A Removable Visible Watermark for Digital Imagescsroc.org.tw/journal/Published...

A Removable Visible Watermark for Digital Images

Chin-Chen Chang1, 3, Chia-Chen Lin2, ＊, and Kuan-Ming Li3

1 Department of Information Engineering and Computer Science,

Feng Chia University,

Taichung 40724, Taiwan, ROC

[email protected]

2 Department of Computer Science and Information Management,

Providence University,

Taichung 43301, Taiwan, ROC

[email protected]

3 Department of Computer Science and Information Engineering,

National Chung Cheng University,

Chiayi 621, Taiwan, ROC

[email protected]

Received 5 July 2010; Revised 5 August 2010; Accepted 10 September 2010

Abstract. Visible watermarking schemes are among the techniques being used to protect intellectual property rights (IPRs). To provide authorized users better visibility, a new research issue called removable visible wa-termarking has been proposed recently. In this paper, an enhancement of Huang and Tang’s scheme is pro-posed to achieve removability. Authorized users can simply remove an embedded watermark by using the re-ceived secret stream to reconstruct the original image. However, malicious users can only obtain recon-structed images by using blind guessing, and those will be of poor image quality. The experimental results confirm that the difference in image quality generated by unauthorized and authorized users can range up to 19 dB. Furthermore, the secret stream for reconstructing the original image in our scheme can be shortened to 192 bytes by using our proposed secret stream shortening algorithm.

Keywords: Removable visible watermark, vector quantization

1 Introduction

Owing to the progress in information technologies and the growth of the Internet, vast amounts of data such as texts and images have been digitized for easy storage, processing and transmission over the Internet. Although the Internet allows senders and receivers to transmit data easily without geographical limitations, it also opens the transmitted data to easy attack by malicious users. That is, malicious users can tamper with and grab trans-mitted data from the Internet illegally, manipulate the data and then claim ownership of the manipulated data. In such case, the original ownership is violated and creation is discouraged.

To prevent such behavior and protect intellectual property rights (IPRs), two approaches have been proposed over the past decade. One is legislation that forces violators to pay stiffer penalties for illegal cribs and manipu-lations. The other one is based on information technologies such as watermarking, copy detection or digital signatures. By using watermarking, the original owners of digital media can embed their own logos, such as portraits or trademarks or even secret information into their work. Later, the embedded logos or secret informa-tion can be easily extracted by the real owners or authorized users who have the necessary confidential data, also called keys, to prove ownership.

To ensure the owner’s rights, most watermarking schemes must satisfy the following criteria: The watermark can only be extracted by the original owner. The watermarked image must withstand various attacks that contain A/D transfer, filtering, compression,

rotation, and the like.

＊Correspondence author

Journal of Computers Vol.21, No.3, October 2010

38

According to the degree to which watermarking can withstand various attacks, scholars classify watermarking schemes into one of these three categories.

Robust watermarking schemes: Embedded watermarks for copyright protection are expected to survive different types of manipulations to some extent, provided that the manipulated media are still valuable in terms of commercial importance or significant in terms of visual quality.

Fragile watermarking schemes: Embedded watermarks are used for authentication and content integrity verification. A tiny manipulation will cause significant damage to the embedded watermark to demon-strate that the integrity of digital media has been violated. That is, fragile watermarking schemes focus on sensitivity to attacks or even incidental manipulation in some cases [1].

Semi-fragile watermarking schemes: Embedded watermarks are robust to mild modification such as JPEG compression and channel additive white Gaussian noise (AWGN), but fragile to malicious attacks [2].

Excepting these three classifications, watermarking schemes also can be categorized as visible watermarking and invisible watermarking according to the visibility of the embedded watermarks. It is obvious that in the former, anyone can see the embedded watermarks and ownership can be easily determined without requiring computation. On the contrast, invisible watermarking requires computations for a legal owner to extract the watermark and prove ownership. In essence, invisible watermarking schemes offer better protection of IPRs by taking advantage of their invisibility. The major reason is attackers cannot remove invisible watermarks without detecting them in advance. Certainly, each scheme offers different levels of robustness and visibility for its em-bedded watermark. However, invisibility and robustness influence each other, as shown in Figure 1.

