each pair of them. The pair which gives the biggest NC value contains the most similar two sub-images (close to be identical). These two sub-images can play the same role of the two identical copies of the original image in the ordinary non-blind watermarking. Instead of embedding the watermark image in the whole original image, the embedding is performed in sub-image of it as shown in figure 1. As a result, there is no need to another copy of original image in the watermark image recovery (the main goal of this paper). For the pervious equations (2, 3): I and I* will be replaced by the most similar two sub-images, I' by the watermarked sub-image and I'' by the attacked watermarked sub-image.

Fig. 1: Proposed method. (a) At the sender, and (b) At the receiver.

In the proposed method, the watermarking operation is performed after the interlacing operation directly which means that both operations must be performed on the same colour band (Red or Green or Blue). Also means that our work must be done in two steps: The first step is the selecting of this colour band (interlacing only) and the second step is the whole algorithm (interlacing and watermarking) performed on it. The selected colour band must be the colour band that gives the most similar two sub-images (biggest NC value) for number of test images. Finally to present our solution a comparison is made between an ordinary non-blind watermarking algorithm without image interlacing and the same algorithm with interlacing.

IV. EXPERIMENTAL RESULTS In this paper eight famous standard testing images are used, each of size 512 x 512: “Lena, Baboon, Peppers, F-16, House, Sailboat, Splash &Tiffany”, and a gray scale image “Evil Inside” [11] of size 32 x 32 is used as a watermark image as shown in figure 2.

A. First Step (Interlacing Only) In this step there are two levels of interlacing (two MATLAB codes) are proposed: 1)

One Level Interlacing: For any image of size (rows x columns), there are two types of one level interlacing: By rows only and By columns only. By rows only produces pair of sub-images (Even Rows (ER) and Odd Rows (OR)) each of size (rows/2 x columns) as shown in figure 3 (a). By columns only produces also pair of sub-images (Even Columns (EC) and Odd Columns (OC)) each of size (rows x columns/2) as shown in figure 3 (b). Table 1 presents the NC values between each two sub-images in each pair in different colour bands, the values in bold font indicates the biggest values for each image.

94


2)

Two Level Interlacing: Interlacing an image of size (rows x columns) by rows first then by columns (or in the opposite order) will produce four sub-images: Even Rows Even Columns (EE), Even Rows Odd Columns (EO), Odd Rows Even Columns (OE), and Odd Rows Odd Columns (OO) each of size (rows/2 x columns/2) as shown in figure 3 (c). Also, Table 2 presents the NC values between each two sub-images.

Fig. 2: (a) Test images, and (b) Watermark image.

Fig. 3: Interlacing. (a) One level by rows, (b) One level by columns and (c) Two level. Table 1: One Level Interlacing

Color Band Red Green Blue

SubImages

Lena

Baboon

Peppers

ER - OR EC - OC ER - OR EC - OC ER - OR EC - OC

0.9894 0.9798 0.9824 0.9690 0.9578 0.9329

0.8597 0.9228 0.7579 0.8657 0.8777 0.9072

0.9646 0.9585 0.9774 0.9780 0.9608 0.9648

NC F-16

House

Sailboat

Splash

Tiffany

0.9432 0.9719 0.9652 0.9447 0.9150 0.9625

0.9533 0.9544 0.9279 0.9341 0.9591 0.9751

0.9512 0.9533 0.9622 0.9701 0.9686 0.9697

0.9918 0.9924 0.9837 0.9762 0.9707 0.9785

0.9186 0.9542 0.9200 0.8809 0.9100 0.8961

From the one level interlacing results (Table 1) there are two notes: First, for each test image, the NC values between each two sub-images in each pair are different in the same colour band and in different colour bands, but in general, all values are very close to one (Max = 0.9924 ≈ 1, Min = 0.7579 and Avg = 0.945021 ≈ 1) which indicates that all sub-images in all pairs are very similar to each other. Second, the biggest NC values (bold font) for both test images do not exist in the same colour band. For five test images from the selected eight test images the biggest NC values exists in the red colour band (Lena,Baboon,F-16, Splash, and Tiffany), and for other two test images (Peppers and Sailboat) in the green colour band, and for the rest test image (House) in the blue colour band. So, the colour band that gives the most similar two sub-images in almost test images is the red.

