Double Watermarking and Turbo Coding for Robust Image Watermarking

10 pages
37 views
of 10
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Share
Description
This paper is an attempt to describe the concept of double watermarking. The latter term refers to a new watermarking scheme based on embedding a mark (signature) in both spatial and multiresolution domains. This scheme is able of embedding 2000 bits
Transcript
   PROOF COPY [JTE101051] 001804JTE P    R   O   O   F     C   O   P    Y      [    J    T    E    1   0   1   0   5    1    ]     0   0   1   8   0   4   J    T    E      Chokri Chemak, 1  , 2  Mohamed Salim Bouhlel, 1 and Jean Christophe Lapayre 2 Double Watermarking and Turbo Coding forRobust Image Watermarking ABSTRACT: Thispaperisanattempttodescribetheconceptofdoublewatermarking.Thelattertermreferstoanewwatermarkingschemebased onembeddingamark(signature)inbothspatialandmultiresolutiondomains.Thisschemeisableofembedding2000bitsofmarkinmedicalimageswithdimensions256by256pixels.Experimentsonadatabaseof30medicalimagesindicatethewatermarksarerobusttonoises,filterattacks,JPEGcompression, and cropping. For the purpose of increasing the image watermarking robustness against attacks of an image transmission and to perform a large number of bits to hide into images we encode with a turbo code an image-embedded mark. Fidelity of images is improved byincorporation of the relative peak signal-to-noise ratio as a perceptual metric to measure image degradation. We demonstrate by some experimentalresults that this unit of measurement is the best distortion metric which is correlated with the human visual system to evaluate the quality of imagesafter the watermarking process. We show that each of these three components improves performance substantially. KEYWORDS: double watermarking, multi-resolution field, spatial field, turbo code, RPSNR, robustness Introduction Image watermarking allows owners or providers to hide an invis-ible and a robust signature inside images, often for security pur- poses and more precisely for particular owner or content authenti-cation[1,2].Medicinehasbenefitedfromwatermarkingresearchto preserve medical deontology [3,4] and facilitate distant diagnosis[5]. Watermarking is a solution to conserve the intellectual proper-ties of a diagnostic image, as well as to maintain the perceptualfidelity. There are three parameters in digital watermarking: data payload, fidelity, and robustness. The reader is directed to [6] for adetailed discussion of these concepts. Data payload can be defined as the number of useful bits that can be hidden in an image. Itshould be noted that, most of the time, data payload depends on thesize of the host image. Fidelity can be seen as the perceptual simi-larity between the srcinal images and watermarked ones. Themodifications introduced by the watermarking process should beimperceptible. Finally, robustness means that the retriever is stillable to recover the hidden mark even if the watermarked image has been altered after different attacks.In this paper we develop a new scheme of a robust image water-marking able to embed large data payloads, entitled “double water-marking.” This scheme attempts to attain an optimal trade-off among robustness data payload, and estimates of perceptual fidel-ity.Weproposetoembedthemarkintwodifferentfields:themulti-resolution domain and the spatial one. The resulting image will benamed a “double watermarked image.” This new approach willallow us to benefit from the advantages of both domains and to es-cape the drawbacks of each. The embedded mark is coded with anerrorcorrectingcode(ECC):theturbocode.Fortheaimofimprov-ing the perceptual fidelity after watermarking process, we use therelative peak signal-to-noise ratio (RPSNR) as a distortion metricto estimate image degradation.This scheme of image watermarking is able to embed 2000 bitsin medical images with 256 by 256 sized pixels. Results of experi-ments, carried out on a database of 30 medical images, will dem-onstrate that our double watermarking scheme is robust against dif-ferent attacks such as noises, filtering, JPEG compression, and geometric transformation such as cropping.In our paper, images are extracted from DICOM library, and