Robustness

Invisibility

Fig. 1. The curve of the invisibility and robustness factors of invisible watermarking schemes [3]

Although visible watermarks make it easy for authorized owners to claim copyright protection or ownership of digital media without computation, visible watermarks may influence the image quality of digital media, especially if the embedded watermarks cross the entire digital media. Because of these weaknesses, only a few visible watermarking schemes have been proposed [4]. In general, two approaches have been adopted for de-signing visible watermarks. One is to hide watermarks in the spatial domain, such as hiding watermarks in the least significant bits (LSBs) in an image, and the other is to embed watermarks in the frequency domain of an image. The latter method is more robust than the former because it can withstand the mild compression involved in saving storage space better than the former. Recently, a new branch of visible watermarks, called removable visible watermarking, has been proposed. The motivation behind this new technology is to maintain the image quality of digital media and allow authorized owners to remove embedded watermarks by using secret keys purchased from sellers and then reconstructing the original images once they need to print out the digital media.

Basically, the reversible visible watermark concept was proposed by IBM in 1997 [5], and several reversible invisible schemes based on this concept have since been proposed [6]. Recent work in reversible invisible wa-termarks was proposed by Alattar based on Tain’s pixel difference expansion [6]. In essence, those algorithms designed for invisible watermarks are hardly extended and applied to visible watermarking schemes because they use different perceptual models and embedding strategies to meet different requirements for robustness. Recently, user-key–dependent removable visible watermarking was proposed by Hu et al. [4]. In their scheme, if an on-line shop wants to sell an image to three different customers, a common watermark will be embedded into three images, respectively, by using different user keys, which are provided by different customers. In other words, three customers can receive the same watermarked image and the embedded watermarks will be the same, although the user keys used to generate an embedding template in the embedding process will be different. In other words, in their scheme, the user key can decide the embedded subset of the watermark and calculate cer-tain embedding parameters. With a correct user key, an owner can remove the embedded watermark and restore the high-quality unmarked image. Based on their experimental results, on the average, a legal user can restore an

Chang et al: A Removable Visible Watermark for Digital Images

39

original image with an image quality of 44.022 dB, while an illegal user can achieve an image quality of only 37.984 dB. Although there is a significant difference between the two images generated by a legal user and an illegal user, an illegal user still can get image quality up to 37 dB through blind guessing. For the human vision system, an image quality of 37 dB will not show a significant difference, which means illegal users still can achieve sufficiently good image quality through blind guessing. In this paper, we try to extend the visible wa-termarking scheme proposed by Huang and Tang [3] to propose a removable visible watermarking scheme in which the image quality derived by an illegal user is decreased and the copyright of authorized users is guaran-teed.

The rest of this paper is organized as follows. In Section 2, we briefly describe Huang and Tang’s contrast-sensitive visible watermarking scheme. The proposed reversible embedding scheme is presented in Section 3. Several experimental results are illustrated and discussed in Section 4. Finally, concluding remarks as well as future work are stated in Section 5.

2 Review of the Contrast-sensitive Visible Watermarking Scheme

In 2006, Huang and Tang proposed a visible watermarking scheme that varies the strength of a watermark in different blocks of the image according to the human visual system (HVS) and the wavelet coefficient contrast-sensitive function (CSF) [3]. By using the embedding strategy described above, they successfully made the embedded watermark visible without significantly obscuring the host image. Figure 2 is a flowchart illustrating Huang and Tang’s scheme.

Fig. 2. Flowchart of Huang and Tang’s scheme

As Figure 2 shows, the original image and the watermark are preceded by a 5-level wavelet transform to gen-erate the discrete wavelet transform (DWT) coefficients separately. Later, a DWT block of the original image is classified into one of the three classes: plane, edge and texture blocks, according to the entropy Hx and variance

2xσ of the DWT block to distortions in order to maintain better image quality in the watermarked image.

1

* ( )n

i ii

x x p x=

= ∑ , (1)

2 2

1

( ) * ( )n

x i ii

x x p xσ=

= −∑ , (2)

1

( ) * log[ ( )]n

x i ii

H p x p x=

= −∑ , (3)

where xi is the DWT coefficient and x is the mean value. Two predetermined thresholds, 1T and 2T , are used

to decide to which DWT block oA belongs. Basically, if H( oA )< 1T , then block oA is determined to be a plane

block because the entropy is lower for a plane block. If H( oA )> 1T and 22hA Tσ < , then block oA is determined to

be a textured block. If a DWT block does not belong to either of these two cases, it is judged an edge block. Later, Huang and Tang modified the DWT coefficients of the original image to hide a watermark by using

Equation (4).


