REVERSIBLE VISIBLE WATERMARK TECHNIQUE IN DCT COMPRESSED DOMAIN

Kazuki Minemura and KokSheik Wong

Faculty of Comp. Science & Info. Technology,University of Malaya, Malaysia.

kazuki.minemura@gmail.com, koksheik@um.edu.my

ABSTRACTMany watermark algorithms were proposed for protectingdigital media from unauthorized use. Visible watermarkscheme indicates the copyright of a digital media by em-bedding an inconspicuous but recognizable pattern into themedia. However, the embedding process often results in se-rious distortion in the protected content, and the embeddingprocess is not reversible. This paper proposes a reversiblevisible watermarking method for JPEG compressed image.A contrast adaptive strategy is proposed to maintain bothtransparency and recognizable sign. This method can alsoguarantee complete restoration and global protection. Exper-imental results show that the proposed method outperformsrelated methods in terms of quality preservation of the recon-structed image.

Index Terms— visible watermark, contrast-adaptive,transparency, JPEG, DCT

1. INTRODUCTION

Visible digital watermark techniques [1-8] have recently beenutilized to protect multimedia content from unauthorized use.Although they are applied in different applications and pos-sess the same importance, some of them are not completelyreversible, i.e., the original multimedia content cannot be re-constructed from its watermarked counterpart. This paperproposes a novel solution for visible watermark mechanismin JPEG image that is completely reversible.

In this paper, our research focus on digital image becauseit is one of the mostly considered type of media in our dailylife. For commercial consideration, valuable images are digi-tized and posted over the Internet by resource provider. Nat-urally, people can browse and download these media freely.Thus, protecting the digital media from illegal use is an im-portant issue. Visible watermark mechanisms [1][2][3] areproposed to indicate the copyright of digital image by embed-ding an inconspicuous but recognizable logo into the image.The embedded visible watermark, however, can seriously dis-tort the protected image. Thus, removable visible watermark(RVW) mechanisms are proposed to resolve this issue [4].In addition to the prevention of unauthorized use, removable

watermarking method allows authorized users to remove thevisible watermark and reconstruct an unmarked version ofthe image with acceptable quality. Hu and Jeon proposed animproved method in which the embedded visible watermark(i.e., logo) can be completely removed [5]. Nevertheless, thisvisible watermark can only protect a local area of the image.It fails to satisfy the global protection requirement of a visi-ble watermark mechanisms as stipulated in [6][7]. Mohantyproposed an adaptable visible watermarking method by clas-sifying block in DCT (discrete cosine transform ) domain [8].Although it can survive unintentional modifications or mali-cious attacks, the method is irreversible.

The main concerns in designing an RVW mechanism aretransparency and restoration [6][7]. Transparency focuses onhow the strength of the watermark influences the perceptualquality of the host image. To obtain good perceptual translu-cency from a smooth test image, we must lower the strengthof the watermark. On the other hand, to ensure visibility ofthe watermark in image of high spatial activity, we must in-crease the strength of the watermark, which may obscure theimage. Restoration addresses the quality of the reconstructedimage. Without loss of generality, the host image is seriouslydistorted after a visible logo is embedded into it. RVW meth-ods must permit authorized users to restore watermarked im-ages to an unmarked version with sufficiently high quality.Otherwise, the reconstructed image may become too severelydistorted and hence of low (commercial) value.

Conventional methods considered crossed domain ap-proach to realize removable visible watermarking in digitalimage. In particular, watermark is embedded based on thestatistical features induced by the image in frequency domainwhile the final output is stored in spatial domain. In thiswork, we proposed a completely reversible watermarkingmethod in JPEG compressed image that operates solely inthe DCT domain. To ensure perfect restoration, we proposea contrast-adaptive strategy to consider strength of the wa-termark logo during embedding. In particular, the proposedscheme considers an adaptive weight that allows resourceprovider to adjust the strength of the watermark according tothe visual requirement of the protected image. Experimentalresults show that the proposed scheme is more capable ofrestoring watermarked image to it original counter part at

Proceedings of the IIEEJ Image Electronics 　　　　and Visual Computing Workshop 2012

Kuching, Malaysia, November 21-24, 2012

!"#$%&"'$%() !"#$%&"'()

!"#&"'$%() !"#&"'()

Fig. 1. DC position Fig. 2. Watermark logo

higher visual quality when compared to the existing RVWmethods considered.

