An adaptive digital image watermarking technique for copyright protection

AN ADAPTIVE DIGITAL IMAGE WATERMARKING TECHNIQUE FOR COPYRIGHT PROTECTION

Chang-Hsing Lee' and Yeuan-Kuen Lee'

Department of Computer Science, Chinese Culture University' 55 Hwa Kang Rd., Yang Ming Shan, Taipei, Taiwan 11114

Department of Computer and Information Science2

National Chiao Tung University 1001 Ta Hsush Rd., Hsinchu. Taiwan 30050 Ahtrtict- A n atlaptivc digital image

watcrmarlting tccliniquc i s prop(iscd i n this paper. Tlic proposed method exploits the sensitivity of human cycs to atlaptivcly cmhctl a visually

rccognizablc watcniiarli i n an image witliotit affecting tlic pcrccptual quality o f the tinderlying host image. I n addition, the watcrinarli w i l l s t i l l be present i f sonic lossy image processing o p c ~ i t i o n s such a s low-pass liltcring, mctlian filtering, rcsampliiig, rcc~ii~iiitiz;ition, and lossy JPECi iniagc compression arc applied to the

watcrtiiarlictl image. I(xpcrimcntal results show the cflcctivcncss of the proposed watcrmarlting mcthotl.

Index T e r m - copyright protection, digital watcrniarlting, 1 I'EG compression

1. I N ' I ~ I W I ~ U C I ' I O N

'l'hc rapid evolution o f the Intcrnct ni~11~cs casicr the transmission o f digital miiltinicdia

content stich as text, audio, image, inid video. Digital incdia can bc accessed 01- tlistributctl

through the network. As a rcsult, rcplications of digital nicdia arc simplc with no loss 01' litklity, that is, the copy of ti digital mcditnn i s itlcntical

to the original otic. 'I'hrougli the network, an unlirnitetl number o f identical copies o f digital mcdia can hc illcgally protlucctl, this i s a serious threat to [tic iiitcllccttial property rights of the

media owner. Thcrcforc, to protcct aiitl enforce ititellectual property rights o f tlic iiictlia owiicr i s an important issuc in the digital world

To protect digital media, traditional encryption algorithms sticli 21s DES or RSA arc widely adopted [ I ,

21.

111 tlicsc encryption algorithms, the digital mcdia is encrypted into

scramhlcd data using a prctlctermincd encryption ltcy (private ltcy or ptihlic Itcy). l'lic encryption

algoritlims hide tlic mc;niing of the original inicdia. An oppcincnt who iicccsscs tlic cncryptctl data docs not know what t h e original tncssagc is.

Only t h e Iioltlcr who knows the correct

dccryption ltcy can tlccrypt the encrypted daki and recover the original iiictlia. 1 lowcvcr, the dccryptctl iiicdiii, wliicli is idcntical to tlic original one, may be illcgallp tlis~ributctl or copied. The digital watcrmarlting tccliniqtic provitlcs a good way to solve this isstic. A digital watcrniark can be tisctl to claim the owticrsliip of the digital ~ i i c d i n and protect tlic intcllcctual

properly rights o f the media creator or owncr. Tlic tligilal \vatcriii~irltiiig tccliniquc ciiibcds ii digital signature or digital waterinark, which asserts the owiicrsliip or intcllcctiial propci-ty rights of the media creator or owner, i n the

digital mctlia sticli a s text, audio, itiingc, iintl

vitlco. l'lic watermark can then be extracted from thc watcrn~arlictl nicdiii mtl i s used to itlcnti fy the author or tlistrihutor o f the media. The principle oT digital watcrmarlting i s the rohust mid sccrct cmbcdding o f copyriglit informatioii in

a digital tiicdium. 'lo he rcally cl'l'cctivc for copyriglit cnlhrcciiicnt, ii digital \vatcrmarking

tccliniquc inust satisfy the Ibllowing rcquircmcnts:

( 1) Peiwptuol t~vr~~.~/~cire~ic'),y

l ' l i c cinhcdtlctl watermark intist he perceptually invisihlc or inautlihlc to maintain tlic cl~ialily o f the host mctlia tnidcr typical

pel-ccptual conditions. 'l'hat is, hiiman ol~scrvcrs

cannot distinguish the original host nictlia froiii

the watcrniarlictl media. As a rcsult, the cxistcticc

o f t h c watcrniark is liidtlcn to Iiuman ohscrvcrs. (2) Una,uhixui/y

l h c retrieval oT ii watermark should uiiambiguotisly itlcntiry the owncr. Lii addition,

