Semi blind rgb color image watermarking using dct and two level svd

Signal & Image Processing : An International Journal (SIPIJ) Vol.4, No.5, October 2013

SEMI-BLIND RGB COLOR IMAGE
WATERMARKING USING DCT AND TWOLEVEL SVD
Bokan Omar Ali1 and Ghazali Bin Sulong2
1, 2

Department of Computer Science, Universiti Teknologi Malaysia

ABSTRACT
This paper presents semi blind RGB color image watermarking using DCT and two-level SVD. First, RGB
image is divided into red, green, and blue channels. The blue component is divided into blocks according
to the watermark size. Second, DCT is applied to each block to form a new block in the transform domain.
DC component is retrieved and assembled from each block to form a new matrix of 128x128 pixels. SVD is
applied to the resultant matrix to obtain matrices, U, S and V. The watermark is embedded into the S
matrix. The watermark can be extracted without original host image, however, matrices U1, S and V1 are
required. Experimental results indicate that the proposed algorithm can satisfy imperceptibility and it is
more robust against common types of attacks such as filtering, adding noise, geometric and compression
attacks.

KEYWORDS
Semi blind watermarking; Digital image watermarking; Frequency domain; Discrete Cosine Transform;
singular value decomposition; attacks.

1. INTRODUCTION
Traditionally, developers have used analog technology to create multimedia applications.
Unfortunately analog technology made it difficult to manipulate multimedia applications because
of limited capacity; however, digital technologies are more flexible as well as more reliable,
allowing easier handling [1].
After further progress in the field of multimedia applications and multimedia distribution, users
find it difficult to protect their own content. Anyone could obtain and easily use their content as
unauthorized copy. Owners need to protect their media content against theft and poor
reproductive performance. Because of the wide use of the internet these types of multimedia have
become unsecure. Anyone can get data from different sources and change this data without the
original owner’s permission. For this reason, many copyright issues have emerged recently.
Digital watermarking is a technique to embed and extract watermark information into digital data
such as (images, Video or audio) to make a statement about the data. A watermark is designed to

DOI : 10.5121/sipij.2013.4501

1


protect multimedia data against copyright infringement or abuse [2]. Watermark information can
be an image or text information about the author.
This paper is organized as follows: the problem background is discussed in Section II. In Section
III the methodology is described. Experimental results are described in Section IV. The
conclusion is drown in Section V.

2. PROBLEM BACKGROUND
Watermarking techniques can be classified into two parts. The first includes the spatial domain
methods, which embed watermark information by directly modifying the host image pixel values.
Using image's LSB to store watermark information is an example of such methods [3]. The
second includes the use of transformation techniques to embed the watermark information in the
frequency domain representation of the host image. The two most commonly used methods in the
field of transformation are DCT and DWT [4]. DWT has better compression ratio without losing
image information but it requires more processing power. DCT needs less processing power but
loses a bit of image information because of blocked artefacts [5].
In a recent watermarking scheme, singular value decomposition (SVD) transformation has been
used. SVD is a powerful tool used to analyze numerical matrices, resulting in a minimum least
squares error truncation [6]. Image watermarking based on SVD provides secure and reliable
owner identification because changing the singular values of the host image pixel does not
change the host image quality much [7, 8]. SVD can embed watermark information in two ways.
The first is by embedding watermark data directly into the host image pixel singular values. The
second is by converting the host image into the frequency domain then embedding the watermark
data in singular coefficient values [9].
SVD may be used in conjunction with DCT to improve and obtain good watermark results in
terms of imperceptibility and robustness [10]. However, most previous work in this area has been
non-blind, and embedded data cannot be extracted without the original host image. [10] This
study proposes a non-blind watermarking schema using a combination of DCT and SVD to
embed a grayscale watermark image within a grayscale host image. This method showed good
performance against some attacks, however it is weak against robust cropping and median filter
attacks.
Another digital image watermarking method is proposed in [13] using DWT and SVD to embed a
16x16 pixel watermark image within a RGB color host image of 200x200 pixels. This schema is
robust against some types of attacks but at the same time performs poorly against Gaussian noise.
A combination of DCT, SVD and edge detection image watermarking technique is used in [11].
This method is robust enough against common attacks but it is too weak against median filter and
Gaussian noise. In addition, this technique depends on large blocks which decrease the capacity.