2. JPEG COMPRESSION

A grayscale image is divided into non-overlapping blockseach of size 8 × 8 pixels and each block is transformed us-ing the discrete cosine basis [9]. The coefficients are thenquantized using the divisors determined by the quality factor.The quantized AC coefficients undergo zigzag ordering andentropy encoding. At the same time, the DC coefficients arepredicted from their immediate left neighbors and the predic-tion errors are entropy coded. For the proposed method, weshall be focusing an the manipulation of the DC coefficients,where each of them encodes the average intensity of the blockto which it belongs.

3. THE ESSENTIALS

The features and requirement of digital watermark are listedin [6][7]. In particular, the authors considered: (a) Restora-tion, (b) Transparency, (c) Global Protection, (d) Multiple Au-thorizations, and (e) Robustness. In this work, we considerthe first three requirements.

4. EMBEDDING PROCEDURE

The proposed reversible visible watermarking method in-cludes two subroutines: weight estimation and watermarkembedding.

4.1. Weight Estimation Subroutine

Assume that the logo to be embedded (i.e., visible watermark)M is a binary image and let DC(m,n) denote the originalDC value located at the (m,n)-th block form = 1, 2, · · · ,Mand n = 1, 2, · · · , N for an image of dimension 8M × 8Npixels. According to the requirements of RVW mechanisms,the embedded visible watermark should be as translucent aspossible. In addition to ensuring that the visible watermark isrecognizable, we aim at maintaining quality of the image to beprotected. In particular, we follow a simple rule where we in-crease the strength (i.e., weight) of the watermark to highlightthe logo in bright region and reduce the strength accordinglyat dark region. We estimate the weight as follows:

Step 1: we compute the (m,n)-th DC coefficient valueX and the mean value of tis neighboring DC values X̄ asfollows:

X(m,n) = DC(2m, 2n) (1)

X̄(m,n) =

DC(2m− 1, 2n− 1) +DC(2m, 2n− 1) +DC(2m− 1, 2n)

3(2)

Fig. 1 shows the position of each DC coefficient.Step 2: The integer difference dm,n between the DC co-

efficient values X and X̄ is calculated as

dm,n = |X(m,n)− X̄(m,n)| (3)

Step 3: We obtain the variance σm,n of the coefficientvalues X and X̄ by computing

σm,n =dm,n − dmin

dmax − dmin(4)

where dmin and dmax are the minimum and maximum differ-ence values among all dm,n’s, respectively.

Step 4: The weight wm,n for the watermark bit Mm,n isdefined as

wm,n = wu + bwv × σm,nc (5)

The secret value factor wu adjusts the strength of the vis-ible watermark by exploiting the human visual system (HVS)while wv controls the strength of the significant texture (i.e.the variance dm,n) of the host images.

4.2. Watermark Embedding Subroutine

Since the content energy of watermark M mainly gathers inthe dark regions, we must highlight these regions of M inthe watermarked image to emphasize its visibility. In ourproposed method, we only embed the dark regions of M toprotect copyright of the image. The bright regions shown inFig. 2 (which posses less energy) are not embedded into theoriginal image. Thus, we can reduce the number of alterationsin the original image to preserve quality of the watermarkedimage while enhancing visibility of the watermark. Obeyingthe above concept, if a watermark bit Mm,n belongs to thedark regions (i.e. the watermark bit Mm,n = 0), we employthe corresponding weight wm,n to adjust the (m,n)-th and itsneighboring DC coefficient values by the following equation:

DC ′(2m+ i, 2n+ j) =

DC(2m+ i, 2n+ j)− wm,n for i, j ∈ {0, 1}(6)

DC shows a DC coefficient which is limited between [-1024, 1023] in JPEG compression. Thereforewu andwv haveto meet following equation.

(a) Airplaine (b) Baboon

(c) Elaine (d) House

(e) Lake (f) Lenna

(g) Milkdrop (h) Peppers

(i) Sailboat (j) Tiffany

Fig. 3. Example of processed image

−1023 ≤ |DC − wu + bwv × σm,nc | ≤ 1024 (7)

These modified DC coefficients are utilized to generate thewatermarked image. The resource provider therefore can is-sue G′ to the individual user along with the secret factors wu

and wv and logo M . Moreover, to avoid unauthorized re-construction of the original image, the secret parameters wu

and wv can be encrypted by the individual user’s secret keyprior to transmission. With weight wm,n, the watermarkedimage G′ is capable of highlighting the energy of the water-mark. Specifically, each weight is adjusted based on varianceof each block strength (i.e., weight) of the image to be wa-termarked. That is, the perceptual quality of the watermarkedimage is sensitively adapted according to the watermark logo,feature of the host image, and user-defined wu and wv .