95


Table 2: Two Level Interlacing

Color Band

Red

Green

Blue

SubImages

Lena

Baboon

Peppers

F-16

NC House

EE - EO EE -OE EE -OO EO -OE EO-OO OE-OO EE - EO EE - OE EE -OO EO -OE EO-OO OE-OO EE - EO EE - OE EE -OO EO -OE EO-OO OE-OO

0.9797 0.9865 0.9670 0.9748 0.9866 0.9799 0.9690 0.9811 0.9543 0.9636 0.9809 0.9690 0.9326 0.9554 0.9170 0.9233 0.9560 0.9332

0.9215 0.8653 0.8541 0.8508 0.8654 0.9241 0.8640 0.7604 0.7322 0.7203 0.7609 0.8673 0.9061 0.8788 0.8387 0.8391 0.8788 0.9083

0.9589 0.9636 0.9486 0.9506 0.9630 0.9582 0.9782 0.9767 0.9605 0.9630 0.9770 0.9779 0.9653 0.9628 0.9426 0.9429 0.9631 0.9642

0.9715 0.9585 0.9357 0.9360 0.9586 0.9723 0.9446 0.9672 0.9201 0.9171 0.9684 0.9449 0.9610 0.9190 0.8974 0.8897 0.9188 0.9640

0.9520 0.9557 0.9210 0.9227 0.9562 0.9569 0.9329 0.9150 0.8609 0.8641 0.9179 0.9354 0.9751 0.9548 0.9336 0.9352 0.9550 0.9753

Sailboat

Splash

Tiffany

0.9533 0.9541 0.9399 0.9352 0.9546 0.9534 0.9703 0.9629 0.9483 0.9467 0.9633 0.9699 0.9696 0.9688 0.9513 0.9496 0.9689 0.9698

0.9925 0.9973 0.9905 0.9903 0.9973 0.9924 0.9762 0.9824 0.9617 0.9617 0.9826 0.9762 0.9784 0.9696 0.9513 0.9525 0.9694 0.9787

0.9518 0.9385 0.9043 0.9122 0.9378 0.9564 0.8773 0.9235 0.8381 0.8438 0.9343 0.8847 0.8920 0.9164 0.8397 0.8604 0.9194 0.9000

From the two level interlacing results (table 2) there are the same notes as table 1: First, all NC values are very close to one (Max = 0.9973 ≈ 1, Min = 0.7203 and Avg = 0.935341 ≈ 1). Second, the colour bands that give the biggest NC values for each test image are the same, which indicates that if the one level interlacing for any image gives the most similar two sub-images in specified colour band, the two level interlacing for this image will give the most similar two sub-images in the same colour band. As a summary from both tables, the selected colour band is the red. More than two levels of interlacing can be done, but there is a limitation: More levels of interlacing means more small size sub-images. In order to embed any watermark image in any image (or sub-image of it, or subband of sub-image of it as in this paper), the watermark image size must be equal to the size of this image (or smaller than it and then resized to be equal to it) but not bigger than it.

B. Second Step (Watermarking Algorithm With and Without Interlacing) In this step there are three Matlab codes are proposed: 1) Ordinary Non-Blind watermarking algorithm (without interlacing). 2) The same algorithm with one level interlacing. 3) The same algorithm with two level interlacing. Both three cases are divided into the following parts according to the sequential operations: Fist: At the Sender: 1. Read the cover image, separate the RGB color bands then select the red color band. 2.

Read the watermark image and encrypt it using Arnold Transform (equation 1) by number of iterations less than the Arnold Period according to the watermark image size. This number will be the first part of the secret key between the sender and the receiver.

3.

For both second and third codes, interlace the cover image and get the most similar pair of sub-images. This pair will be the second part of the secret key.

4.