weuse MATLAB and its library as software to experiment with our approaches and to validate with our results.The paper is organized as follows: We first introduce the prob-lem in watermarking field such as the embedding of mark in onefield, and briefly give an overview of previous countermeasures proposed in the literature; we follow that by describing the baseline(formal model) of our watermarking scheme. Furthermore, we ex- plain the srcinal contribution of the approach based on the choiceof the multi-resolution field with 5/3 wavelets image decomposi-tion as well as the choice of the spatial field in our new watermark-ing scheme and the utilities of using powerful ECC such as turbocode.The fidelity problem is introduced in “Perceptual Quality Met-rics” section, where we incorporate perceptual shaping, based ondistortion metric for perceptual image quality: the RPSNR. Thismetric is able to reduce the perceptual distance between attacked watermarked images and unmarked ones. We list some simulationresults showing that RPSNR is the best metric correlated with thehuman visual system (HVS) to evaluate image degradation after different attacks. The following section proceeds to present themain steps relevant to our new watermarking embedding algorithmand the detecting one. The “Preliminary Results” section reports aset of selected simulation results that clearly demonstrate that, evenafter rather severe addition of noise, filtering, and JPEG compres-sion, and cropping, all 2000 bits are correctly detected from water-marked images while preserving image fidelity after the differentattacks on images. Finally, in the last section we offer some con-cluding remarks reporting the performance of the proposed ap- Manuscript received February 12, 2007; accepted for publication February13, 2008. 1 Research Unit: Sciences and Technologies of Image and Telecommunica-tions, Higher Institute of Biotechnology of Sfax, Sfax University, Tunisia.Electronic mail: medsalim.bouhlel@enis.rnu.tn, [cchemak, jean-christophe.lapayre]@univ-fcomte.fr  2 Computer Science Laboratory of Franche-Comte (L. I. F. C), Franche-Comte University, France.  Journal ofTesting and Evaluation , Vol. 36, No. 4Paper ID JTE101051Available online at: www.astm.org 1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980 Copyright © 2008 by ASTM International, 100 Barr Harbor Drive, PO Box C700, West Conshohocken, PA 19428-2959.  1  PROOF COPY [JTE101051] 001804JTE   PROOF COPY [JTE101051] 001804JTE P    R   O   O   F     C   O   P    Y      [    J    T    E    1   0   1   0   5    1    ]     0   0   1   8   0   4   J    T    E       proach, but also its limitations against large geometrical attacks.We will also present an overview for our future work by giving asolution for image robustness against the geometric attack to ame-liorate our double watermarking scheme. Problem Context In this section, we begin by introducing the related works of our different ideas derived from the literature in order to lead to easyunderstanding and to compare the particular attributes of the ap- proach.Wefollowthatbydescribingthebaseline(formalmodel)of our watermarking scheme. Finally, we explain the srcinal contri- bution of the approach based on the choice of the multi-resolutionfield with 5/3 wavelets image decomposition as well as the choiceof the spatial field in our new watermarking scheme and the utilityof using a powerful ECC such as turbo code.  Related work  In this subsection, we present the outline of the classical methodsusing one or more field in the watermark process.Thus, the robust-nessofimagesdependsonthecharacteristicsofsuchfield.Nikolai-dis et al. [7] as well as Chemak et al. [8] propose to embed the mark in the spatial field to compensate for the geometrical deformationsinduced by cropping. Chemak et al. [8] use the second low symbol byte(LSB2)substitutionofpixelsinthespatialfieldtodemonstrate by some simulation results that embedding the mark in the spatialfield increases the image robustness against additive noises and cropping, but each embedding field is not robust against filteringattack or JPEG compression.Coxetal.