5. EXPERIMENTAL RESULTS AND DISCUSSIONS

For RVW mechanisms, the greatest concerns are the trans-parency of the watermarked image and the quality of the re-stored image. In this section, we demonstrate the viabilityof the proposed method. In particular, we compare relatedwell-known RVW schemes to the proposed method. Struc-tural similarity (SSIM) [10] is utilized to measure the imagequality of the watermarked and restored images with respectto the JPEG compressed unwatermarked image. The size ofthe test images and the visible watermark are 512 × 512 and32 × 32 pixels, respectively. Ten standard test images (i.e.,Airplane, Baboon, Elaine, House, Lake, Lenna, Milkdrop,Peppers, Sailboat and Tiffany) are considered to verify thebasic performance of the proposed reversible visible water-marking method. The quality factor is set at 75.

5.1. Transparency Analysis

We examine the strength of the visible watermark in varioustest images. As shown in Fig. 3, when we set the secret factorsto wu = 60 and wv = 80, transparency of the watermark logois sufficient to maintain visible detail of the host image whileprotecting the image. That is, potential buyers can browsefor the desired images since each host image is not seriouslyobscured. Using Airplane image as the representative exam-ple, we illustrate the effect of using different values of wu togenerate a suitable watermarked image. Given a fixed wv , wechange the value of wu to adjust the strength of the water-mark. As displayed in Fig. 4, the increase of wu results in amore visible watermark logo. Moreover, the transparency ofthe watermark is still acceptable for all values of wu consid-ered. Furthermore, we adopt the secret factor wv in Eq.(5) tohighlight the significant feature of the original image. Thatis, if the protected image is of high spatial activity (i.e., com-plex), then wv can be utilized to increase the strength of wa-

(a) wu = 20 (b) wu = 40 (c) wu = 60 (d) wu = 80

Fig. 4. Watermarked image using various wu and wv = 80

!"

!#$"

!#%"

!#&"

!#'"

!#("

!#)"

!#*"

!#+"

!#,"

$"

!" %!" '!" )!" +!" $!!"

!!"#

$

%&$

"-./012.1304567189.7:;" "-./012.13045<.=>>:;"

"-./012.13045?9.7:0;" "-./012.13045@>AB0;"

"-./012.13045C.30;" "-./012.13045C0::.;"

"-./012.13045D79341>8;" "-./012.13045E08801B;"

"-./012.13045F.79=>./;" "-./012.13045G7H.:I;"

(a) Watermarked

!"

!#$"

!#%"

!#&"

!#'"

!#("

!#)"

!#*"

!#+"

!#,"

$"

!" %!" '!" )!" +!" $!!"

!!"#

$

%&$

"-./01.234567895:.;" "-./01.23<9=00:;"

"-./01.23>895:.;" "-./01.23?0@A.;"

"-./01.23B9C.;" "-./01.23B.::9;"

"-./01.23D58C2607;" "-./01.23E.77.6A;"

"-./01.23F958=09G;" "-./01.23H5I9:J;"

(b) Removed

Fig. 5. SSIM of various removed/watermarked images for wv = 80

(a) original image (b) watermarked image

(c) recoverd image

Fig. 6. Visual comparison of original, watermarked and restored images

termark so that to improve recognition of the watermark logo.However, larger values of wu and wv may cause underflow(i.e., coefficient are out of the range handled in JPEG com-pression standard). Nevertheless, since the range for DC coef-ficient handled in JPEG is [-1024,1023], overflow/underflowis very unlikely to occur. Hence, the proposed method pre-serves transparency.

5.2. Analysis on Reversibility

In general, distortion of the restored image is often causedby the embedded watermark. The quality of the restored im-age is an important concern in evaluating a RVW mecha-nism. SSIM of value unity implies that a perfect reconstruc-tion is achieved, and anything below unity signifies that thereare some discrepancies between the original and the recon-structed images. As illustrated in Fig. 5, the SSIM value forall restored images for the proposed method is 1 after thewatermark is removed. For our simulation with the valuewu ∈ [20, 100] and wv = 80, the proposed method alwaysachieves perfect reconstructions. Fig. 6(b) and (c) present thewatermarked image and reconstructed images, respectively,with wu = 60 and wv = 80. Compared with Fig. 6(a) , therecovered image shown in Fig. 6(c) is exactly the same as theoriginal image. This indicates that proposed method is capa-ble of removing the embedded watermark logo such that therecovered image is completely identical to the original image.Therefore, the proposed scheme can realize perfect restora-tion.