3. PROPOSED METHOD
In this paper, semi-blind RGB color image watermarking using a combination of DCT and two
levels of SVD is proposed to generate a high degree of imperceptibility and a robust watermarked
image. RGB color image of 512x512 pixels is used as a host image and a grayscale image of
2


128x128 pixels is used as a watermark image. This methodology consists of three main stages
including pre-processing, embedding, extracting.
First, the pre-processing stage procedure is as follows: The RGB color image is divided into three
components Red (R), Green (G) and Blue (B). The Blue components have been chosen for
embedding purpose as it is less sensitive to HVS. The chosen component is divided into square
blocks of (4x4) pixels and DCT is applied to each block to form a new block in transform
domain.
Second, in the embedding stage, The DC component is retrieved and assembled from each
transformed block to form a new matrix of 128x128 pixels. Next, SVD is applied to the resultant
matrix to get three matrices U, S and V using equation (1) below:
svd (A) = USVT

(1)

Where A represents transformed matrix, U and V are two orthogonal matrices and S represents
diagonal matrix. S matrix is stored for reuse in the extraction stage. Grayscale watermark image
is embedded into the S matrix according to equation (2).
W' = S+αW

(2)

Where W' represents watermarked matrix, S is a diagonal matrix, αis scaling factor (in this paper
value ofα=0.1) and W represents the watermark image. Once the process is completed, the SVD
is applied to W' to obtain three matrices U1, S1and V1. U1 and V1 and then stored for reuse in
the extraction stage. Next, three matrices U, S1 and VT are recombined to make a new matrix and
each square block of B component is replaced with a pixel of the obtained matrix. The B channel
is brought back to spatial domain using Discrete Cosine Transform (IDCT) and merged with the
remaining components to form RGB color watermarked image.
Finally, the extraction stage includes extracting watermark image from the watermarked image
without requiring the original host image. However, the stored matrices S, U1 and V1 are
required. The procedure is as follows: RGB watermarked image is divided into three channels R,
G and B. B channel is selected and divided into square blocks of (4x4) pixels and DCT is applied
to each block. Extract and collect DC component from each transformed block to form a new
matrix of (128x128) pixels. Perform SVD on the resultant matrix using equation (1) to create
three matrices U2, S2 and V2 then obtain a new matrix by combining matrices U1, S2 and V1
transfer according to equation (3).
D = U1S2V1T

(3)

Where U1 and V1 are two orthogonal matrices which were stored in the embedding stage and S2
represents a diagonal matrix. Extract watermark image pixel values by using equation (4).
W* = D - S/α

(4)

where W* represents extracted watermark image, D represents obtained matrix by using equation
(3), S represents diagonal matrix stored during embedding stage andα represents scaling factor,
which is = 0.1.
3


Detailed algorithms for the embedding and extracting stages are as follows:

A. WATERMARK EMBEDDING PROCESS
The proposed embedding process is divided into nine steps, which are briefly described below:
Input: a dataset of standard RGB color image used as the host image is denoted by H (512x512);
a grayscale image used as a watermark image is denoted by W (128x128).
Output: RGB color watermarked image.
Step-1: Extract blue component from RGB host image H (512x512) and the selected channel is
then split into square blocks of (4x4) pixels.
Step-2: Apply DCT to each block to obtain DCT_matrix (512x512).
Step-3: Divide DCT_matrix (512x512) into square blocks of (4x4) pixels and then extract and
assemble the DC component of each block to form a new matrix A (128x128).
Step-4: Perform SVD on matrix A to obtain matrices (U, S, and V) then store matrix S to reuse
during watermark extraction.
Step-5: Embed a grayscale watermark image W (128x128) pixels value in matrix S according to
E=S+αW where E represents the obtained matrix, S represents the A singular values, α represents
the predefined embedding strength factor. In this paper the values of α=0.1 and W represent the
watermark image pixels.
Step-6: Perform SVD on E to obtain (U1, S1, and V1) then store U1 and V1 to reuse in the
extraction process.
Step-7: Get matrix A1 according to

A1=US1VT.

Step-8: Replace DC component of each block of DCT_matrix (512x512) with the A1 (128x128)
and apply inverse DCT to each block to obtain matrix Blue_w (512x512).
Step-9: Recombine red, green and Blue_w components to get the RGB color watermarked image
H*(512x512).