[9]useaspreadspectrumtoembedawatermarkinthediscretecosinetransform(DCT)andprovethattheembeddingpro-cess in the DCT domain is robust against JPEG compression and image filtering attack. Hernandez et al. [10,11] use 2-D multi-pulseamplitude modulation and spread spectrum to embed bit sequencesin digital images and prove the watermark robustness in the DCTdomain against Gaussian noises, whereas Hsu et al. [12] demon-stratethattheDCTwatermarkingfielddomaingivesusfrequencieslocalization but not a spatial localization of image. Consequently,the image is not robust against geometric attacks such us cropping.Inoue et al. [13] demonstrate that embedding the mark after image wavelet decomposition is robust against JPEG compressionand filtering attacks. Furthermore, Jianming et al. [14] use an adap-tive threshold method to improve image robustness against Gauss-ian noises in wavelet domain. Chemak et al. [15] propose to embed the mark in a multi-resolution field by image 5/3 wavelet decom- position and prove the spatial localization of the multi-resolutionfield and therefore the robustness of the watermark against crop- ping in this field. However, in the multi-resolution field, the imageis not robust against additive noises [15]. Each embedding field ismainly robust against some image processing, but is still weak against other attacks. For this reason, some of the recent literaturesuch as Shih et al. [16] propose a combinational watermarkingimageinthespatialandDCTdomaintoallowustobenefitfromtheadvantages of both domains and to escape the drawbacks of each.However,theDCTdomainisnotrobustagainstgeometricalattackssuch as cropping [12]. Therefore, we need to combine the spatialfield with another field, such as the multi-resolution field that givesus spatial and frequencies localization of image, to assure imagerobustness against this attack. For this reason, in our watermarkingscheme we proposed to combine with the spatial field the multi-resolution one.What is more, it is quite easy to see that data payload, fidelity,and robustness are often in conflict. One may want to increase thewatermarking strength in order to enhance the robustness, but thisalso results in a more perceptible watermark. Further, one can in-crease the data payload by decreasing the redundancy of each hid-den bit, but this is counterbalanced by a loss of robustness. As aresult, a trade-off has to be found and it is often tied to the targeted application. For this reason, we incorporate an ECC: the turbo codein the formatting of the watermark to increase the number of bits tohide. In more recent studies, one can find watermarking techniquesthat use more powerful error correcting codes, such as convolu-tional codes [17,18], BCH codes [19–21], or concatenated codes based on a convolutional code followed by a Reed-Solomon code[22] or even convolutional turbo codes [23–25] to ensure that alarge number of bits to be hidden into images and to increase imagerobustness against some attacks. For example, Rey et al. [26] usethe concatenation of a BCH product code and a repetition code toincreasetherobustnessofimagewatermarkingagainstphotometricattacks such as additive noise or lossy compression. Relevant to theabove-mentioned aim, we incorporate the turbo code in the format-ting of the watermark to increase the number of bits to hide and toincrease image robustness against attacks.To keep image fidelity after embedding a large data payload, weneed a perceptual quality metric correlated with the HVS. Nowa-days, the most popular distortion measures in the image field arethe weighted peak signal-to-noise ratio (wPSNR) and the peak signal-to-noiseratio.However,itiswellknownthatthesedistortionmetrics are not correlated with the HVS [27]. For this reason, Reyet al. [27] proposed a novel measure adapted to the HVS: themasked peak signal-to-noise ratio (MPSNR) to compensate for theweaknesses of usual image quality metric.Wang et al. [28] proposethe index of Structural SIMilarity (SSIM) as a perceptual metricnear to HVS characteristic based on the structural information onimage to evaluate image quality. In our paper, we will use theRPSNR to estimate image degradation and to keep image fidelityafter embedding a large number of bits into images.The RPSNR isused byTao et al. [29] to estimate the image quality of video trans-mitted over a network. Our scheme is perceived as a new one in thatit combines the spatial and the multi-resolution field in the embed-dingprocess.Themarkcanbeextractedfromthespatialfieldorthemulti-resolution one and images will be robust against noises,JPEG compression, filtering, and cropping. Our watermarkingschemeallowsattainmentofanoptimaltrade-offbetweendatapay-loads and fidelity by using the turbo code to increase the number of  bit to hide into images and incorporating the RPSNR, which is cor-related with the HVS as a metric to evaluate image degradationafter the watermarking process.  