40

( ) ( ) ( ) ( ) ( )( , | ) ( , | ) ( , | )k k k k kc o wy i j p y i j p y i j pα β= × + (4)

Here, ( ) ( , | )kcy i j p is the composite coefficient of the visible watermarked image, and ( ) ( , | )k

oy i j p and ( ) ( , | )kwy i j p are the decomposed wavelet coefficients of the original image and the watermark, respectively. In

addition, k={1,2,3,4,5} indicates the image orientation, p={1,2,3,4} indicates the decomposed levels, and ( i , j ) represents the spatial location. Meanwhile, ( )kα and ( )kβ are the scaling and embedding factors, which are de-

fined as ( )kα and ( )kβ as follows.

2

2

(7.20 )1

7.20

kk

i

rDα

−= − + (5)

2

2

(7.20 )0.01 2

7.20

kk

i

rDβ

−= + + × (6)

Here, kr is the wavelet coefficient CSF of the perceptual importance weight, as shown in Figure 3. iD is the

distortion index, where {0, 0.01, 0.02}iD ∈ .

Fig. 3. Discrete wavelet transform contrast-sensitive function (CSF) mask with 10 unique digits

Because Huang and Tang’s scheme carefully considers the characteristics of blocks in an original image as the watermark is embedded in it, the watermarked image maintains good image quality as judged by human eyes. Moreover, the visibility of the embedded watermark is better than can be achieved with Hu and Kwong’s scheme [7]. Understanding that authorized owners may want to remove the embedded watermark for later use, we extended Huang and Tang’s scheme to propose a removable visible watermarking scheme in this paper to fit potential commerce usage. Detailed descriptions of our proposed scheme are given in the next section.

3 The Proposed Scheme

The removable visible watermarking scheme proposed here is a special variant of visible watermarking that not only allows an original image to contain a visible watermark for copyright protection purposes but also allows an authorized owner to remove an embedded watermark under certain requirements. Such an application bene-fits authorized owners by allowing them to clearly view the content of the watermarked image without interfer-ence by the embedded watermark. Because the visibility of the embedded watermark proposed by Huang and Tang’s scheme is better than with other schemes, and because the watermarked image generated by their scheme can also be clearly seen by the human eye, we then expand on Huang and Tang’s scheme to achieve removabil-ity. Following Hu et al.’s concept [4], a secret stream is generated by the on-line shop for the authorized user after purchase. Later, the authorized consumer can use the secret stream issued to remove the embedded water-mark and restore the original image if the authorized owner does not want the embedded watermark to interfere. In our scheme, the image quality of the original images reconstructed by authorized users can maintain good image quality. However, unauthorized users will not be able to achieve image quality as high as that of Hu et


41

al.’s scheme. The details of watermark embedding and removing processes are described in the following sub-sections.

3.1 Visible Watermark Embedding

In our scheme, a watermark (W) is embedded into an original image (O) and both are M N× in size. Figure 4 illustrates the flowchart for the watermark embedding procedure in our scheme. The upper part of the flowchart shows that the visible watermarked image is a combination of the watermark (W) and the original image (O). The lower part of the flowchart illustrates that the watermarked block is the input to generate the secret stream that enables the authorized user to remove the embedded watermark and reconstruct the original image.

Fig. 4. Flowchart of watermark embedding procedure

In essence, our watermark embedding procedure is similar to Huang and Tang’s. The original image and wa-termark perform a 5-level wavelet transform operation first. Next, each DWT block of the original image is

classified into one of plane, edge or texture blocks, according to the entropy Hx and variance 2xσ of the DWT

block. Both Hx and 2xσ can be obtained by Equations (2) and (3), which are defined in Section 2. At last, the

DWT coefficients of the original image are modified to hide a watermark as shown in Equation (4), which is defined in Section 2. However, to make sure the watermarked image can preserve better image quality, we modified Huang and Tang’s scaling and embedding factors, ( )kα and ( )kβ , as Equations (7) and (8), respectively. The significant difference between Equations (5) and (6) in Huang and Tang’s scheme and Equations (7) and (8) in our scheme are their coefficients of the distortion index Di. We changed the coefficients of the distortion index Di to those in Equations (7) and (8) to increase the difference in image quality that can be derived by au-thorized users and by malicious users.

