Proc. of the International Conference on Computer & Communication Engineering 2014 (ICCCE 2014)

23-25 September 2014, Kuala Lumpur, Malaysia

Robust Image Watermarking Scheme by Discrete

Wavelet Transform

Mohammad Abdullatif1

Othman O. Khalifa

2 Akram M. Zeki

3

R.F. Olanrewaju4

1,2,4Department of Electrical and Computer Engineering, International Islamic University Malaysia, Kuala Lumpur, Malaysia

3Department of Information System, International Islamic University Malaysia, Kuala Lumpur, Malaysia

Abstract— This paper proposes a robust digital image

watermarking techniques by using discrete wavelet transform.

The watermark is embedded in 2-level diagonal coefficient to

achieve high performance in term of imperceptibility. The

watermark is embedded by selecting a seed that generating

pseudorandom noise sequence. Then the Watermark is

extracted back by using the same pseudorandom noise

sequence seed in the embedding. The experimental results

show that the proposed algorithm provides good level of

imperceptibility. Moreover, the watermarked image is robust

against several attacks.

Keywords: Digital Image Watermarking, Discrete Wavelet

Transform, Pseudorandom Noise Sequence, Watermarking

Attacks.

I. Introduction

Since the Internet has become very popular, and people

can share whatever they want to share such as images,

videos, documents, etc., there has been a need to protect

publishing copyright. In addition, there has been also a

significant demand for information security. For these

reasons and other reasons, digital image watermarking has

become very popular recent years as a good solution for

these cases. Many researches have gone through this field

to create new techniques, and to enhance current

techniques as proper solutions for previous problems.

Digital image watermarking techniques stand on

embedding a host image with information which is called

watermark, then the watermarked image will be

transmitted, and can be extracted at the receiver. There are

two kinds of detection types at the receiver. The first type

is called blind watermarking, because the detector doesn’t

need the original cover image to detect the watermark. The

second type is called non-blind and it needs the original

cover image to extract the watermark [1].

The basic model of digital image watermarking

consists of two parts; the first part is the watermark

embedding process which shown in Fig. 1, and the second

part is the watermark detection process which shown in

Fig. 2.

Fig. 1. Watermark Embedding

Fig. 2. Watermark Detection

In Fig. 1, which represents sender, the Watermark is

embedded into the Cover Image with the Secret Key that

ensures the security of watermarking process. The output

is the Watermarked Image. In Fig. 2, at the receiver side,

the detector detects the watermark from the Watermarked

Image by using the Secret Key to recover the Watermark

[2].

Digital image watermarking concerns to solve some

issues properly, thus, this paper highlights the main

requirements of watermarked image as following:

Cover Image Watermarked

Image

Watermark

Embedding

Watermark

Secret

Key

Watermarked

Image Watermark

Watermark

Detection

Secret

Key

Proc. of the International Conference on Computer & Communication Engineering 2014 (ICCCE 2014)

23-25 September 2014, Kuala Lumpur, Malaysia

A- Robustness:

The robustness is the ability of detecting the watermark

after some signal processing modification such as spatial

filtering, scanning and printing, lossy compression,

translation, scaling, and rotation [1], and other operations

like digital to analog (D/A), analog to digital (A/D)

conversions, cutting, image enhancement [3]. In addition,

not all watermarking algorithms have the same level of

robustness, Some techniques are robust against some

manipulation operations, however, they fail against other

stronger attacks [4]. Moreover, it’s not always desirable

for watermark to be robust, in some cases; it’s desired for

the watermark to be fragile [1]. Therefore, the robustness

can be classified as following:

Robust: The watermark is designed to be able to

survive against intentional and incidental attacks

[5]. This kind of watermarking can be used in

broadcast monitoring, copyright protection,

fingerprinting, and copy control [6].

Fragile: The watermark in this type is designed to

be destroyed at any kind of modification, to detect

any illegal manipulation, even slight changes,

involving incidental and intentional attacks.

