IEEE Paper Template in A4 (V1) - International Journal of Computer ...

International Journal of Computer Applications in Engineering Sciences[VOL III, ISSUE I, MARCH 2013] [ISSN: 2231-4946]Medical Image Watermarking using MultiRidgelet and Fast ICAA. Umaamaheshvari 1 , K.Thanushkodi 21 Department of ECE, SSEC, Anna University of TechnologyCoimbatore, India2 ACET, Tamilnadu, India1 ums612@gmail.comAbstract— The medical field along with computer fieldgrows in a rapid manner. In telemedicine and telediagnosisauthentication of medical images are veryimportant to the patient ,doctor and Insurance company.The youngsters are also keen in knowing about thetreatment methods and procedures. But no one likes to tellabout the disease in open hall. So there is always an urgefor watermarking of medical images. So in this work weuse the concept of multi ridge let and fast independentcomponent analysis. Since multi ridge let is a good way toenhance edges and reduce the noise. In this method wehave calculated the structural similarity index measure(SSIM), mean square error(MSE) and peak signal to noiseratio(PSNR) .The experimental results are compared withthe existing methods and found to be very effective.Keywords— fast ICA, mean square error(MSE) , multiridgelet , peak signal to noise ratio(PSNR) structuralsimilarity index measure (SSIM),telemedicine , telediagnosisI. INTRODUCTIONMultimedia and communication technologyprovided new ways to store, access and distributemedical data in a digital format. Watermarking usuallyconsists of the use of perceptually invisibleauthentication techniques. That is a predefined―controlled‖ distortion is introduced in the multimediadata. The goals are i) verification of the owner and ii)the detection of forgeries of an original image. MedicalImagery is a field where the protection of theintegrity and confidentiality of content is a critical issuedue to the special characteristics derived from strictethics, legislative and diagnostic implications [1].Medical images should be kept intact in anycircumstance and before any operation they must bechecked for [2].Many classifications are available for digital imagewatermarking. The common taxonomy is embedding inspatial and frequency domain. Spatial domain methodsare not robust against various attacks and are lesscomplex [3,4].In spatial domain the intensities of pixelare modified and embedded. In transform domain whenimage is inverse transformed watermark is distributedirregularly and this makes the attacker confused.Generally transform domain methods are superior tospatial domain methods. Some of the transform domainmethods are Fourier transform [5], Discrete cosinetransform, fast Fourier transform, wavelet transform[6].The transform domain coefficients are altered to embed thewatermark and finally inverse transform is applied to obtainthe watermarked digital data.Schyndel et al. [3] have proposed two methodsusing bit plane manipulation of the LSB and the othermethod based on the linear addition of the watermark tothe image data, which is more difficult to decode,offering inherent security. The watermarking schemeemployed in spatial domain using hash functions and insome works are presented in the most popularwatermarking schemes based on the Spread Spectrum.The first k highest magnitude DFT/DCT coefficients ofthe image is used for watermarking and extraction isdone by comparing the DFT/DCT coefficients of thewatermarked and the original image. Shinfeng et al.[6]proposed a robust DCT technique watermark isembedded in low frequency. Barni et al. [7] haveproposed a watermarking algorithm, which operates inthe frequency domain, embeds a pseudo-randomsequence of real numbers in a selected set of DCTcoefficients. The fractional Fourier transform basedwatermarking scheme for the multimedia copyrightprotection is also done .After decomposing the imagevia FRFT, transformation coefficients are reordering innon increasing sequence and the watermark is embeddedin the middle coefficients. Feng et al. [8] have proposeda blind watermarking algorithm in which multiple chirpsare used as watermark and embedded in the spatialdomain directly but detected in the FRFT domain. Yu etal. [9] have used the same logic proposed by Feng et al.