2

2

(7.20 )1 10

7.20

kk

i

rDα

−= − + × (7)

2

2

(7.20 )0.01 8

7.20

kk

i

rDβ

−= + − × (8)

Here, kr is the wavelet coefficient CSF of the perceptual importance weight, the largest value is 7.20 and the smallest value is 1.00, as shown in Figure 3. iD is the distortion index, where {0, 0.01, 0.02}iD ∈ .

Although the difference in image quality obtained by authorized users and malicious users can be enlarged by the modifying the coefficient of the distortion index Di, we still want to maintain the image quality of the wa-termarked image. Therefore, we try to achieve our objective by coefficient classification in the DWT block. Generally, each DWT block is further classified into three frequency bands, low-, medium-, and high-frequency, to judge the distortion index iD , as shown in Figure 5. In this figure, the gray area is high-frequency, the white

area is medium-frequency and the blue area is low-frequency. In other words, each DWT block belongs to one


42

of three frequency bands. Later, the maximum and minimum bounds for a DWT block can be determined by sorting its absolute values for all coefficients. The interval between the maximum and minimum bounds is then divided to three regions mapped to P3, P2 and P1, respectively, from the largest absolute coefficients to the smallest.

Fig. 5. Example of three frequency bands in a DWT image

Such an arrangement is necessary because the importance of a coefficient is positive to its absolute value. Therefore, the modifications to coefficients with large absolute values caused during the watermark embedding procedure should be the lowest possible so that the image quality of the watermarked image can be preserved.

Based on experimental results, we set the corresponding Di of 1P as 0.01 and the Di’s of 2 3,P P as 0 in the low-

frequency band. In the high-frequency band, the relative Di of 1P is 0.02 and the Di’s of 2 3,P P are 0.01. The

corresponding Di’s of 1P , 2P and 3P are 0.02, 0.01 and 0, respectively, in the medium-frequency band accord-

ing to our experimental data. Based on this arrangement, each ( )kα and ( )kβ for coefficients in the original im-age and watermark can be calculated and the visible watermarked image can be calculated as well. Furthermore, the factor ( )kα is restricted within the interval [0.86, 1] and the factor ( )kβ is restricted within the interval [0.01,0.14]. Both intervals were also derived from experiments.

Fig. 6. Example of watermark embedding procedure

Figure 6 presents the DWT coefficients of the original image and a watermark on the left and medium side, respectively. The DWT coefficients of the original image and the watermark are multiplied by individual ( )kα

and ( )kβ , respectively. Then, the two modified DWT coefficients are added to generate a DWT coefficient of the watermarked image. Later, the DWT coefficients of the watermarked image are inversed into a visible water-marked image by an inverse discrete wavelet transform (IDWT) operation.

However, there is still a problem. That is, each DWT coefficient of the original image is mapped to a distor-tion index iD , which is required to remove the embedded watermark and restore the original image. Because the

distortion index iD is determined by the classification of each original DWT block in an original block and such

information cannot be obtained from the watermarked image, these distortion indices must be recorded in ad-vance for use later in the extraction process. In other words, the distortion indices are a secret stream for an


43

original image, and only the authorized user can own it and use it for later watermark extraction and original image restoration. In essence, each distortion index iD requires two bits for representation.

Therefore, the length of the secret stream is 4

M N× bytes. However, this stream is too long to serve as a key

for removing an embedded watermark and restoring an original image. In this paper, a Linde-Buzo-Gray (LBG) algorithm [8, 9] designed for vector quantization (VQ) [10] is applied to shorten the stream. The lower part of Figure 7 demonstrates the process for generating the secret stream. Because a codebook is required in our pro-posed process to shorten the secret stream, the algorithms for generating the codebook and secret stream short-ening are described below, in order.

Codebook Generation Algorithm

Step 1: Five training images are used to generate a codebook. Each training image performs a 5-level wavelet transform, individually.

Step 2: Each DWT image is divided into nonoverlapping blocks w w× pixels in size. Each block is treated as an initial vector.

Step 3: K vectors are randomly chosen from these blocks to form an initial codebook CB. Step 3.1 Each residual block is calculated for its Euclidean distance with each of the K initial vectors. Step 3.2 Residual blocks are grouped into K initial vectors as K groups, respectively, with the minimum Euclid-

ean distance linked to the corresponding initial vector. Step 3.3 The new centroid of each group is computed to obtain K codewords. These K codewords form a new

codebook CB. Step 3.4 Steps 3.1 to 3.3 are repeated until the variation of each centroid of each group is converged. Once all

centroids are converged, the codebook CB is outputted for later secret stream generation.