[lie accuracy o f owner identification sliould

tlcgratlc gracefully under attacks. (3) Robustness

As a watcrmarlt is used to idcntily the owner o f digital nicdia, removal of the cnibcddctl watermark should be difficult [or iin altackcr or any tniauthorizcd uscr. In practice, any watermark can he rcmovcd i l sufficient insertion is known. I~lowcvcr, if only parLiiil information is available, atlcnipting to rcniovc or destroy (lie watermark should produce a remarkable degradation in nicdia quality before tlic watermark is lost. In general, lossy signal processing opcrations that damage the waterniarltctl media may also dainagc tlic watcrmarlt. Therefore, tlic watcrmark m u s l slill hc present if the watcrmarltcd nictlia arc processed by sonic cotninon signal proccssing opcratioiis. Tlicse operations include rcsainpling, requantization, lossy compression (e.g., JI'EC;, MPEG, wavclet compression), linear filtering (e.g., low-pass and high-pass filtering), nonlinear filtering (e.g., iiicdian liltcring), geometric distortions (e.g., scaling, translation, rotation, ancl cropping), as well as digital-to-analog and analog-to-digital conversion. I n general, the robustness ollcn conflicts with Iransparcncy requircnicnt. To be robust, ii watcrinarli should bc cmbcddcd in pcrccptually signilicant regions of thc host mcdia. On the other hand, to be transparent to huniaii observers, a watermark slioulti be cmbctldcd in pcrccptually insignificant regions of the host nictlia. 'I'hcrcforc, to propose a transparent and robust watermarking schctnc is an inipcdinient to

many

researchers.

(4) Tatnper-i-e.Fi.Ftaiice

l'hc ctnbetldetl watcrmarli must be resistant to tanipcring through collusion by comparing tiitiltiplc copies of tlic incdia cnibcddctl with di nbrcnt watcrniarlcs.

A transparent and robust digital iniagc watermarking approach is proposed in this paper. The cmbeddcd witcriiiark is invisible to Iiuman cycs and is robust i f the watcrmarkcd image is processed by some lossy iniagc processing operations, such 21s low-pass filtering, nicdian liltcring, resampling, rcquantizalion and JI'EG image compression. In the next section, we will ltnowlcdgc about the process of watcrmar1i

give 21 survey o f existing walermarlting mctliotls.

'I'hc proposed digital image watcrniarlting approach is tlescribcd i n Section 3 . I n Section 4, we present the cxpcrimentd results and show Lhc robustncss o l the proposed approach. I'inally, ii

brier conclusion antl tliscussion is presented i n Section 5 .

I I . I'KEVIOUS WORKS

111 this scclion, wc will give a rcvicw on

digital image watcrinarlcing tccliniqucs. A tlctailctl rcvicw on niultiincdia data cnibctlding antl watcrniarlting techniques can also bc l'ound in [ 3 , 4:l. l'hc digital imagc watcrniarlting techniqucs can lie classi ficd into two categories: spatiel-domain tcchniqucs (spatial watcrinarlis) 15-14] and ficciucricy-domain tcchiiiqttcs (spectral watcrmrlts)

I

15-26], The spatial- domain techniques directly inotlily the intensities or color valucs of some sclcctcd pixels while thc frcqticncy-[loiiiaiii tccliiiiclucs ti1otliCy tlic valtics of some transfornictl coclficicnk.

l'hc simplest spatial-tloinain itiitigc watcrmarlting ~ccliniqttc is to cnihcd a watermark in the least significant bits (LSBs) of sonic randomly sclcctcd pixels [3-51. 'i'hc watcrniark is actually invisible to human eyes. However, the watcrinark can be easily destroyed if the watcrmarltctl imagc is low-pass lilterctl or JI'EG coniprcsscd. To increase tlic security of the watermark, Matsui antl Tanalta I:6] proposcd a method that uscs a secret ltcy to sclcct the locations whcrc ii watermark is cmbctldcd, e.g.,

the use of a pseudo-random number gcncrator to dctcrminc the sequence of locations on the image plane. Vopatzis and Pitas used a toral autoniorpliisni [7] approach to s c ~ i i n b l c tlic digital watcrinarlc before it is inserted into an image. To increase (lie robustncss of the watermark, many approziclics have been proposed to modify sonic properties o f sclcctcd pixels or blocks 18-14], Wolfgang e/

d.

reshaped

a n in-scquencc into two-dimensional watcrniarli

blocks, which are added antl dctcctcd on ii block-

by-block basis

1x1.

Pitas proposcd a nicthotl that shills some pixel valtics for (lata cnihcdtling [9,

IO]. I n his nicthotl, a digital watermark

,S

is a spccilic binary pattern o l s i x N x M wlicrc the number o f 1's cquals the number ol 0's.

Dcpcntling 011 Lhc binary paltcrii of

s,

the gray scale image

f

is split into two subsets, A and /I, of cqn~il size I'=NxM/2. 'I'Iic intcnsitics oT pixels iii subset A are added hy a quantity I< while the intcnsitics of pixels i n subset I) arc not alterctl. Uryuntloncltx et ul. proposed ii block-based

spatial watcrmarking mctliod tliat modi lies tlic average luminancc valnc ol'ii block 11 I , 121. An image is lirst (1ccoiirpc)scd into a set ol' IIXII

hloclts. A ltcy is iiscd to dctcrminc the cmbcdding bloclcs or locations. 1':acIi bloclc is classilicd into one of the tlircc types of contrast: li:ird, progressive, antl noise contrast. The pixcls in a block arc assigned to one o f two zones, xonc I and zone

2.

Each zonc is ftirtlicr dividctl into Lwo categories: A and 11. 'Ilic division is hascd on

a grid dctcrminctl by tlic cotlcr. Enrbcdding of ii

bit h is pcrforincd hasctl on the Tollowing cmhcdtling rulc:

if/)=O: m,,,' - mlA' z. 1.