For both second and third codes, select first sub-image from the selected pair and convert it using three level DWT then embed the encrypted watermark image in HL3 sub-band according to equation 2. For the first code, this operation is performed on the cover image directly. The scaling factor α will be the third part of the secret key.

5.

For both second and third codes, de-interlace the red color band of the watermarked image from the watermarked sub-image and other sub-images.

96


6.

Generate the watermarked image using the green and blue color bands of the cover image and send it to the receiver. For the first code, send the cover image and only the first and third parts of the secret key with it. For both second and third codes, send all the secret key parts with it.

Second: At the Communication Channel: During the channel the watermarked image is attacked. These attacks are categorized into five categories [12], one or more example of each category is used as shown in figure 4 and as the following: 1. Geometric Attacks: Cropping by 64 x64 (or 1/8 of the image size) at intermediate and upper left. 2. Noising Attacks: Gaussian and Salt-&-Pepper Noise of ratio 0.001, 0.01 and 0.05 of the image size. 3. De-Noising (Filtering) Attacks: Median Filtering. 4. Format-Compression Attacks: JPEG Compression of quality 70, 50 and 30. 5. Image-processing Attacks: Brightening, Darkening and Sharpening. Finally: At the Receiver: 1. Read the received attacked watermarked image, separate the RGB color bands then select the red. 2. For both second and third codes, interlace the attacked watermarked image then select the pair of subimages as in the second part of the secret key (first one is the attacked watermarked sub-image). 3. For both second and third codes, convert this pair using three level DWT then extract the encrypted watermark image from HL3 sub-band according to equation 3 using the scaling factor (the third part of the secret key). For the first code this operation is performed using both the attacked watermarked image and the received copy of the cover image. 4. Decrypt the extracted watermark image using Inverse Arnold Transform by a number of iterations as in the first part of the secret key. A comparison between the three codes is made in table 3 to represent the proposed solution. From the results there are two notes: First, the results of the three codes are close to each other which indicate that the goal of this paper is achieved. Second, the results of both second and third codes are close to each other which indicate that there is no need for more levels of interlacing because there is no enhancement in results in two level interlacing over one level interlacing. But, in the view of code complexity and processing time the one level interlacing is more preferable than the two level interlacing.

Fig. 4: Attacks.

V. CONCLUSION This paper has proposed a new technique for non-blind image watermarking. The main aim of the proposed scheme is to achieve non-blind watermarking without need of original image at the receiver to save memory and bandwidth. This technique depends on image interlacing. We have tested our proposed scheme by many different

97


standard test images, which gave perfect results. We have compared the performance of our proposed scheme with ordinary non-blind watermarking scheme, the proposed method achieved without any dropping in the performance evaluation of the watermarking system. Furthermore, simulation results showed that the proposed algorithm can achieve excellent robustness against different attacks. Table 3: Comparison between the Ordinary Non-Blind Watermarking and the Proposed Method

NC

Images Interlacing Level Sub-Images Scaling Factor (α) PSNR No Attacks Crop. 1/8 Intermediate Crop. 1/8 Left Corner Gaussian 0.001 Gaussian 0.01 Gaussian 0.05 Salt & pepper 0.001 Salt & pepper 0.01 Salt & pepper 0.05 Median Filtering JPEG 70 JPEG 50 JPEG 30 Brightening Darkening Sharpening

No

F-16 One

Two

1 34.6429 0.9998 0.9533 0.9472 0.9957 0.9588 0.8271 0.9984 0.9850 0.9327 0.9941 0.9570 0.9256 0.8645 0.8446 0.8915 0.9772

EC-OC 1.5 31.1584 0.9981 0.9796 0.9778 0.9941 0.9619 0.8501 0.9968 0.9854 0.9367 -0.9703 0.9376 0.7890 0.1525 0.9981 0.9981 0.9863

OE-OO 1.5 31.1874 0.9971 0.9752 0.9839 0.9927 0.9586 0.8378 0.9960 0.9855 0.9401 -0.3582 0.8484 0.1810 0.1229 0.9971 0.9971 0.9815