Formal Model of theWatermarking Scheme In this subsection, we briefly describe the baseline of our water-marking algorithm. The main steps of our watermarking schemewill be introduced in “The Double Watermarking Algorithm” sec-tion. The considered approach is derived from the advantages of combining more than one field in the watermarking process, suchas the spatial and the DCT field [16].The main idea is to embed themark in the spatial and multi-resolution fields to ensure watermark robustness against usual attacks such as JPEG compression, filter- 2  JOURNAL OF TESTING AND EVALUATION 81828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194  PROOF COPY [JTE101051] 001804JTE   PROOF COPY [JTE101051] 001804JTE P    R   O   O   F     C   O   P    Y      [    J    T    E    1   0   1   0   5    1    ]     0   0   1   8   0   4   J    T    E      ing, noises, and cropping, and to escape the drawback of the use of one field in the embedding process. The formatted model of thespatial watermarking scheme and the multi-resolution scheme aredocumented in [8,30]. The approach provided in these last refer-ences gives us the advantages and the drawbacks of each of the twofields and the principle of the embedding scheme adopted in eachdomain.The extraction mode is blind. The embedded mark is extracted from the multi-resolution field by an inverse 5/3 wavelet decompo-sition using the formatted model explained in [30]. After that, it isextracted from the spatial field by the formatted model presented in[8]. Use of Multi-resolution and Spatial Fields in the DoubleWatermarking Scheme In this subsection, we will describe the srcinal contribution of combining the spatial and the multi-resolution fields. This idea isdue to the characteristic of each field of watermarking. For this rea-son, we will present the aim of using the multi-resolution field bythe 5/3 wavelets image decomposition and the aim of using the spa-tial field to guaranty the robustness of the approach against usualattacks: JPEG compression, filtering, noises, and cropping. On theone hand, our choice of investigating the 5/3 wavelet decomposi-tion in our double watermarking algorithm is motivated by manyreasons:• The 5/3 wavelet decomposition is an integer-to-integer trans-formadaptedforJPEG2000compressionandcomesinfrontfor its frequent use in JPEG2000 norm. Consequently, theimage watermarked is robust against JPEG compression at-tacks [30,31].• The 5/3 wavelet is used for its conservative characteristic[32]. This field gives us spatial and frequency localization.For this reason, the image keeps spatial contain after a 5/3wavelet decomposition. Fortunately, the image will be robustagainst geometrical attacks such as cropping.To keep image fidelity after the watermarking process, we need a perceptual metric correlated with the HVS. Nevertheless in themulti-resolution field image decomposition in sample bands is near to perception canal decomposition, so we can easily choose a psycho-visual model to measure image degradation [3].The 5/3 wavelets image decomposition equations introduced byLe Gall are:  d   n   =  d  0  n   −  ⌊ 1 2  d   n   +  d   n  − 1  ⌋   1   s  n   =  s 0  n   + ⌊ 1 4  d   n   +  d   n  − 1   +12 ⌋   2  where  d   n   is the lowpass subband signal,  s  n   is the highpass sub- band signal,  s 0  n  =  x  2 n  ,  d  0  n  =  x  2 n +1  , and   x  n   is the inputsignal.We use an embedding function (or secret key) to hide informa-tioninsideimages.Thefollowingfunctionisusedinourwatermark image: Y  i  =  X  i  1 +   W  i   (3)where  Y  i  is the watermarked image,  X  i  is the srcinal image,  W  i represents bits of information to be embedded