- 112?,,' I'

i f b = l : inl,,-mlil*=/.

Ill2,,* - lllZlI* = I,,

wiicre 1 1 1 ~ , ~ * , inl,,*, ni2il*, and iii2,,* arc the avcragc

luminancc valucs after cinl~ctltlii~g a bit b ~iiitl L is tlic cmhcdding level. Kuttcr c l d . proposed a n

amplitutlc modulation approach for color images watcrmarlcing [ 181. The locations, wlicrc the watermark is cmhctldctl, arc tlclcrminctl hy nsing a sccrct lccy. l'he watcmiarlc bits arc cmbcdtlctl in tlic bluc cliaiinel since tlic human visual system ( I W S ) is relatively less scnsitivc to this color information. A single bil s is cinbcddcd in a I)scLid(~-raiid[)1iiIy sclcctctl pixel at location (i, , j ) by inotlifying the bluc clianiicl 8 by a fraction of the luminancc L as:

wlicrc q is a constanl v;iltic used to tlctcrminc tlic strcnglh o f the signature antl is sclcctcd to optimize rolxistncss antl invisibility. 1)arrwi ~ m t l Scott proposed a fractal-hascd stcguiographic mcthotl Lo cmbcd binary messages 1,141. I'ractal analysis is first tisctl to identify similar bloclcs. The set o f similar bloclts is then tlivitlctl into two catcgorics. '1'0 cmbctl ii "0" bit, tlic bloclts in tlic

a " I " hit, the blocks in the second category arc

used.

TIic frccluciicy-domain tcclinic~ncs lirst / I j j t B , j ' l - q(2s - l ) / , j J ,

first category arc usctl. On the contral-y, to cmbcd

translorni an iiitage into the ficqucncy tloniain coefficients. 'I'Iic transforination may hc I'ouricr transfom Lis.1,

L)cr

116-231, or \wivcict transroriii [24-26:1, etc. 'I'Iic watcrmarlt is t11c11 cmbcddcd in tlic tr;nisforiiictl coclTicicnts according to LIIC pcrccptual significance of Lhc

transform cocfficicnts. Thcrcforc, the waterinark is irrcgnlarly distribiitctl over tlic cntirc image. l~;indly, the cocflicicnts arc invcrsc-tr~iiisTormctl to form the watcrmarltcd image, which is identical to the original iinagc. As a result, the watcriiiarlt is invisible Tor iiii cncniy to dccodc o r to read and is more rohust to sonic iinagc ~xoccssing operations. O'litianaitlh cl id. 151 ciiibccldcd the waLcrniark in the pliasc information i n the tliscrctc Fourier translhrm doiliain siiicc tlic ~pliasc clistorlion is more scnsitivc to I I I I I ~ ~ ~ I I I visual systcin than the magnitntlc tlistortions. Therefore, it is iiiorc robust to tainpcring wlicii ccirnparctl 10 magnitutlc ~notlulation. Cox e/ trl. [ I71 l~roposcc'l

a s c c ~ i r c spread spcctrum watcrmarkitig method [or cnihctltling a waLcrmark in the 1)C1' domain.

111 the algori~hm, tlic watermark is inserted i n the

pcrccptual significant portion of tlic image i n orclcr to provide greater robustness. The watermark W is ii scqncncc

01'

normally

distrihutcd, zero-incm and tinit-variance random numbcrs. 'l'liat is, W = ( i q ,

w2,

. . ., w,J, wlrcrc c a c l ~ \vi is clioscii according to N ( 0 , I ) . A DC1' is Grst pcrformctl on thc entire iinagc antl ilic cocfticicnts with thc largest magnitutlcs ai-c itlcntilictl as the perceptually significant portion of the imagc. Then, the watermark is inserted into tlrcsc sclcctctl cocrTicicnts by setting each frcqncrrcy cocflicicnt C, 21s:

where U; is a sciiliir factor. I'inally, the itivcrsc

DCT oT the watcriuarkctl coefficients will form a transparcntly w;itcrmarlced image. Both the original and wiifcriiiarltctl images arc nccdcd to extract the cmbcdtlcd watermark. A similarity nrcastirc is then iisctl to comparc the cxiractctl watcrmarlt with the original one to test whether a watcrmarlt is prescnt in the image. l ' h c r c s u l ~ s sliow tliat the tcchniquc is crfcctivc in terms oT

transparency antl rohustncss. A block-basctl U C I w;itcrmarlting xpproach was pi-oposctl by I I s u and Wu [18-20]. The waterinark is 21 visrially

rccognizablc pattern SLICII as ini image of a seal

with Chinese characters. An image is first divided into bloclcs and DCT is per~ornicd on each block. 'The watcrmark is thcn cmbeddcd by selectively modifying tlie mitldle-frequency DCT coefficients. Sincc the cmbeddcd watcriiiark is an image, hunian eyes can easily identify thc cxtractctl watermark. 'l'liey claim that thc watcrniarkcd iinagc is robust to general image operations and lossy JPEG compression. Tang and Aolti 1211 also proposcd a block-based middle-band cmbcdding algorithm, which is similar to L18-20:I. A differential pulse code modulation is uscd to pcrniutc a watermarlc image. Similar to some inclhods [IC,, 22, 231, their approach cxpolits tlic characteristics of huinan visual system in the cmbcdding proccss. Thc aim is to insert inore cnibctlding bits wlicrc they are most robust to attack mid arc least noticeable.