B. WATERMARK EXTRACTION PROCESS
The proposed extraction process is divided into six steps which are briefly described below:
Input: RGB color watermarked image; S, U1 and V1 matrices.
Output: Extracted grayscale watermark image.
Step-1: Extract blue component from RGB watermarked image H* (512x512) and then divide
the selected channel into square blocks of (4x4) pixels.
4


Step-2: Apply DCT on each block to obtain DCT_matrix (512x512).
Step-3: Divide DCT_matrix (512x512) into square blocks of (4x4) pixels and then extract and
collect DC component from each block to form a new matrix B (128x128).
Step-4: Perform SVD on the matrix B to obtain matrices (U2, S2, and V2).
Step-5: Obtain matrix D according to D = U1S2V1T.
Step-6: Finally, extract watermark pixel values according to W*=D-S/α.

4. EXPERIMENTAL RESULTS AND DISCUSSIONS
In this study, a dataset of standard RGB color image of (512x512) pixels is used as a host image
and a grayscale image of 128x128 pixels is used as a watermark image. Peak signal to Noise
Ratio (PSNR) is used to measure the quality ratio between cover the image and its watermark.
Normalized cross correlation (NCC) is used to define the similarity between the original and
extracted watermarks. Due to limited space, here we just show the results obtained using two
selected dataset images which are (Lena.jpg and Pepper.jpg). Different attacks such as filtering
(median filter, sharpen filter, motion blur), adding noise (salt & pepper noise, Gaussian, Poisson,
speckle), geometric attacks (rotation and cropping) and compression attacks (JPEG compression)
are used to test the proposed algorithm in term of robustness. Figure 1 shows the original
watermark image. Figures 2 and 3 the selected dataset of standard RGB color host images and
watermarked images with their respective PSNR value. The extracted watermark images under
various types of attacks with their NCC values are shown in Figures 4, 5 and 6. The comparison
between the proposed method and existing methods in [12] and [13] is shown in Table 1.

Fig 1. Original watermark image

5


Fig 2.

Standard dataset of RGB host image

PSNR = 52.7846 dB

PSNR = 54.1647 dB

PSNR = 53.9252 dB

PSNR = 52.0589 dB

PSNR = 54.3248 dB

PSNR = 53.5848 dB

Fig 3. RGB watermarked images with their PSNR values
6


PSNR values for the watermarked images are very high which means that the cover image and its
watermarked are almost identical.

Attacks

Extracted watermark
image from Lena.jpg
host image

NCC

Extracted watermark
image from Pepper.jpg
host image

NCC

Sharpen filter

0.9902

0.9900

Median filter 3x3

0.9855

0.9849

Motion Blur

0.8317

0.8032

Fig 4. Extracted watermark image with their NCC value after attacking watermarked host image with
various types of filter attack.

Attacks

Extracted watermark
image from Lena.jpg
host image

NCC

Extracted watermark
host image

NCC

Salt & Pepper
noise 0.01

0.9889

0.9830

Gaussian noise
0.01

0.9884

0.9614

Speckle noise
0.01

0.9945

0.9964

Poisson noise

0.9954

0.9939

various types of adding noise attack.
7


Extracted watermark
image from Lena.jpg
host image

Attacks

NCC

Extracted watermark
host image

NCC

Rotate 90o

0.9995

0.9971

Cropping 25%

0.9681

0.8460

JPEG (50) quality

0.9916

0.9819

JPEG (80) quality

0.9941

0.9808

various types of geometric attack.

Normalized cross correlation is used to analyze the differences between the original watermark
images and the recovered watermark images which have passed various attacks such as filtering
attacks, adding noise attacks and geometric attacks. NCC results show that the proposed method
has excellent resistance to most common attacks such as sharpening, median filter, salt and
pepper, Poisson noise, cropping, rotation, and JPEG compression.
Table 1: Robustness performance: proposed method verses methods in [12] and [13].

Methods

Proposed
method

Reference[12]

Reference[13]

Sharpen filter

0.9902

0.6128

n/a

Median filter 3x3

0.9855

0.6842

0.8968

Motion Blur
Salt & Pepper noise 0.01
Gaussian noise 0.01
Speckle noise 0.01
Poisson noise
Rotation 90o
Crop 25%
JPEG 50
JPEG 80

0.8317
0.9889
0.9884
0.9945
0.9954
0.9995
0.9681
0.9916
0.9941

n/a
0.9492
0.8365
n/a
n/a
n/a
0.8473
0.8150
0.8610

n/a
n/a
0.7415
n/a
n/a
n/a
0.9031
0.9402
n/a
8


Based on the comparison of the proposed method and the existing works in [12] and [13], the
proposed method is more resistant to the most common attacks when compared to the previous
works.

5. CONCLUSION
This paper offered semi blind color image watermarking based on DCT and two-level SVD. The
proposed method has an excellent resistance to common types of attacks such as filtering, adding
noise and geometric attacks. Without referring to the original host image an embedded
watermark can be extracted from an attacked watermarked image. The experimental results have
shown that the proposed algorithm has a strong resistance to common image processing
operations. A comparison of the proposed method and the existing methods in [12] and [13]
shows that the proposed method has better performance.

REFERENCES
[1]
[2]

[3]
[4]
[5]
[6]

[7]

[8]
[9]

[10]

[11]

[12]

[13]

Friedman, G. L. (1993). The trustworthy digital camera: Restoring credibility to the photographic
image. Consumer Electronics, IEEE Transactions on, 39(4), 905-910.
Dharwadkar, N. V., and Amberker, B. (2010). Watermarking Scheme for Color Images using
Wavelet Transform based Texture Properties and Secret Sharing. International Journal of Signal
Processing, 6(2).
Chang, C.-C., Hsiao, J.-Y., and Chan, C.-S. (2003). Finding optimal least-significant-bit substitution
in image hiding by dynamic programming strategy. Pattern Recognition, 36(7), 1583-1595.
Barni, M., Bartolini, F., Cappellini, V., and Piva, A. (1998). A DCT-domain system for robust
image watermarking. Signal processing, 66(3), 357-372.
Nadenau, M. (2000). Integration of human color vision models into high quality image compression.
École Polytechnique Fédérale De Lausanne, Thesis No, 2296.
Liu, F., and Qian, Y. (2011). A Novel Robust Watermarking Algorithm Based On Two_Levels DCT
and Two_Levels SVD. Paper presented at the Measuring Technology and Mechatronics Automation
(ICMTMA), 2011 Third International Conference on.
Liu, F., Han, K., and zheng Wang, C. (2009). A novel blind watermark algorithm based On SVD
and DCT. Paper presented at the Intelligent Computing and Intelligent Systems, 2009. ICIS 2009.
IEEE International Conference on.
Shi, F., Shi, Y., and Lai, L. (2011). Optimization on digital watermarking algorithm based on SVDDWT. Paper presented at the Granular Computing (GrC), 2011 IEEE International Conference on.
Li, Z., Yap, K.-H., and Lei, B.-Y. (2011). A new blind robust image watermarking scheme in SVDDCT composite domain. Paper presented at the Image Processing (ICIP), 2011 18th IEEE
International Conference on.
Quan, L., and Qingsong, A. (2004). A combination of DCT-based and SVD-based watermarking
scheme. Paper presented at the Signal Processing, 2004. Proceedings. ICSP'04. 2004 7th
Rajani, A., and Ramashri, T. (2011). Image Watermarking Algorithm Using DCT, SVD and Edge
Detection Technique. International Journal of Engineering Research and Applications (IJERA),
1(4), 1828-1834.
Golea, N. H., Seghir, R. and Benzid, R. (2010). A blind RGB color image watermarking based on
singular value decomposition. In Computer System and Applications (AICCSA), IEEE/ACS
Yin, C. Q., Li, L., Lv, A. Q. and Qu, L. (2007, August). Color image watermarking algorithm based
on DWT-SVD. In Automation and Logistics, 2007 IEEE International Conference on (pp. 26072611).
9


AUTHORS
Bokan Omar Ali was born in Jan 1, 1985, Kalar, Kurdistan-Iraq. He received his B.Sc.
Computer Science 2006 University of Duhok, Kurdistan-Iraq. Currently he is M.Sc.
student in University Technology Malaysia (UTM), Johor, Malaysia.

Ghazali Bin Sulong was born in May 21, 1958 Malaysia. He received his Ph.D.
Computing 1989 University of Wales College of Cardiff (UWCC), Wales, U.K., M.Sc.
Computing 1982 University of Wales College Cardiff (UCC), Wales, U.K., B.Sc. Statistic
1979 UKM, Malaysia. Currently he is a Professor in Image Processing.

10

Semi blind rgb color image watermarking using dct and two level svd

Related slideshows

More Related Content

Semi blind rgb color image watermarking using dct and two level svd