Secret Stream Shortening Algorithm

Step 1: The watermarked image, which has been performed by DWT, is divided to nonoverlapping blocks ww× pixels in size.

Step 2: For each watermarked DWT block, a look-up operation is performed to find a codeword from code-book CB with the minimum absolute value of distortion to the block. The distortion measure is defined in Equation (9).

2( )jiiD X X∧

= − , where ( )jiiX X∧

− >0

( )iiD X X∧

= − , otherwise. (9)

Here, Xi is the coefficient of the watermarked DWT block and ˆjiX is the ith element of jth codeword

for 0 ( 1)j K≤ ≥ − . Step 3: Output the corresponding index of codeword found in Step 2. Step 4: Steps 2 and 3 are repeated until all watermarked DWT blocks have been processed.

The indices generated by the secret stream shortening algorithm serves as the secret key to be given to the au-thorized user instead of the original distortion indices. Typically, our secret stream shortening algorithm can

shorten a secret stream to log2

8

KM N

w w

× ×

× × bytes, where M and N are the width and height of the original image,

respectively. K is the number of codewords in codebook CB, and w w× is the dimension of each codeword. The length of the secret stream is decreased greatly with our proposed shortening algorithm.

3.2 Visible Watermark Removing

The removing of an embedded watermark requires the original watermark, the codebook CB, the DWT CSF mask shown in Figure 3 and the secret stream. Only the secret stream is in the care of the authorized user; other components are obtained from the on-line shop. Figure 7 presents the flowchart for removing an embedded watermark.


44

Fig. 7. Flowchart for removing an embedded watermark

First, the watermarked image and watermark perform a 5-level DWT transformation, respectively. Later, the user uses his or her secret stream and the received codebook CB to restore the distortion index iD for each

DWT watermarked block. Furthermore, Equations (7) and (8), which are given in Subsection 3.1, are calculated to obtain the scaling ( )kα and embedding ( )kβ factors, respectively. Finally, the original image can be recon-

structed and the embedded watermark W−

can be removed by Equation (10)

( ) ( ) ( )( )

( )

( , | ) ( , | )( , | )

k k kk c w

o k

y i j p y i j py i j p

β

α

− ×= , (10)

where ( ) ( , | )kcy i j p is the composite coefficient of the visible watermarked image, and ( ) ( , | )k

oy i j p and ( ) ( , | )kwy i j p are the decomposed wavelet coefficients of the original image and the watermark, respectively.

Fig. 8. Example of removing an embedded watermark and restoring an original image

As shown in Figure 8, by inversing the embedding procedure, the DWT coefficients of an original image can be generated based on the combination of watermarked image, watermark, secret stream and CSF mask. Notably, the generated DWT coefficients are exactly the same as those shown in Figure 6. The original image, which was successfully restored by IDWT on these DWT coefficients, can be used for other purposes.


45

4 Experimental Results

The visible watermarking scheme was coded by using Borland C++ Builder and was tested on a Microsoft Win-dows XP platform with Pentium M and 1.73 GHz. The codebook CB is trained by five gray-scale test images, “Boat”, “Lena”, “Jet”, “Toys”, and “SailBoat”, using an LBG algorithm. Each test image is 256×256 pixels and is divided into 20480 nonoverlapping blocks, each 4×4 pixels in size. One logo watermark, which was used in the following experiments, is shown in Figure 9. Three gray-scale 256×256 test images, “Baboon”, “Lena”, and “Pepper” served as host images and are shown in Figure 10.

Fig. 9. Logo used as a watermark 256×256 pixels in size

Fig. 10. Three gray-scale test images 256×256 pixels in size

In the following experiments, peak signal-to-noise ratio (PSNR) is used to measure the quality of the water-marked image and the restored image. The PSNR is defined as follows:

2

10

25510 logPSNR dB

MSE= , (11)

where 255 represents the maximum value of a pixel and the mean square error (MSE) of two M×N gray-scale images is defined as follows:

1 12

0 0

1( ) ( )

M N

ij iji j

MSE y yM N

− −

= =

= −×

∑∑ (12)

Here, ijy represents the value of the original image in the (i, j) location and ijy represents the modified value

in the (i, j) location. A larger PSNR implies that the difference between the original image and the modified one is small, and vice versa. In addition, once the PSNR is equal to or greater than 30 dB, the human visual system is not sensitive to the distortion caused by some modifications.