Ohnishi and Matsui cmbcddcd a watermarli in the Harr wavclct lransform domain 1241. To embed the watermark, one is addcd o r subtracted koni some sclcctcd transform cocflicicnts. Xia et

LII. introduced a iiiultircsolulion watcrmarlcing mcthod for digital images based on digital wavelet translorm [25]. I n tlic method, a watermark

W(m,

71) is a Gaussian noise with

7,cro-nicaii and unit-variance. The watermark is inserted into the largc cocrficicnts at the high and middle frequency bands o f an image according to

Lhc following equation:

P"!J(m, n ) = IIJ(m, 71)

+

(x [I,@, /1)] W(m, n ) ,

where I,,,(rn, n ) and P,,(nz, n ) refer to the original and watcrtnarlied wavelet cocf'ficicnis at position (m, n ) in I-csolntion lcvcl

I

and frcqucncy orientation ,f; and U is a conskant thai is maximized under the Lraiisparcncy constraints. Kunder and Hatzinalcos uscd multircsolution fiision techniques and incorporaicd a model of the human visual system to cnibcd a watcrmarli

P G I .

In general, a spatial-domain watcrmarliing method has larger capacity than that of a

fi.equcncy-tloniaiii method. That is, niore (lata can bc cmbcddcd in the spatial domain than i n the lrcqncncy domain. However, data cinhcddcd in the frcqticiicy domain is more robust to coininon image processing operations. 'l'hcrcforc, there is a trade-off between the capacity and

robustness. Most of the watermarlting algorilhms use a serial number, a set of normally distributed randoni numbers, a Gaussian distribution, or an author ID as a watcrmarli. In thcsc algorithms, a quantitative nieasnrc is required to verify the extraction results. Usually a similarity, y, between the original watcrniarli and cxtractcd watcrmark is computed. 'The valuc of y is thcn lcstcd against a thrcshold 7'. If q > T, it is assunicd that the iniagc is watcrmarkcd, otherwise tlie image has no watermark. However, tlic determination of the threshold valuc T protluccs ambiguity. A small valuc o r 7' will accept the cxistcncc of a watermark although there is iionc. On the other hand, a large valuc of 7 will reject the existence o l a watermark although there is oiic. l'licrcforc, how to decide a proper threshold valuc becomes a serious problem. A bcttcr solution is to L W a visually meaningful watermark (e.g., a small iniagc) [ 18- 201. Hunian eyes can then easily verify the cxtraclion rcsults. However, a large quantity of data must bc embedded in the host inxigc if a

visually mcaningful watermark is adopted. T ~ U S the cmbcdtling algorithm must adapt its insertion strategy to accommodatc a largc quantity of data in the host image.

As dcscrihcd above, to provide larger capacity for waterinark insertion, a spatial- domain watcrinarking method is preferable. In fact, cmbcdding a watcrmark in the least significant bits of a pixel is less sensitive to human cycs. However, the watermark will be destroyed if some coininon image operations such as low-pass liltcring are applied to the watcrniarlicd iniagc. Thcrclore, to nialcc the cmbcdded watermark inore resistant to any attack, the watermark must be cinhcddcd in the most significant bits. liowcvcr, this will introduce more distortion to the host image and conflicts with the invisible rcqnircmcnt. To meet both invisibility and robustness, wc will proposc a

nicthotl that adaptively modifies the intensities of some sclccted pixels as much 21s possible and this iiiodification is not noticeable to human cycs.

In next section, wc will describe an adaptive image watcrinarking approach. The proposcd approach utilizes the sensitivity of the human visual system to adaptively modify the intensities of sonic pixels in a block. Tlic modification of

pixel intensities dcpcnds on the cotrtcnl or ii block. If the contrast o f the block is large, the intcnsitics can he changed grcatly without introducing any distortion. On the other hand, il" the contrast is small, the intcnsitics can only bc changcd slightly.

111. 'I'IXIC PROPOSED AI'I'ROACII

In this section, wc will tlcscribc the p~-oposc"d adaptive image watcrmarlting tccliniqnc. The watermark iiscd is ii visually rccogtii7ablc binary

image rather than a raiitlomly gcncratcd scqucncc ol' bits. Thus, htnnan cycs can easily idcnticy the extracted watcrmark. l'hc proposed techtiiquc adaptivcly modifies the intensities o r sonic selected pixels as much as possible antl the niodilicalion is not noticcablc to human cycs. In addition, to prcvcnt tampering or tinautliorizccl R C C ~ S S , the watcrtiiark is first pcrmutctl into scrambletl data. The block diagram ol' the proposed watermarking s y s h n is dcpictctl

in

Fig. 1. In thc hollowing subsections, wc will lirsl givc the pertiintation algorithm, ;itid then tlcscribc the watermark cmbcdtling and cxtl'action ~ ~ r o c c s s c s .