5.3. Analysis on Robustness

The robustness of the proposed method against common wa-termark attacks is examined in this section. StirMark [11][12]has a variety of watermark attacks but the difference in visualappearance is most significant after applying the Median fil-ter [3]. Since the embedded watermark is visible, it is ableto withstand common attacks such as rotation, shearing, andresizing. For that, we only focus on the results of Median fil-ter. Here, median filtering is carried out using window of size3 × 3 pixels and 5 × 5 pixels in StirMark [11][12]. Due tospace limitation, only the results for 5 × 5 are shown. Fig. 7shows the watermarked image for wu ∈ {20, 40, 60, 80} and

(a) wu = 20 (b) wu = 40

(c) wu = 60 (d) wu = 80

(e) wu = 20 (f) wu = 40

(g) wu = 60 (h) wu = 80

(i) wu = 20 (j) wu = 40

(k) wu = 60 (l) wu = 80

Fig. 7. Median filtered watermarked image using various wu and wv = 80

wv = 80 after filtering. These images still have outline of theembedded watermark logo. Similar results are observed forthe case of 3× 3 window. Therefore, the proposed method isrobust against Median Filter.

5.4. Functional Comparison

Since the proposed method operates solely in the DCT do-main, it is of lower complexity than any of the cross-domainapproaches. The proposed method also requires less mem-ory to operate since the operations are carried out on a block-by-block basis. Complete reversibility is an important fea-ture in RVW method and the proposed method offers this fea-ture.Table. 1 compares the operational function between theproposed method and those of the existing methods. It ob-served that the proposed method offers more features than theexisting methods considered.

6. CONCLUSIONS

A novel reversible visible watermarking method was pro-posed in the DCT compressed domain. DC coefficients weremanipulated adaptively according to the watermark logo andthe host image. Transparency and reversibility were simul-taneously achieved. Simulation results showed that the pro-posed method outperforms the existing methods consideredin terms of the reversibility.

To recover the original image from the watermarked coun-terpart without extra storage for side information will be themain focus of our future works. We also want to apply theproposed to video.

7. REFERENCES

[1] Yongjian Hu and Sam Kwong, “Wavelet domain adap-tive visible watermarking,” Electronics Letters, vol. 37,no. 20, pp. 1219–1220, 2001.

[2] Biao-Bing Huang and Shao-Xian Tang, “A contrast-sensitive visible watermarking scheme,” Multimedia,IEEE, vol. 13, no. 2, pp. 60–66, 2006.

[3] M.-J. Tsai, “A visible watermarking algorithm basedon the content and contrast aware (cocoa) technique,”Journal of Visual Communication and Image Represen-tation, vol. 20, no. 5, pp. 323–338, 2009.

[4] Yongjian Hu, Sam Kwong, and Jiwu Huang, “An al-gorithm for removable visible watermarking,” Circuitsand Systems for Video Technology, vol. 16, no. 1, pp.129–133, 2006.

[5] Y Hu and B Jeon, “Reversible visible watermarking andlossless recovery of original images,” ACM Trans. on

Table 1. Functionality comparison among related worksProposed Ref [5] Ref [3] Ref [8]

Reversibility Yes Yes No NoTransparency High Low High High

Muti-Authorization Yes No - YesUnderflow/Overflow No/No No/No No/No No/No

Grobal/Local Protection Global Local Global GlobalOperating domain DCT Spatial Cross Cross

Output format JPEG RAW RAW RAW

Modeling and Computer Simulation, vol. 16, no. 11, pp.1423–1429, 2006.

[6] F. Mintzer and G. Braudaway, “Effective and ineffectivedigital watermarks,” IEEE International Conference onImage Processing, vol. 3, pp. 9–12, 1997.

[7] Minerva M. Yeung and Fredrick C. Mintzer, “Digitalwatermarking for high-quality imaging,” IEEE FirstWorkshop on Multimedia Signal Processing, pp. 357–362, 1997.

[8] S. P. Mohanty, ISWAR: An Imaging System with Wa-termarking and Attack Resilience, Ph.D. thesis, CornellUniversity, 2012.

[9] William B.Pennebaker and Joan L.Mitchell, JPEG: still

image data compression standard, Van Nostrand Rein-hold, 1992.

[10] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simon-celli, “Image quality assessment: From error visibilityto structural similarity,” IEEE Transactions on ImageProcessing, vol. 13, no. 4, pp. 600–612, 2004.

[11] Markus G. Kuhn Fabien A. P. Petitcolas, Ross J. An-derson, “Attacks on copyright marking systems,” In-formation Hiding, Second International Workshop, pp.219–239, 1998.

[12] Fabien A. P. Petitcolas, “Watermarking schemes evalua-tion,” IEEE Signal Processing, vol. 17, no. 5, pp. 58–64,2000.