Because the codebook concept is used to shorten the secret stream, in our first experiment we explored the in-fluence of codebook size on the image quality of a watermarked image. Six codebooks with different combina-

(a) Baboon (b) Lena (c) Pepper


46

tions of codebook size and codeword dimensions were tested: ( , )iCB size dim for 1 6i≤ ≤ , 256, 128, 64{ }size∈

and {16, 64, 256}dim∈ . Table 1 presents the PSNRs for the watermarked image and restored image with all six codebooks. Figures 11 and 12 demonstrate the image quality of watermarked and restored images with code-book 1 (256, 16)CB and the image quality of watermarked and restored images with codebook 6 (64, 256)CB ,

respectively. Notably, no matter which codebook is adopted for watermark embedding, the restored images are nearly the same, and all are up to 46 dB. The tiny difference is mainly caused by the round-off operation during the image reconstruction process. Although the PSNRs of the watermarked images are smaller than 30 dB, the embedded watermarks are still visible, as shown in Figures 11 and 12. Even though codebook CB6 contains only 64 codewords and each codeword has 256 elements, the visibility of the embedded watermarks is exactly as clear as those generated by codebook CB1, which has 256 16-dimension codewords.

Table 1. The PSNRs of watermarked and restored images with six codebooks.

Watermarked images Restored images Images

Codebook Baboon Lena Pepper Baboon Lena Pepper

1 (256, 16)CB 24.9172 23.3708 23.7481 46.0209 46.1301 46.1121

2 (128, 16)CB 24.8883 23.3794 23.7502 46.0476 46.1327 46.1033

3 (64, 16)CB 24.9524 23.2904 23.6055 46.023 46.1286 46.1017

4 (128, 64)CB 24.9351 23.3823 23.7244 46.0382 46.1574 46.1193

5 (64, 64)CB 24.9351 23.3823 23.7244 46.0382 46.1574 46.1193

6 (64, 256)CB 25.1558 23.3597 23.9046 46.1125 46.1884 46.2033

The secret stream lengths generated by the six codebooks are listed in Table 2. As the table shows, a secret

stream can be shortened to 192 bytes if CB6 is used during the procedure for embedding watermarks.

Table 2. The size (bytes) of secret stream generated by the six codebooks.

Original images Codebooks

Baboon Lena Pepper

1 (256, 16)CB 4096 4096 4096

2 (128, 16)CB 3584 3584 3584

3 (64, 16)CB 3072 3072 3072

4 (128, 64)CB 896 896 896

5 (64, 64)CB 768 768 768

6 (64, 256)CB 192 192 192

In our second experiment, we explored whether the proposed scheme could enlarge the difference between

the restored images generated by authorized and unauthorized users. In essence, unauthorized users cannot ac-cess a secret stream and remove an embedded watermark through blind guessing. Figures 13 and 14 shows that randomly assigning a distortion index for each DWT watermarked block results in the image quality related to the restored images under CB1, CB2, CB5 and CB6, respectively.


47

(a) Watermark image: Baboon,

PSNR:23.370 dB (b)Watermark image: Lena

PSNR:24.917 dB (c)Watermark image: Pepper

PSNR:23.748 dB

(d) Restored image: Baboon

PSNR:46.020 dB (e) Restored image: Lena

PSNR:46.130 dB (f) Restored image: Pepper

PSNR:46.112 dB

Fig. 11. Image quality of watermarked and restored images with codebook 1 (256, 16)CB

(a)Watermark image: Baboon

PSNR: 25.155 dB (b)Watermark image: Lena

PSNR: 23.359 dB (c) Watermark image: Pepper

PSNR: 23.907 dB

(d) Restored image: Baboon

PSNR: 46.112 dB (e) Restored image: Lena

PSNR: 46.188 dB (f) Restored image: Pepper

PSNR: 46.203 dB

Fig. 12. Image quality of watermarked and restored images with codebook 6 (64, 256)CB


48

(a) Attacked image : Baboon

PSNR:28.495 dB

(b) Attacked image : Lena

PSNR:26.610 dB

(c) Attacked image : Pepper

PSNR:26.909 dB

(d) Attacked image : Baboon

PSNR:28.000 dB

(e) Attacked image : Lena

PSNR:26.931 dB

(f) Attacked image : Pepper

PSNR:27.063 dB

Fig. 13. PSNRs of restored images by blind guessing under codebooks CB1 and CB2, respectively