riiiiitiitioii

A. M/atennnrk l'ermutrrtion Algorillzm

1'0 prcvcnt the watermark from tampcring or unauthorized access by attackers, the watermark image is first pcrmutcd to bc scranihlctl data hcrorc insertion. l'hc walcrmark pcrmu~atio~i strakgy is tlic same as that proposed in [ 20'1. A two-dimensional psetitlo-random iiLniibcr traversing method is used to permute tlic watcriiiark. Let W antl W,, hc the original and pcrniiitcd walcrmark image, thal is,

W,, = { i ' ~ , , ( i , , j ) = ~ ( ~ ' , ~ ~ l O ii,i'<M and Osjjkiv}, where pixel at ( i ' , j ' ) is mappctl to pixel at (i,,j) i n a psetitlo-r;indoiii order, iL1 and N arc the height and width o r the watermark image, rcspcctivcly. 'The pcrmnhtion ;ilgoritIim is implcmcntctl a s Ibllows: Step I : Numhcr each pixel l'rom 0 lo (MxN-I)

it, 'I I taster ., ~,

Stcp 2: Cicticratc ii scqncncc of ( M x N I.;intlom

nttmbcrs bctwccn 0 inid ( M x N - I ) using linear feedback shill rcgister [28]. liacli pixel / J is tlicn niappcd to a I.;intloiii

I<cplacc tlic pixel value O ~ / J with y.

scan ortlcr of tlic image.

V ~ l l l l c (J. Step 3:

:

Host Im;igc Ilccoiistructed watcriiiark K e y

Watermark liisertioii Watcrinark ICxtraclion

1

1

T

1

T

Fig. 2 An illustration of the rnlcs for pixel intensity.

B. Wnterninrk Etnhedding

After the binary watcrniark iinagc is pcrmutcd, tlic scramhled data scqucncc is then inserted into the host image. The cmbctldcd watcrtnark mist be invisible to human eyes and robust to most iinagc processing operations. To incct these requiremcnts, a bit of pixel value (0 or I ) is cmbcdded in a block of the host iinagc. Depending on thc contrast of the block, pixcls

i n

this block arc adaptivcly niodilied to maximizc robustness and gnarantcc invisibility. Hcforc insertion, the host image is lirst tlcconiposcd into blocks of s i x n x n . The position or block for cmbcdding is selected by a pscudo-random number generator using a scctl valuc k. The valuc of /c is similar to the sccrct key of a scct~rc DL(S system. Let I3 be a selected hlock, thc watermark insertion method is described as follows:

Step I:

s-2:

Step 3:

Sort thc pixcls in block B i n an ascending order of pixel intcnsitics. Compute the avcragc intensity g,,,,

,,,,,

rnaxiinal intensity gl,,zlA, and ininimal intensity

s,,,~,,

of the block. That is,

g,,,;,, = max (tiq, 0 < i , j < n ) , and g,,,,,, = niin (/I;,, 0

<

i, j < n ) ,

whcrc b, rcprescnh the intensity of thc (i,j)-th pixel in block U.

Classify cvcry pixel in B into one o f the following calcgorics by using g,,,,;,,,:

bo c Z,l if b,j

g,,,,;,,,

b, t Zl, irtJii<g,,,c:,,l,

whcrc Z,, and ZL rcprcscnt high- intcnsity category and low-intensity calcgory, rcspcctivcly.

Cotiipntc the incini values, inll and nz,, of these two catcgorics.

Define the contrast valuc of block I3 as Stcp4:

Stcp-5:

G I = m4C,,,i,,>

4s,?lt,h

-

g,,,i,,))>

Whcrc

a

is a constant, and is a constant valuc which determines the minimal valuc a pixel should bc modilicd.

Assuming that the cmhcddcd valtie b, is 0 or 1. Motiiry the pixcl valucs in block B according to the followiiig rnlcs:

s'

=

s,,,;,,

i f g >

m,,,

s'

=

st,,,;,,,

if mI.~s<~,e,,,,,,, s'=g+6 othcrwisc,

s'

=

smi,,

g'=s-cy othcrwisc, SJep6: if / I , = 1 : if h," = 0: il'g < in,,, . . fi' = &,,,, It g,,,,;,,,%<%,

whcrc g' is thc inoditkd intensity and 6 is a randomly gcneratctl value bctwccn 0 and Cll. Fig. 2 dcpicts thc modification of pixel intcnsitics according to the rnlcs de

'llic cnihecitling of tlic watermark bit dcpcnds on the cotitclit of cacli block. If the block is of larger contrast, thc intensities of pixels will bc changed greatly. Otherwise, the intensities arc tuncd slightly. In the cxtrcinc case, if a smooth

block, whcrc all tlic pixels lhavc the smic pixcl intcnsiiy, is clioscn for tlnta cmhcdtling, the pixel intcnsitics w i l l be tutictl by ii sinal1 rantloinly

gcneratctl value. This modi lication can avoid the blocking a r t i f k t sincc the pixel intcnsitics arc addctl or subtracted by sonic randomly gcncratctl valncs instead o f a constant valuc. l'litis tlic pixel

intcnsitics can l i e atlaplivcly motlifietl tlcpcntling

011 thc contrast 01' 11 hlock. Lcl bloclt I( illld 1%'

dcnotc the original ancl iiiodilictl bloclts, rcspcctivcly. T l i c sum of pixcl intcnsitics of 1%' w i l l bc larger than that o f H i f the inserted waterinark pixcl valuc b, is I . On tlic contrary, if tlic inserted watcrmarlt pixcl valuc h , is 0, the sitin of pixel intensities o r I)' will be sillaller than that o1'B.

C. Wutei-nicidc Exxtr-uction

l'he extraction of ii watcmiark is similar to

the cmbcddiug process wliilc i n a rcvcrsc ordcr. I n the proposccl algorithm, tlic extraction 01' a watermark m i s t rcfcr to the original hosl imagc. First, wc LISC tlic seed valuc, k , to get a scqncncc

of positions or bloclis wlierc thc waterinark is cmbcddcd. For each selectcd position, let H and U' represent the corrcspontling blocks of the original host iinagc and watcrmarkctl imagc, rcspectivcly. Coniputc the SLIIII of pixcl intensities,

S,, ant1 S , , of U and 1%

'.

The rctricvctl w;ilcrmark bit valuc b, is dctcrmincd by comparing S,, and S,\ :

/I,!, = 1 if S, >

S,],

h, = 0 otherwise,

The cxtractcd watcrmark bit val~ics, h,'s, iirc then invcrscly permuted to gct the rcconstructcd watermark.

1V. EXl'lr:KIMENTAL, I<ICSULTS In tlic cxpcriincnts, the host iinagc is o r size

S 12x5 12 with 256 gray Icucls. The watermark is

ii visually rccogni7ablc hinary image 01' sizc

128x128. Figs. 3 ( a ) and 3(h) show a 512x512 host inlagc ant1 a 128x 128 binary watcrm;irlt

imagc, respcctivcly. I'ig. 3(c) shows the watermarkcd image that is derived by cmhcdtling the watermark i n tlic host imagc. From Figs. 3(a) antl 3(c), wc can sec that tlicsc two images look almost the same. I'ig. 3(tl) shows the rcconstrnctctl watermark, wc can sec that it is the same as Fig. ~ ( I I ) . 'I'Iic siinilarily bctwccn these two images is qitantitativcly mcasurctl by tlic normalizctl cross corrclation [ 181 dclincd as:

c c

W ;; W'ii

whcrc W , nntl W'ii rcprcscnt thc pixcl valucs at location (i, , j ) in the original and cxtractctl \vatcrniarlc images, rcspcciivcly.

algorithm under c o i n i i m i i iniagc processing operations, we liavc proccssctl tlic watcrmarlcctl iniagc nsing the following operations: liiiciir low- pass filtering, median liltcring, rcsampling, rcqiiantization, mid lossy ,Il'E(; compression. Fig. 4(a) shows ihc ~ C S U I ~ 01' applying 11 linear IOW-

~ I I S S filtcriiig lo the walcrmarlictl imagc. The filter is ii ~nciglihorliootl avcraging operation with

a iiiask ol'sizc 3x3. Fig. 4(b) shows the cxtractctl watermark liom I'ig. 4(a). llic nornializcd cross corrclatioii v;iluc

NC bclwccn

tlic cxtractctl watcrmarlc antl original one is 0.9658. b'ig. S ( a )

shows the rcsult of applying a 3x3 ncigliborliootl mctlian filtering to the watcrmarlted imagc. TIic cxtractctl wmtcrmarlt is shown i n Pig. S(b). 'l'lic normalizcd cross correlation valuc NC,' is 0.8879. From Figs. 4(b) antl 5(h), wc can easily verify the existence o f tlic watcrmarlt altliotigli thcrc is sonic distortion i n tlic cxtractcd watcrmarlc.

Fig. 6 ( a ) sliows tlic rcsitll of applying rcsampling operation to tlic watcrmarltctl image. l'hc waterniarkctl is l i n t scaled to be 114 of its original s i x by using a 2x2 subsampling operation. l'licn the snbsamplctl iniagc is intcrpolatcd to the s i x o f the original onc. Fig. 6(b) shows tlic cxtractctl w;itcrtn;irk from I'ig. 6 ( ~ ) . l'hc normalizctl cross correlation viilt~c N(,' is 0.989 I, I'ig. 7(a) shows the rcsult ol' applying rccltiantization operation to tlic watcrmarltctl imigc. The watct-mai-ltctl itniigc with 256 guiy lcvcls is rcqttantizctl to bc of 32 gray Icvcls. Tlic cxtractcd watermat-It Trom this 32-lcvcl image is shown in Pig. 7(b). Tlic norm;tlizctl cross '1'0 SllOW thc robuslllcss of the proposctl

corrclation value is 0.9475. From Figs. 6(b) and 7(b), we can see that thc extracted watermark is almost tlic samc as the original one. llius, the proposed mcthotl is very robust to iniagc rcsampling and rcquantization.

To show the robustness of tlie proposctl algorithm under lossy JI'EG compression, we first compress the watcrmarkcd image and then extract the watcrniark kom the compressed image. Fig. 8(a) shows the coinprcsscd iinagc with a compression ratio (CK) of 7.27. The cxtracted watcrniark is shown i i i Fig. 8(b). 'Ihc nornializcd cross corrclation valuc NC is 0.9929. I'ig. 8(c) shows tlic compressed iniagc with a coniprcssion ratio (CR) of 14.39. The cxtracletl watermark is shown in Fig. 8(d). The normalized cross corrclation valuc NC is 0.9103. Fig. 9 comparcs the normalizcd cross corrclation values for di€feretit JPEG coniprcssion ratios. Prom this figurc, we can see that thc normalizcd cross corrclatiori values range from 0.9103 for a high JPEG compression ratio to 0.9998 for a low image compression ratio.

From the above cxpcrimental rcsults, we can scc that tlic extracted watermarks can be easily used to identify the owiicr of the host image sincc it is a visually recognizable binary image. In addition, the proposed algorithm is robust to conitnon image proccssing opcrations such as low-pass filtering, median filtering, rcsampling, requantization, and lossy JI'EG comprcssion.

V. CONCLUSIONS

In this paper, wc have proposcd an adaptive watcrniarking algorithm for images. The watcrniark adopted in this paper is a visually mcaningful image such that human eyes can easily judge the extraction rcsull. To embed a watermark in tlic host image, the proposed approach utilizcs the scnsitivity of human visual systcni to adaptivcly modily Lhc contents oi' a set o f blocks. The pixcl intensities in a block are changcd adaptivcly dcpending on the contrast of the block. The modification of pixel intensities o r the block is large (e.g., an ctlgc block), the intensities can bc changcd grcatly without introducing any distortion. On tlic other hand, if dcpcnds on the contcnt 01'8 block. II'thc contrast

the contrast is sinall (e.g., a smooth block), thc intensities can only hc tuned slightly. Expcrimcntal results show that tlic proposed algorithm is robust to common image processing opcrations such iis low-pass filtering, mcdian Ciltcring, rcsampling, requantization, aiid lossy JI'EG coniprcssion.

REFERENCES

l3. Schncicr, Applied Cryptography, 2nd Ettition, John Wilcy & Sons, I996

I). R. Stinson, Ctyptography: Theory and Prtrctice, CRC Press, 1995.

M. 0. Swanson, M. Kohayashi, and A. 11. Tcwfik, "Mullinictlia tlata-cmbctlding and watermarking teclinologics", Proceedings of'

//w

IEEE, Vol. 86, No. 6, pp. 1064-1087, Junc 1998,

W. Bender, U. Gruhi, N. Morinioto, and A. 1 .U, "Tcchniqiics for data hiding", 1Hh.l

Sy,ytenz.s Journd, Vol. 3 5 , Nos. 3&4, pp. 3 13-335, 1996.

R. G . van Schyndcl, A. Z. Tirkel, and C. F. Osbornc, "A digital watermark", in Proceeding.s uf ICIP'94, Vol. I , pp. 86-90,

1994.

I<. Matsui and K. Tanalta, "Vidco- stcganography: how to cmbcd a signaturc in a piclurc", i n Proceedings of IMA liitellectuul I'roper<y, Vol. 1, No. 1, pp. 187- 206, Jan. 1994.

G. Voyatzis and 1. Pitas, "Applications of toral automorphisms in image waterinarking", in Proceeding.s qf'CIP'96, Vol. 2, pp. 237- 240, 1996.

R. U. Wolfgang and E. J. Dclp, "A watermark for digital images", in Proceedings OfICIl'96, Vol. 3, pp. 219-222,

1996.

I. Pitas, "A method for signature casting on digital images", in Proceedings of' IUP'96, Vol. 3, pp. 215-218, 1990.

I . Pitas, " A mctliotl for watermark c;istiiig on digital images", lBBB Tmnsaction on

Circuits mid Sy.vtems f o r Video Technology, Vol. 8, No. 6, pp. 775-780, Oct. 1998.

Macq, "Spatial method for copyright labclitig o r digital images", in l'roceedings

of

/EBB Nonlinear Sigizul Prucassing W o h h o p , pp. 456-459, 1995,

[ 121 V. Darmstacdtcr, .I.#. I)claiglc, J. J . Quisquatcr, atid R. Maccl, "Low cost spatial watcrmarking", C u i i t ~ ~ z i / e r and Gniplzic.y, Vol. [I31 M. Kuttcr, F, Jordan and I:. Uosscn, "Digital

watcrmarlting o f color images using amplitutlc motlul;atioii", .lout.rial of

332, April 1998,

1

1411 1'. Ilavcrii anti M. Scolt, "Fractal based imagc stcgaiiography", in Procee~litzgs of

First 1iiter.ncrtionul Wurkslioyi on ln/brnic/tioii Ifiding, pp, 279-294, 1097.

[I51 J J K. O'fluanaidh, W J. Dowling, and I; M. Uoland, "Phasc watcrmarlting 01' digital images", in I'tweeedin~s

oi'

ICIP'96, Vol. 3, pp. 239-242, 1996.

[161 J J I<. O'lluanaidh, W J . L)owling, a i d t; M. Uoland, "Watermalting digital iiiiagcs for copyright protection", [Eh; Proceedings: Vision, Image & Signul I'roccxring, Vol. 143, No. 4, pp. 250-256, h u g . 1996.