[8], the only difference is that the embedding is done inFRFT domain where watermarkPosition and the transform order are used as theencryption keys. Xia et al. [10] have added apseudorandom sequence to the largest coefficients of thedetail bands where perceptual considerations are takeninto account by setting the amount of modificationproportional to the strength of the coefficient itself.Watermark detection is achieved through comparison9 | P a g e

Umaamaheshvari et. al.with the original un-watermarked image. Barni et al.[11] proposed a method based on the characteristics ofthe human visual system operating in wavelet domain.Based on the texture and the luminance content of allimage sub-bands, a mask is accomplished pixel by pixel.Kundur et al. [12] proposed the use of gray scale logo aswatermark. They addressed a multi resolution fusionbased watermarking method for embedding gray scalelogos into wavelet transformed images via saliencefactor. Wang et al. [13] and Zhang et al. [14] proposed anew watermarking algorithm based on wavelet treequantization. The detailed survey on wavelet basedwatermarking techniques can be found in [15].A multi ridge let transform along with fast ICA isproposed in this paper. The experimental results of theproposed method are compared with the alreadyavailable watermarking paper.In the organization of the paper the history ofwatermarking methods is given in section I , a briefnote on ridge lets is given in section II, a brief overviewof fast independent component analysis is given insection III, Watermark embedding and extraction isgiven in section IV. Results and conclusion are given insection V.II.RIDGELET TRANSFORMRidgelet transform deals with line or super-planesingularities. The general overview is given below.Let Ψ be in L2(R) with sufficient decay andsatisfies the admissibility condition,KΨ := ∫ |Ψˆ(ξ)| 2 /|(ξ) 2 dξ ⊂ ∞ ……………..………..(1)That KΨ = 1.For a> 0,b ∈ R and Θ ∈[0,2π].In 2-D, points and lines are related via the Radontransform, thus the wavelet and ridgelet transforms arelinked via the Radon transform. More precisely, denotethe Radon transform asRf( Θ ,t)=∫f ( x1,x2, ) δ( x1cos Θ +x2 sin Θ −t) dx 1 dx 2 …….(2)then the ridgelet transform is the application of a 1-Dwavelet transform to the slices (also referred to asprojections) of the Radon transform.CRTf (a, b, Θ)=∫Ψ a,b (t)R f (θ,t)dt………………………….(3)For f ∈ L1 (R2 ) is represented as a continuoussuperposition of Ridgelet coefficient and is given byRf(a,b, Θ)= ∫ f(x)Ψa,b,Θ(x)dx ……………………..(4)Any Function f ∈ L1(R2 ) ∩ L2(R)2 is represented as acontinuous superposition of Ridgelet functions.F(x) = ∫∫∫Rf(a,b,θ)Ψ a,b, θ(x)da…………………………(5)A. Discrete Ridgelet TransformA continuous ridgelet transform is calculated byapplying 1D wavelet transform to the slices of radontransform R f (θ). In radon transform a famous projectionslicetheorem is used. This theorem says that the Radontransform can be obtained by applying the onedimensionalinverse Fourier transform to the twodimensionalFourier transform of function restricted toradial lines through the origin. The relation among theFourier, radon and ridgetet domain is depicted shown inFig. 1.Fig .1 Relations between DomainsTo complete the ridge let transform, apply a onedimensional wavelet transform along the radial variablein Radon space. The sum up of above procedure isshown in Fig. 2 in the form of flow chart. The DRT ofan image of size n×n is an image of size2n×2n,introducing a redundancy factor equal to 4[25,26].F(ωcosθ,ωsinθ) = ∫R f (θ,t)e -2πiωx dt ………………..(6)B. Multiscale Ridgelet TransformMultiscale ridgelets based on the ridgelet transformcombined with a spatial bandpass filtering operation toisolate different scales .Algorithm‣ Apply a trous algorithm with M scales‣ Apply Radon Transform on detail sub - bands of Mscales.‣ Calculate the ridgelet coefficients by applying‣ 1D wavelet transform on radon coefficients.‣ Get the multiscale ridgelet coefficient of M scales.III.RidgeletdomainRadon DomainFAST INDEPENDENT COMPONENT ANALYSISIndependent component analysis is a method ofsignal processing and used for data analysis. It is basedon blind source separation (BSS) method. Without thesource signals the coefficients are used for the recoveryof the signal.ICA was introduced by Comon, Pierre [16].Hyvarinen proposed famous Fast ICA based on negativeentropy principle [17].A. The Mixing Model of ICA2D Fourierdomain10 | P a g e