(a) Attacked image : Baboon

PSNR:28.255 dB

(b) Attacked image : Lena

PSNR:26.767 dB

(c) Attacked image : Pepper

PSNR:26.935 dB

(d) Attacked image : Baboon

PSNR:28.614 dB

(e) Attacked image : Lena

PSNR:26.911 dB

(f) Attacked image : Pepper

PSNR:27.213 dB

Fig. 14. PSNRs of restored images by blind guessing under codebooks CB5 and CB6, respectively


49

Notably, even when using codebook CB6 (64, 256), restored images such as “Lena” and “Pepper” still retain partially embedded watermarks. Compared with the PSNRs presented in Table 1, which are for images restored by authorized users, the degree of distortion can be as large as 19 dB with codebook CB6 (64, 256). Certainly, the visibility of embedded watermarks in complex images such as “Baboon” is not as clear as the watermarks in smooth images such as “Lena” and “Pepper”. However, the experimental results confirm partial achievement of our objective. Making our proposed scheme work with complex images will be our future work.

5 Conclusions

In this paper, we enhance Huang and Tang’s scheme to achieve removability so that authorized users can re-move visible watermarking to gain better visibility without interference from embedded watermarks. Consider-ing that attackers may try to remove embedded watermarks through blind guessing, the proposed scheme care-fully modifies the parameters of our embedding function and designs a secret stream shortening algorithm. Therefore, the image quality of restored image obtained by both attackers and authorized users can be as much as 19 dB, which is significantly higher than that achieved by Hu et al.’s scheme [4]. Furthermore, except for very complex images, embedded watermarks still remain in the restored images generated by malicious attack-ers. As for authorized users, the PSNRs of restored images can be up to 46 dB with very good image quality.

Certainly, the proposed scheme does not work as well on very complex images as it does in other cases, and the size of the secret stream requires further shortening. However, the proposed scheme has achieved greater success in degrading the image quality of restored images that can be achieved by illegal users than has been achieved in existing work. Solving the weakness described above will be our future work.

References

[1] H. Lu, R. Shen, F. L. Chung, “Fragile Watermarking Scheme for Image Authentication,” Electronics Letters, Vol. 39,

No. 12, pp. 898-900, 2003.

[2] K. Ding, C. He, L. G. Jiang, H. X. Wang, “Wavelet-Based Semi-Fragile Watermarking with Tamper Detection,”

IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, Vol. E88-A, No. 3,

pp.787-790, 2005.

[3] B. B. Huang and S. X. Tang, “A Contrast-Sensitive Visible Watermarking Scheme,” IEEE Multimedia, Vol. 13, No. 2,

pp. 60-66, 2006.

[4] Y. Hu, S. Kwong, J. Huang, “An Algorithm for Removable Visible Watermarking,” IEEE Transactions on Circuits

and Systems for Video Technology, Vol. 16, No. 1, pp. 129-133, 2006.

[5] D-Lib Magazine, http://www.dlib.org/dlib/december97/ibm/l2lotspieth.html.

[6] A. M. Alattar, “Reversible Watermark Using the Difference Expansion of a Generalized Integer Transform,” IEEE

Transactions on Image Processing, Vol. 13, No. 8, pp. 1147-1156, 2004.

[7] Y. Hu and S. Kwong, “An Image Fusion-Based Visible Watermarking Algorithm,” Proceedings of IEEE International

Symposium on Circuits and Systems, Vol. 3, pp. 25-28, 2003.

[8] Y. C. Lin and S. C. Tai, “A Fast Linde-Buzo-Gray Algorithm in Image Vector Quantization,” IEEE Transactions on

Circuits and Systems II: Analog and Digital Signal Processing, Vol. 45, No. 3, pp. 432-435, 1998.

[9] Y. Linde, A. Buzo, R.M. Gray, “An Algorithm for Vector Quantization Design,” IEEE Transactions on Communica-

tions, Vol. 28, No. 1, pp. 84-95, 1980.

[10] N. M. Nasrabadi and R. A. King, “Image Coding Using Vector Quantization: a Review,” IEEE Transactions on Com-

munications, Vol. 36, No. 8, pp. 957-971, 1988.

Date post:	18-Aug-2020
Category:	Documents
Upload:	others
View:	1 times
Download:	0 times

A Removable Visible Watermark for Digital Imagescsroc.org.tw/journal/Published...

Documents