No

House One

Two

1 34.6497 0.9997 0.9750 0.9189 0.9958 0.9601 0.8397 0.9987 0.9885 0.9371 0.9837 0.9612 0.9266 0.8621 0.8758 0.9136 0.9677

EC-OC 1.8 29.6274 0.9969 0.9784 0.9765 0.9942 0.9724 0.8807 0.9963 0.9893 0.9577 -0.9104 0.9306 0.8800 0.4006 0.9968 0.9969 0.9827

OE-OO 1.5 31.1996 0.9960 0.9739 0.9830 0.9916 0.9545 0.8376 0.9947 0.9855 0.9382 -0.4935 0.8375 0.4076 0.1714 0.9960 0.9960 0.9805

No

Splash One

Two

1 34.6457 0.9998 0.9813 0.9546 0.9956 0.9609 0.8542 0.9989 0.9874 0.9350 0.9972 0.9574 0.9322 0.8678 0.9610 0.9739 0.9876

EC-OC 1 34.6582 0.9990 0.9805 0.9785 0.9910 0.9253 0.7272 0.9958 0.9763 0.8852 -0.9376 0.8689 0.3212 0.1934 0.9990 0.9990 0.9864

EO-OO 1.2 33.2714 0.9919 0.9719 0.9723 0.9841 0.9270 0.7692 0.9896 0.9718 0.9009 -0.1740 0.2381 0.2043 0.1136 0.9917 0.9920 0.9753

REFERENCES [1] Prabhishek Singh, R S Chadha, “A Survey of Digital Watermarking Techniques Applications and Attacks”, International Journal of Engineering and Innovative Technology (IJEIT), Volume 2, Issue 9, March 2013. [2] AKSHYA KUMAR GUPTA and MEHUL S RAVAL,"A robust and secure watermarking scheme based on singular values replacement", Vol. 37, Part 4, August 2012, pp. 425–440, Indian Academy of Sciences. [3] Navnidhi Chaturvedi, “Various Digital Image Watermarking Techniques And Wavelet Transforms”, International Journal of Emerging Technology and Advanced Engineering, ISSN 2250-2459, Volume 2, Issue 5, May 2012. [4] Baisa L. Gunjal, Suresh N. Mali, “Comparative Performance Analysis of DWT-SVD Based Color Image Watermarking Technique in YUV, RGB and YIQ Color Spaces”, International Journal of Computer Theory and Engineering, Vol. 3, No. 6, December 2011. [5] Chittaranjan Pradhan, Vilakshan Sexena, Ajay Kumar Bisoi, “Non Blind Digital Watermarking Technique Using DCT and Cross Chaos Map”, 2012 International Conference on Communications, Devices and Intelligent Systems (CODIS). [6] http://en.wikipedia.org/wiki/Interlacing [7] http://www.designerdigitals.com [8] http://beej.us/blog/data/image-interlacing [9] B.Sridhar, Dr.C.Arun, “On Secure Multiple Image Watermarking Techniques using DWT”, IEEE-20180, ICCCNT'12, 26th _28th July 2012, Coimbatore, India. [10] Ensaf Hussein, Mohamed A. Belal, “Digital Watermarking Techniques, Applications and Attacks Applied to Digital Media: A Survey”, International Journal of Engineering Research & Technology (IJERT), ISSN: 2278-0181, Vol. 1 Issue 7, September – 2012. [11] Kaushik Deb,Md. Sajib Al-Seraj, Md. Moshiul Hoque, and Md. Iqbal Hasan Sarkar, “Combined DWT-DCT Based Digital Image Watermarking Technique for Copyright Protection”, 2012 7th International Conference on Electrical and Computer Engineering, 20-22 December, 2012, Dhaka, Bangladesh. [12] Chih-Chin Lai, Chih-Hsiang Yeh, Chung-Hung Ko, and Chin-Yuan Chiang, “Image Watermarking Scheme Using Genetic Algorithm”, 2012 Sixth International Conference on Genetic and Evolutionary Computing.

98