Medical Image Watermarking using Multi Ridgelet and Fast ICASourcesS 1S 2S NMixingmatrixAFig.3 Watermark EmbeddingBDemixingmatrix3.The iterative algorithm finds the direction of weightvector W maximizing the non-gaussianity of theprojection W T X for the data X.4. Randomize the initial weight vector W.5. letFig.2 Mixing Model6. Let7.If not converged go to step 5.imageMultiWat Inversridermegelarkridgelet Fig.4 Watermark emb etExtractiontraeddi ktransfnsfng eormoryWatermark mFAST ICAed image WATERMARKING METHODIV.B. Watermark EmbeddingIn the proposed method ,the image and watermarkis decomposed by multiscale ridgelet transform. Thecoefficients of mulriscale ridgelet for watermarkimageW MR is embedded in the low resolution part of theimage I MRI MR (low) = I MR (low) (1+αW MR (low)) .........................(7)From the spectrum analysis of the image most ofthe useful information is located in the low resolutionpart of the image. The invisibility of watermark isenhanced by using proper value for α. On the contrary,in case of attacks low resolution part of the image ispreserved in the low representation of the image ,makingthe watermark robust.The other coefficients are embedded in the highfrequency components of the image representing theedges and textures. To increase the robustness of thewatermark we use the equationI MR (hi)= I MR (hi) + βW MR (hi) .........................................(8)To maintain the invisibility of the image.C. Watermark ExtractionFor extraction process Fast ICA is used.1. Before applying fast ICA the input vector datashould be centered.2. The data should be centered by computing themean of each component of X and subtracting thatmean. This has the effect of making each componenthave zero mean.XX-E(X)C. Quality Measures of proposed methodAny new method is evaluated based quality measures.For this new method we have used four qualitymeasures.i)Mean Square ErrorMSE =...........(9)ii) Peak Signal to Noise RatioThe ratio between the maximum power of a signaland the power of corrupting noise .It is defined via themean squared error for which two I*J images f and zwhere one of the images is noise and the other is givenasPSNR is given by..............(10)iii) Structural Index Similarity Measure(SSIM)The extracted watermarks can be compared withoriginal watermark subjectively. Beside subjectivelyjudgment for the watermark fidelity, we have defined anobjective measure of similarity between the originalwatermark and the extracted watermark in the followingway:...............(11)V. CONCLUSIONSThe computer tomography, MRI and ultrasound imagesare taken and watermarked. We have taken 50 images ineach image and studied. The Gaussian, rotation, scaling,salt key and pepper noise and speckle noise are added andthe peak to signal noise ratio is calculated. The11 | P a g e

S.NoAttacksRT (LenaImage)Ref 18RT(Medicalimage)proposedRT (LenaImage)Ref 18RT(Medicalimage)proposedUmaamaheshvari et. al.corresponding structural similarity index and meansquare error are also calculated and found to besignificantly higher. So we can conclude that a abovemethod gives a acceptable result when compared withthe previous techniques.CT ImagesPERFORMANCE MATRICESTABLE 1: FOR CT IMAGESIMAGE3 58.5229 0.0914 0.9973IMAGE4 8.5978 0.0898 0.9972IMAGE5 59.7017 0.0696 0.9981TABLE II: FOR MRI IMAGESULTRAImagesPSNR MSE SSIMIMAGE1 59.2181 0.0779 0.9979IMAGE2 58.5565 0.0907 0.9973MRIImagesPSNR MSE SSIMMAGE1 67.9157 