composed of 1000 bits coded with 1/2 ratio turbo coder, and     is the visibilitycoefficient.On the other hand, using the spatial domain in our paper is mo-tivated by the fact that the image watermarked is robust againstgeometrical attacks such as cropping [33,8], contrary to the fre-quencies field, where the image is very sensitive to geometric trans-formations. In this field, the image is also robust against additivenoise [8]. Use ofTurbo Code in the DoubleWatermarking Scheme In this subsection, we will describe the srcinal contribution of using the turbo code. We begin by presenting the turbo code struc-ture, after that we present the hypothesis of using such an ECC.The system transmission of turbo coder or parallel concatena-tion code consists in setting in parallel form of recursive systematicconvolutional coders (RSC) C 1  and C 2  [32].The structural diagramof turbo codes is represented in (Fig. 1) [32], where d  k   is the input bitinformationandX k   issystematicoutputbitformingcodewords.Y 1k   and Y 2K   are the parity bits coming, respectively, from the firstand the second recursive coders after interleaving the systematic bits or input bits.Turbo codes have been successfully used in digital communica-tion systems in order to achieve reliable transmission on a noisychannel [34–36].The idea of using turbo code in our double watermarking algo-rithm scheme comes from the efficiency of the ECC like the prod-uct code (concatenation of a BCH product code and a repetitioncode) to increase robustness against noises, JPEG compression and geometrical attacks [26].Furthermore, the incorporation of turbo code in the formattingof the watermark increases the number of bits to hide, since thenumber of repetitions of each bit of the watermark decreases in thesame proportion [26].In this paper, we use turbo code using the soft output Viterbealgorithm (SOVA) for decoding techniques [18] to achieve higherror-correction capability with reasonable decoding complexity.For the mentioned aim, turbo code is an attractive solution for thisapplication as it achieves powerful error-correction capability and higher payloads. Perceptual Quality Metrics Inthissection,wefirstexplaintheoriginalcontributionofusingtheRPSNR as metric to evaluate image quality in our double water- FIG. 1—  Structural diagram of turbo coder. CHEMAK ET AL. ON NEW WATERMARKING SCHEME FOR SECURITY AND TRANSMISSION  3 195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287  PROOF COPY [JTE101051] 001804JTE   PROOF COPY [JTE101051] 001804JTE P    R   O   O   F     C   O   P    Y      [    J    T    E    1   0   1   0   5    1    ]     0   0   1   8   0   4   J    T    E      marking by making a comparison between other usual perceptualmetrics used in the literature: the peak signal-to-noise ratio (PSNR)and the weighted peak signal-to-noise ratio (wPSNR). After that,we demonstrate by some experimental results that this unit of mea-surementisthebestdistortionmetrictoevaluateimagedegradationafter different attacks. Nowadays, the most popular distortion measures in the imagefield are the PSNR and the wPSNR. They are usually measured indecibels.The PSNR is derived from the mean square error (MSE).The equations are the following:  MSE   =1  MN    m =0  M  −1  n =0  N  −1   x  m , n   −  y  m , n  2 (4)where  M   and   N   are, respectively, the numbers of pixels lengthwiseand widthwise per image.  x  and   y  are, respectively, the gray scale of the srcinal image and the degraded image.  PSNR  = 10 log 10   X  max2 /  MSE    (5)where  X  max  is the maximum luminescence in the image.The weighted peak signal-to-noise ratio is derived from theweighted mean square error. The equations are the following: wMSE   =1  MN    m =0  M  −1  n =0  N  −1   x  m , n   −  y  m , n  1 + Var   m , n    2 (6)where  M   and   N   are, respectively, the numbers of pixels lengthwiseand widthwise per image,  x  and   y  are, respectively, the gray scale of the srcinal image and the degraded image. wPSNR  = 10 log 10   X  max wMSE    (7)where  X  max  is the maximum luminescence in the image.Thepopularityofthesetwometricsisverylikelyduetothesim- plicity of the metric. However, it is well known that these distortionmetrics are not correlated with the