[I71 I. J. Cox, J. Kilian, I:.

I.

Lcighlon, and T. Shanioon, "Scc~irc spread spcctrniii watcrmarking lor multimcd ia", / E R E

Trun.sactions oflrnage Proccxsiirg, Vol. 6 No. 12, pp. 1673-1687, Dcc. 1997.

1

I81 C.-'l', llsu and J.-I,. Wu, "I litldcn signatures in images", in /'roceediiigs ICII"96, pp. 223- 226, 1996.

191 C.-l, lisn and J.-I,. Wu, "UC1'-basctl watcrmarlting for video", /EEL Trirn.scictioir,y

on Consunier Electronic.s, Vol. 44, No. I , pp. 206-216, l'cb. 1998.

1201 C.-T, llsu ancl J.-L. Wu, "Ilitldcn digital watermarks in images", IEKh; kitisactions

on Image Processing, Vol. 8 , No. I , pp. 58- 68, 1999.

1211 W. l'ang and Y . Aolti, "A DCl-based coding of iinagcs in watcnuarlting", in Proceedings qf the International Confererrcc

on Injbrnmtion, Cotntrrzmicaliuns crnd jCigncrl

l'rucc..ssing, Vol. 1, pp. 5 10-512, 1997. 1221 M. D. Swansoti,

U.

Zh~i, ant1 A. 11. 'l'cwfik,

22, No. 4, {)I>. 417-424, 1998.

Blrclro/zic InzUging, Vol. 7, No. 2, 1>1). 326-

~ . ~ ~~~~~~

"'ffiinsparcnt robust imagc watcriiiarlting", i n Procecdii1g.r ojlCIP'96, Vol. 3, pp. 21 1-214,

1996,

1231 C. I'otlilcliiik antl W. Zcng, "Digital imagc watcri1i;irlting using visual inodcls", in

CIi;iiig-llsiiig L c c was born (111

JUIY 24, 1968 in Tainan, l'aiwan, Ilcpublic of China. Hc rcccivcd the U.S. ancl P1i.D. dcgrccs i n Coiiipukr antl Idormation Scicncc from N;itional Ch iiio Tung University i n 1991 a n d 1995,

I

24

125

I'i.ocecding.s

/

II

SI'II/IS& T Electronic It~icr~yiizg'97; Ilnmair Vision rrnd Electronic, Iniuging, Vol. 3022, pp. 3 10-321, J'cb. 1097.

J . Ohnishi and K. Matsui, "Embctlding a seal into a picttirc untlcr orlhogonal wavclct transform", i n IJroceedings qf Midfiinedia, pp. 514-521, 1996,

X.-G. Xia, C. G. Roncclct, and G. 11. Arcc, "A nrultircsolulion watermark for digital images", in Proceedings

.J'

lCII"97, Vol. 3, pp. 548-55 I , 1997.

1261 I). I<tindur and D. tlatzinalios, "A robust digital image w;itcrmarking method using wavelet-basctl fiision", in I'roceediigy of' ICll"97, Vol. I , pp. 544-547, 1997.

12711 1'. A. Wilson, S. I<. liogcrs, and L. I<. Myers, "Pcrccptnal hascd liypcrspcctral iinagc liision using niriltircsolution aiialysis", Opical liizgineeritzg, Vol. 34, N o . 1 I , pp. 3154-3164, N o v . 1995.

1281 H. Sklar, lIigi/ol (:OtiztiintiiciitiDns; Fundunieil/til.s cmd Appliccitiuirs, I'rcnlicc-

Yeuan-l<ueti Lee rcccivcd both worlted as a lcctiirc i n the Taiiisui Oxlbrd thc

13,s.

a n d M.S. dcgrccs i n Univcrsity College, Taipei. I IC is currcnlly a Pl1.D. Computer aiid Iiilhrmation student in the Dcpartmciit of Computer and Sciciicc lrom National Chiao Inroriiiatioti Sciciicc at National Chiao Tung 'Itiiig University, in l9X9 and University. I lis research interests include image I99 I , rcspcctivcly. From processing, image cryptography, and data hiding. September 19'93 to July 1995 Iic

(b) (21) (h)

Fig. 4 Iksull o r low-pass lilkring. ( a ) Low- pass l i l t e r c t l iiiiagc. (b) Extracted watermark

with N C 0 . 9 6 5 8 .

( e )

( 4

( a ) (b)

Fig. 3 Ai1 cxainplc to illustrate the proposed mcthod. (a) [lost image o f a i r c 5 12x5 12. (h) Biliary watermark image o r size 128x 128. (c)

Watertnarlccd image. (d) Uxtractcd binary watcriiiai-IC.

Fig, 5 llcsult of mcdian filtering. (a) Mcdian liltcrcd image. (b) F,xtractctl watcriiiarli with

5.3s (1.33 1.27

Fig. 8 I<csults ol'JI'IIG coiiiprcssion. (a) Jl'liC;

comprcsscd imagc wilh CR-7.27. (11) E x t ~ i c t c t l watcrmark with N C 0 . 9 9 2 9 , (c) JI'I:C; coinprcssctl image with CIl~=14.30. (cl)

Ixlraclctl walcrmark wilh NCI 0.0 103.

~~