0.0105 0.9997IMAGE2 71.3853 0.0047 0.9996IMAGE3 71.1752 0.0050 0.9989IMAGE4 69.9087 0.0066 0.9997IMAGE5 67.9319 0.0105 0.9997TABLE III: FOR MRI IMAGESPSNR MSE SSIMIMAGE1 80.4523 0.0006 0.9993IMAGE2 72.3905 0.0037 0.9998IMAGE3 70.4234 0.0059 0.9996IMAGE4 70.6903 0.0055 0.9990IMAGE5 77.1617 0.0013 0.9990Ct images PSNR MSE SSIMGaussian 56.4556 0.0798 0.9046Rotation 56.4679 0.0773 0.9597Scaling 55.4093 0.0793 0.9505Salt and pepper 57.2532 0.0792 0.9791speckle 56.2267 0.0787 0.9754MRI images PSNR MSE SSIMGaussian 67.9985 0.0032 0.9265Rotation 66.9370 0.0026 0.9316Scaling 66.9895 0.0027 0.9316Salt and pepper 67.5509 0.0002 0.9297speckle 67.7110 0.0001 0.9419ultra images PSNR MSE SSIMGaussian 72.3957 0.0023 0.9934Rotation 72.8094 0.0029 0.9475Scaling 70.9152 0.0002 0.9162Salt and pepper 72.0169 0.0006 0.9009speckle 72.8900 0.0026 0.9724TABLE IV: COMPARISONTABLESCREEN SHOT:TABLE V: EFFECT OF DIFFERENT NOISES ON DIFFERENT IMAGESPSNRSSIM1. Gaussian22.32 67.9985 0.9001 0.92652. Rotation23.18 66.9370 0.8764 0.93163 Scaling - 66.9895 - 0.93164 Salt andpepper21.29 67.5509 0.8884 0.92975 speckle - 67.7110 - 0.941912 | P a g e

Medical Image Watermarking using Multi Ridgelet and Fast ICAREFERENCES[1] J. B.Planitz and A. Maeder ―Medical Image Watermarking: Astudy on Image Degradation‖ e- Health research Center, ICTCSIRO, Brisbane, QLD 4000[2] J. Nayak, P. Subbanna Bhat, R. Acharya U and Niranjan UC―Simultaneous storage of medical images in the spatial andfrequency domain: A comparative study‖ BioMedicalEngineering OnLine, pp. 3-17, 2004[3] Schyndle, R.G.V., Tirkel, A.Z., Osbrone, C.F.: A DigitalWatermark. In: Proc. of IEEE Int. Conf. on Image processing,vol. 2, pp. 86–90 (1994)[4] Hwang, M.S., Chang, C.C., Hwang, K.F.: A watermarkingtechnique based on one-way hash functions. IEEE Transcationson ConsumerElectronics 45(2), 286–294 (1999)[5] Djurovic, I., Stankovic, S., Pitas, I.: Digital watermarking in thefractional fourier transformation domain. Journal of Networkand Computer Applications 24(4), 167–173 (2001)[6] Shinfeng D. Lin, Chin-Feng Chen, ―A robust DCT-basedwatermarking for copyright protection‖, IEEE Trans.Consumer Electron. 46 (3) (2000) 415–421[7] Barni .M,Bartiloni F , Cappellini, V., Piva, A.: A DCT DomainSystem for Robust Image watermarking. Signal Processing66(3), 357–372 (1998)[8] Feng, Z., Xiaomin, M., Shouyi, Y.: Multiple-chirp typed blindwatermarking algorithm based on fractional Fourier transform.In: Proc.of Int. Sym. on Intelligent Signal Processing andCommunicationSystems,[9] Yu, F.Q., Zhangi, Z.K., Xu, M.H.: A Digital WatermarkingAlgorithmfor Image Based on Fractional Fourier Transform. In: Proc. ofFirstIEEE Conf. on Industrial Electronics and Applications,pp. 1–5(2006)[10] Xia, X., Boncelet, C.G., Arce, G.R.: A multiresolutionwatermark fordigital images. In: Proc. Fourth IEEE Int. Conf. on ImageProcessing, Santa Barbara, CA, vol. 3, pp. 548– 551 (1997)[11] Barni, M., Bartiloni, F., Piva, A.: Improved wavelet basedwatermarkingthrough pixel wise masking. IEEE Transcations on ImageProcessing10, 783–791 (2001)[12] Kundur, D., Hatzinakos, D.: Towards robust logo watermarkingusingmeltiresolution image fusion. IEEE Transcations on Multimedia6,185–197(2004)[13] Wang, S.H., Lin, Y.P.:Wavelet tree quantization for copyrightprotectionwatermarking. IEEE Transcations on Image Processing 13(2),154–165(2004 )[14] Zhang, X.D., Feng, J., Lo, K.T.: Image watermarking using treebasedspatial-frequency feature of wavelet transform. J.Visual Comm.Image Representation 14, 474–491 (2003)[15] Umaamaheshvari A and Thanushkodi .K ―Survey ofwatermarkingalgorithms for medical images .IJETT,vol.3 pp 401-410,2012.[16] Comon, Pierre(1998) Independent componenet analysis A newconceptSignal Processing, 36(3):287– 314[17] Hyvärinen, A.; Karhunen, J.; Oja, E. (2001): IndependentComponentAnalysis, New York: Wiley, ISBN 978-0-471-40540-5[18]Maggayarkarasi P. And Arulselvi .S:A new Digital imageWatermarkingbased on Finite Ridgelet transform and Extraction using ICA.Proceedings of ICETECT (2011)13 | P a g e