HVS [27].This might be a prob-lem for their application in digital watermarking since sophisti-cated watermarking methods exploit in one way or another theHVS. In addition, using the above metric to quantify the distortioncaused by a watermarking process might therefore result in mis-leading quantitative distortion measurements [37]. Furthermore,thesemetricsareusuallyappliedtotheluminanceandchrominancechannels of images [38], and they give a distortion value for allcolor channels.In this paper, we introduce the relative peak signal-to-noise ratio(RPSNR),whichhasnorelationtothecontentcharacteristicsoftheimage that fits to the HVS and therefore is more suitable for digitalwatermarking. In addition, this metric allows comparison even if the distortion is in a different color channel [39]. The estimationerror for RPSNR is a function of packet loss rate and average loss burst length metric, which represents path quality under differentloss patterns. The RPSNR is used to evaluate the image quality bycalculating the relative mean square error (RMSE) between the im-ages to compare. The equations are the following:  RMSE   =1  MN    m =0  M  −1  n =0  N  −1  2   x  m , n   −  y  m , n   x  m , n   +  y  m , n   2 (8)where  M   and   N   are, respectively, the numbers of pixels lengthwiseand width wise per image,  x  and   y  are, respectively, the gray scaleof the srcinal image and the degraded image.  RPSNR  = 10 log 10   X  max2  RMSE    (9)where  X  max  is the maximum of luminescence in the image.InordertoproperlydemonstratetheperformanceoftheRPSNR in our watermarking scheme and to allow a fair comparison be-tween different perceptual quality metrics, the setup test conditionsare of crucial importance. Table 1 lists different mean values of PSNR, wPSNR, and RPSNR after the most famous types of distor-tionandattacksonadatabaseof30,256by256pixel-sizedmedicalimages. Figure 2 presents some images from the above-mentioned database. The DoubleWatermarkingAlgorithm In this section, we present the different steps of our embedding and detecting image watermarking algorithm. In this paper, we proposeto embed the mark in two different fields.The selected ones are thespatial domain and the multi-resolution one. Here we present the principleoftheembeddingalgorithmandthemainstepsrelevanttoour mark blind detection algorithm. The DoubleWatermarking EmbeddingAlgorithm In the embedding scheme, we start by embedding the mark in themulti-resolution field. For this reason, we begin by decomposingmedical image in 5/3 wavelet decomposition. The 5/3 waveletsimage decomposition decomposes medical images (e.g., Fig. 3( a ))on sample bands (low frequency (LF), medium frequency (MF),and high frequency (HF)) as is shown in (Figs. 3( b ) and 3( c )).After that, pixels with a high intensity are selected from the medium-zone frequencies (MF). Information is then coded with turbo codeand embedded with embedding function (secret key) defined in Eq.3. After the embedding process, we rebuilt the image with an in-verse5/3waveletsimagedecompositionandweobtainedtheimagewatermarked in the multi-resolution field (see Fig. 3( d  )). TABLE 1—   Different means values of PSNR, wPSNR, and RPSNR after differ-ent attacks on image banks. Distortion type PSNR wPSNR RPSNR Mean shift 24.6090 35.6873 64.3123Contrast stretching 24.6003 35.7453 57.913Impulsivesalt and paper noise24.6499 35.4654 63.6996Multiplicative speckle noise 24.6186 34.6206 66.9165Additive Gaussian noise 24.5906 35.5855 62.4892Fuzzy logic 24.6054 45.0642 63.4734JPEG compression 24.7849 38.7652 62.6179FIG. 2—  Some figures from the database of 30 medical images. 