Fragile watermarks are mainly used in content

authentication and integrity verification. They use

blind detection type [6], as it will be discussed in

Detection Types. In addition, the implementation

of fragile techniques is easier than the

implementation of robust ones [7].

Semi-fragile: The watermark in this type is robust

against incidental modifications, but fragile

against malicious attacks [8]. And it is used for

image authentication [9].

B- Imperceptibility:

Imperceptibility (also known as Invisibility and

Fidelity) is the most significant requirement in

watermarking system, and it refers to the perceptual

similarity between the original image before watermarking

process and the watermarked image [1]. In other words,

the watermarked image should look similar to the original

image, and the watermark must be invisible in spite of

occurrence of small degradation in image contrast or

brightness. However, the challenge is that imperceptibility

could be achieved, but the robustness and the capacity will

be reduced, and vice versa, imperceptibility may be

sacrificed by increasing the robustness and the capacity

[4]. Moreover, the watermark not always desired to be

invisible, sometimes, it is preferred to have visible

watermark into the image [6].

C- Capacity:

Capacity (also known as Payload) refers to the number

of bits embedded into the image. The capacity of an image

could be different according to the application that

watermark is designed for [1]. Moreover, studying the

capacity of the image can show us the limit of watermark

information that would be embedded and at the same time

satisfying the imperceptibility and robustness [4].

II. Related Work C. Patvardhan and others [10] proposed a new method

of digital image watermarking using Discrete Wavelet

Packet Transform (DWPT) analysis of the host image. The

binary watermark is embedded in the high frequency

diagonal components of wavelet packet decomposition

tree which have maximum entropy because human eyes

are less sensitive in high frequency bands having

orientation of 450. Watermarking is achieved by

generating a pseudo-random sequence and then embedding

it into wavelet coefficients according to the watermark bit

pattern. The results showed that this algorithm provides a

good level of imperceptibility and robustness against

various types of attacks. However, the watermark

embedded in this method is of size 21x21, and this

watermark is considered very small to be embedded in the

image.

P. Kumhom [11] proposed a new technique based on a

selection of a high frequency range containing large

amount of information. The selected high-frequency range

contributes to the imperceptibility of the watermark while

the robustness against compression is achieved because the

selected frequency range contains large amount of

information. The entropy-based algorithm is adopted to

find the best tree of the wavelet packet transform (WPT).

Such best tree represents the best basis of the WPT whose

corresponding frequency sub-band contains high

information energy.

S. Zhenghua and others [12] proposed a technique

based on a selection of a high frequency range containing

large amount of information. The entropy-based algorithm

is used to find the best tree of the wavelet-based contourlet

packet transform (WCPT). Such best tree represents the

best basis of the WCPT whose corresponding frequency

sub-band contains high information energy. Moreover, the

security of the watermark is added by the pseudorandom

permutation of the bits of the watermark image before

embedding.

III. Discrete Wavelet Transform (DWT)

Wavelet Transform has been used widely since it has

been adopted in the established image coding standard

JPEG 2000 [2], and it produce considerably better quality

for decoded image than JPEG. The main advantage that

DWT has over Fourier transforms is temporal resolution. It

captures both location and frequency information .The

basic idea of DWT is to separate frequency detail, which is

multi-resolution decomposition. One time of

decomposition can divide the image to four sub images at

Proc. of the International Conference on Computer & Communication Engineering 2014 (ICCCE 2014)

23-25 September 2014, Kuala Lumpur, Malaysia

a quarter sizes. They are a low frequency approximate sub

image, and three horizontal, vertical, and diagonal

direction high frequency details sub images [13].

Fig. 3. two-level DWT

Fig. 3 is an example diagram of two-level DWT

decomposition. In the wavelet transform domain, high

frequency parts represent detailed information of image's

edge, contour and texture and so on. It’s not easy to detect

the watermark in these places as people are not easily able

to recognize it. But after processing or attacking, it doesn’t

have good stability. Most energy of image is centralized in

low frequency. Low frequency coefficients are nearly not

changed by common attacks so that watermarking

information embedded in low frequency coefficients has

better robustness [13].

IV. Watermark Embedding and Extraction Method

A- Watermark Embedding Algorithm

1- Input: A Grayscale cover image (I), and a binary

watermark (W).

Output: Watermarked Image.

2- Decompose the cover image in 2-level discrete

wavelet transform (DWT) of Host Image (I).

3- D2 is considered for watermark embedding.

4- Convert the 2D watermark into 1D array.

5- Select a Seed of type RNG to generate a PN sequence

PNS of size equal to the size of selected matrix D2

6- Generate another matrix PN from PNS according to

the relation PN = R1 × (PNS − R2), where, R1 = 2

and R2 = 0.5 to change the PNS from 0s and 1s to 1s

and -1s.

7- If watermark bit is 0 (Black) then modify elements of

the selected matrix as, D2 = D2 + k.PN, Where, k is

embedding strength. If watermark bit is 1 (white) then

wavelet coefficients will not be changed.

8- Repeat the step 6 and 7 for all ‘0’ watermark bits with

every time newly generated PN sequence.

9- Put the modified D2 matrix in its original position in

wavelet tree and take inverse discrete wavelet

transform IDWT to get back the watermarked image

(Iw).

The embedding algorithm is shown graphically in Fig. 4.

Fig. 4. Watermark Embedding Algorithm

B- Watermark Extraction

1- Input: Watermarked Image (𝐼𝑤).

2- Output: Extracted Watermark (𝑊𝑅).

3- Decompose the watermarked image (𝐼𝑤) in 2-level

DWPT and get back the same modified coefficient

metrix D2′.

4- Generate the same PN sequence (𝑃𝑁𝑆), which was

generated in embedding process using the same Seed

of type RNG with same value.

A2 H2

V2 D2

V1 D1

H1

Cover Image

(I)

2-Level

DWT

Seed

PNS

Watermark Embedding

cD2,3 = cD2,3 + k.PN

Watermark

(W)

Convert W

to vector

PN = R1*(PNS-R2)

2-level

IDWT

Watermarked

Image

Proc. of the International Conference on Computer & Communication Engineering 2014 (ICCCE 2014)

23-25 September 2014, Kuala Lumpur, Malaysia

5- Convert the PN sequence (𝑃𝑁𝑆) into 𝑃𝑁 = 𝑅1× (𝑃𝑁𝑆

− 𝑅2) with 𝑅1 = 2 𝑎𝑛𝑑 𝑅2 = 0.5.

6- Compute the correlation coefficients between PN

sequence (𝑃𝑁) and modified coefficient matrix D2′, as

𝑟 = 𝑐𝑜𝑟𝑟2(𝑃𝑁, 𝐷2′) .

7- Repeat step 6 for all watermark bits and compute the

correlation values.

8- Compute the threshold value as (𝑇=1.4 * 𝑚𝑒𝑎𝑛 (𝑟))

and initialize a row vector (𝑊′) having all values ‘1’

equivalent to the size of the watermark.

9- For every watermark bit, compare 𝑟 with 𝑇, then

𝑊′=0 if r > T, and 𝑊′=1 if r < T.

10- Reshape the row matrix 𝑊′ into a matrix equivalent to

the size of original watermark matrix to get recovered

watermark.

Fig. 5. Watermark Extraction Algorithm

V. EXPERIMENTAL RESULTS This section discusses the results of the proposed

algorithm. Five images of size 512 x 512 were used as

cover images, which are Bridge, Peppers, Cameraman,

Boat, and Houses. They are shown in Fig. 6.

a b

c d

e

Fig. 6. Cover images

a: Bridge, b: Peppers, c: Cameraman, d: Boat, e: Houses

Two binary images of size 16 x16 and 32 x 32 were

used as watermarks. They are shown in Fig. 7.

a b

Fig. 7. Watermark images: a: size 16 x16, b: size 32 x32

If r > T

Watermarked

Image Seed

2-Level DWT PNS

PN = R1*(PNS-R2)

Compute Correlation

𝑟 = 𝑐𝑜𝑟𝑟2 (𝑃𝑁, 𝐷2′)

T = 1.4 * mean(r)

Extracted Watermark

W’ = 0 W’ = 1

Yes No

Proc. of the International Conference on Computer & Communication Engineering 2014 (ICCCE 2014)

23-25 September 2014, Kuala Lumpur, Malaysia

The performance of the proposed method is evaluated

in term of imperceptibility and robustness. The

imperceptibility is measured by determining the difference

between the original image and the watermarked image.

Peak Signal to Noise Ratio (PSNR) is commonly used

measure for imperceptibility level. The PSNR is defined

as:

𝑃𝑆𝑁𝑅 (

𝑆 )

For two images having 8 bits per pixel, The PSNR is

defined as:

𝑃𝑆𝑁𝑅 𝑜 (

𝑆 )

Where MSE is Mean Squared Error and it’s defined as:

𝑆

𝑁∑ 𝐼 [ ] 𝐼 [ ]

Where:

N: number of pixels in each image

: Original image

: Watermarked image

The PSNR which is usually expressed in decibel scale

(dB) represents the ratio between the maximum pixel

value of an image and the corrupting noise (defined via

MSE) that affect its representation. The Typical range for

PSNR values is between 30 and 50 dB, considering that

higher value is better where the differences are between

the two images are less and indistinguishable by human

eyes [14].

Table 1: PSNR values for k = 1, 1.5, 2

Image Watermark

size

PSNR (dB)

k = 1 k = 1.5 k = 2

Bridge 16x16 44.066 40.594 38.154

32x32 38.257 34.874 32.512

Peppers 16x16 43.069 39.583 37.133

32x32 37.104 33.699 31.352

Cameraman 16x16 44.091 40.636 38.204

32x32 38.207 34.800 32.415

Boat 16x16 43.973 40.459 37.969

32x32 38.165 34.853 32.560

Houses 16x16 44.141 40.711 38.302

32x32 38.289 34.922 32.575

The robustness is measured by performing some

attacks such as Disk filter, Average filter, and Gaussian

noise. Then Normalized Correlation (NC) between the

embedded and extracted watermarks is calculated for

embedding strength (k = 1, 1.5, 2).

The NC is defined as following:

𝑁 ∑ ∑ 𝑊 𝑊

√∑ ∑ 𝑊 √∑ ∑ 𝑊

Table 2: Normalized Correlation (NC) values under Disk

Filter

Disk Filter Watermark

size

NC

k = 1 k = 1.5 k = 2

Bridge 16x16 0.831 0.948 1.000

32x32 0.835 0.948 0.975

Peppers 16x16 1.000 1.000 1.000

32x32 0.964 0.993 1.000

Cameraman 16x16 0.939 0.982 1.000

32x32 0.936 0.980 0.991

Boat 16x16 0.923 0.991 1.000

32x32 0.911 0.965 0.978

Houses 16x16 0.789 0.925 0.991

32x32 0.768 0.905 0.971

Table 3: Normalized Correlation (NC) values under

Average Filter

Average

Filter

Watermark

size

NC

k = 1 k = 1.5 k = 2

Bridge 16x16 0.713 0.850 0.941

32x32 0.689 0.858 0.927

Peppers 16x16 0.965 0.991 1.000

32x32 0.926 0.975 0.991

Cameraman 16x16 0.865 0.947 1.000

32x32 0.872 0.956 0.982

Boat 16x16 0.875 0.982 1.000

32x32 0.844 0.938 0.975

Houses 16x16 0.753 0.871 0.949

32x32 0.624 0.802 0.888

Table 4: Normalized Correlation (NC) values under

Gaussian Noise

Gaussian Watermark

size

NC

k = 1 k = 1.5 k = 2

Bridge 16x16 0.882 0.982 1.000

32x32 0.910 0.969 0.995

Peppers 16x16 1.000 1.000 1.000

32x32 0.991 1.000 1.000

Cameraman 16x16 0.982 1.000 1.000

32x32 0.971 0.991 1.000

Boat 16x16 0.982 1.000 1.000

32x32 0.956 0.978 0.991

Houses 16x16 0.885 0.991 1.000

32x32 0.858 0.954 0.980

Proc. of the International Conference on Computer & Communication Engineering 2014 (ICCCE 2014)

23-25 September 2014, Kuala Lumpur, Malaysia

Conclusion In this paper, a robust digital image watermarking

scheme by using discrete wavelet transform is proposed.

The watermark is embedded in 2-level diagonal coefficient

to achieve high performance in term of imperceptibility.

Robustness is also evaluated by performing some attacks

such as Disk filter, Average filter, and Gaussian noise. The

experimental results show that the proposed algorithm

provides good level in term of imperceptibility and

robustness. Two different sizes of watermarks were used

16 x 16 and 32 x 32 to show the capacity level of the

proposed algorithm.

References

[1] M. L. M. Ingemar J. Cox, Jeffrey A. Bloom, Jessica Fridrich, and

Ton Kalker, Digital Watermarking and Steganography: Morgan Kaufmann Publishers, 2008.

[2] Y. Yusof and O. O. Khalifa, "Digital watermarking for digital

images using wavelet transform," in Telecommunications and Malaysia International Conference on Communications, 2007.

ICT-MICC 2007. IEEE International Conference on, 2007, pp.

665-669. [3] Z. Yanqun, "Digital Watermarking Technology: A Review," in

Future Computer and Communication, 2009. FCC '09.

International Conference on, 2009, pp. 250-252. [4] R. F. Olanrewaju, "Development of Intelligent Digital

Watermarking via Safe Region," PHD, Electrical and Computer

Engineering, International Islamic University Malaysia, Kulliyyah of Engineering, 2011.

[5] J.-S. Pan, H.-C. Huang, and l. C. Jain, Eds., Intelligent

Watermarking Techniques (Series on Innovative Intelligence. World Scientific, 2004, p.^pp. Pages.

[6] L. Jian and H. Xiangjian, "A Review Study on Digital

Watermarking," in Information and Communication Technologies, 2005. ICICT 2005. First International Conference on, 2005, pp.

337-341.

[7] N. Cvejic, "Algorithms for Audio Watermarking and Steganography," Department of Electrical and Information

Engineering, University of Oulu, 2004.

[8] A. G. Charles Fung, Walter Godoy Junior, "A Review Study on Image Digital Watermarking," presented at the The Tenth

International Conference on Networks, St. Maarten, The

Netherlands Antilles, 2011. [9] S. Jun and M. S. Alam, "Fragility and Robustness of Binary-Phase-

Only-Filter-Based Fragile/Semifragile Digital Image

Watermarking," Instrumentation and Measurement, IEEE Transactions on, vol. 57, pp. 595-606, 2008.

[10] C. Patvardhan, A. K. Verma, and C. V. Lakshmi, "A Robust

Wavelet Packet Based Blind Digital Image Watermarking using HVS characteristics," International Journal of Computer

Applications vol. 36 - Number 9, p. 7, 2011.

[11] P. Kumhom, S. On-rit, and K. Chamnongthai, "Image Watermarking Based On Wavelet Packet Transform With Best

Tree," presented at the ECTI Transactions on Electrical

Engineering, Electronics, and Communications (ECTI-EEC), 2004.

[12] S. Zhenghua, L. Guodong, G. Lixin, Z. Lin, and F. Wenming,

"Image Watermarking Based on Wavelet-Based Contourlet Packet Transform with Best Tree," in Electronic Commerce and Security,

2009. ISECS '09. Second International Symposium on, 2009, pp.

203-207. [13] R. Dubolia, R. Singh, S. S. Bhadoria, and R. Gupta, "Digital Image

Watermarking by Using Discrete Wavelet Transform and Discrete Cosine Transform and Comparison Based on PSNR," in

Communication Systems and Network Technologies (CSNT), 2011

International Conference on, 2011, pp. 593-596. [14] Y. B. Yusof, "Improved Digital Image Watermarking Using

Discrete Wavelet Transform Technique," Master, International

Islamic University Malaysia, 2009.