4  JOURNAL OF TESTING AND EVALUATION 288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367  PROOF COPY [JTE101051] 001804JTE   PROOF COPY [JTE101051] 001804JTE P    R   O   O   F     C   O   P    Y      [    J    T    E    1   0   1   0   5    1    ]     0   0   1   8   0   4   J    T    E      In the second step, we embed the mark in the spatial field. For this raison, we select pixels with high intensity from the image re- built (see Fig. 3( e )). After that, we embed the mark with a second low symbol byte (LSB2) substitution of pixels to be marked. Fi-nally, we obtain the double watermarked image (see Fig. 3(  f   )). The DoubleWatermarking Blind DetectionAlgo-rithm The detection scheme involves two steps:In the first step, we have to extract the mark from the multi-resolution field. Therefore, we decompose medical image in 5/3wavelets image decomposition. After that, pixels with the high-intensity are selected from the medium-zone frequencies (MF).Wethen extract the mark from the multi-resolution field with binaryinversetransformation.Theextractedmarkisdecodedbytheuseof the SOVA. We apply a correlation between the extracted, decoded mark and a dictionary of marks. The latter term refers to a bank of 800 marks containing the srcinal mark embedded into image.Each signature of this bank is supposed to be the information em- bedded in the image. We can determine the embedded mark bycomputing the correlation between the extracted, decoded mark andeveryelementofthebank(dictionary).Thiscomparisonallowsus to identify the degree of similarity between extracted, decoded mark and the srcinal one embedded inside images. If the referencemark (srcinal mark) presents the high value of correlation with theextracted and decoded mark, we can say that the detection of mark has succeeded. However, if the reference mark has not the maxi-mum of correlation with the extracted and decoded mark, detectionof the mark is lost.In the second step, we have to extract the mark from the spatialfield. For this reason, we select pixels with high intensity from themedical images.After that, we extract the mark with detected func-tion:  W  i =  Y  i  mod 4  /2. We then decode the mark using the SOVA.Finally, we compare the mark extracted and decoded with the dic-tionary of marks by the use of the correlation method.To validate robustness of image against attacks, we should choose a mark owner that refers to succeeded correlation after amulti-resolution extract scheme or a spatial extract one (see Figs.4( a )–4( d  )). If we have a successful correlation in every one of thetwo fields, we should choose the maximum of correlation betweenthem(seeFigs.4( i )and4(  j  )).Ifthedetectiondoesnotsucceedafter the two extractions from the two fields, we can say that the medicalimagehaslostitsauthenticityanditisnotexploitable.Applicationsof our detection technique are demonstrated with some simulationresults in the next section. Preliminary Results Firstofall,thissectiondisplaystheexperimentalresultscarriedouton a 30 images from our database. The figures below give a sum-mary of the test results after different attacks of the correlation be-tween extracted, decoded signature and the dictionary after spatialfield detection and multi-resolution field oneThe images are tested after attacks such as noises (Gaussian, additive, and speckle), fil-tering, JPEG compression, and cropping in order to validate thesuggested approach. The mark is extracted from the multi-resolutionandthespatialfields.Theconsideredmarkistheonethathas the successful detection from one of the two fields or that hasthemaximumcorrelationifwehaveasuccessfuldetectionfromthetwo fields.Subsequently, we demonstrate by some simulations results thefidelity of images after the different attacks to be considered bymeasuring the RPSNR between the double watermarked image and the unmarked one after each attack for all images of our database.  RobustnessAgainstAttacks After Gaussian noise and a speckle noise attacks, the marks aredetected from the multi-resolution field shown to us in (Figs. 4( a )and 4( e )). In this way, we explore the advantages of multi-resolution field combined with turbo code to maintain image ro- bustness after Gaussian noise and speckle noise attacks (confirmed  FIG. 3—  (a) Original image. (b) 5/3 image wavelet decomposition. (c) Image decomposed in LF, MF, and HF after a 5/3 image wavelet decomposition. (d) Recon- structed image after an inverse 5/3 image wavelet decomposition. (e) Pixels with a high intensity chosen from image in spatial field. (f) Image double watermarked. CHEMAK ET AL. ON NEW WATERMARKING SCHEME FOR SECURITY AND TRANSMISSION  5 368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434  PROOF COPY [JTE101051] 001804JTE
Related Search
Similar documents
View more...
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks