Oblivious Image Watermarking in Discrete Multiwavelet Domain using QIMM

Document Sample
Oblivious Image Watermarking in Discrete Multiwavelet Domain using QIMM Powered By Docstoc
					JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011                                                                   359

           Oblivious Image Watermarking in
      Discrete Multiwavelet Domain using QIMM
                      N.Leelavathy1, Research Scholar, J.N.T.University, Kakinada, India
                                            Email: nleelavathy@gmail.com1

                E. V. Prasad2, Professor, C.S.E. Department, J.N.T.University, Kakinada, India

             S.Srinivas Kumar3, Professor, E.C.E. Department, J.N.T.University, Kakinada, India

            B. Chandra Mohan4, Professor, E.C.E. Department, Bapatla Engineering College, India
             Email: {, samay_ssk2@yahoo.com3,}

Abstract— This paper proposes a novel oblivious image        transform (DCT) [5-8], Hadamard Transform (HT),
watermarking technique for the copyright protection and      Discrete Wavelet Transform (DWT) [9-11], and
authentication of still images, based on Discrete            Contourlet Transform (CT) [12-14].
Multiwavelet Transform (DMWT) and Quantization Index            DWT has a good property of localization in time and
Modulus Modulation (QIMM). The multiwavelet
                                                             frequency. Hence, the wavelets are useful tool in digital
coefficients of an image are quantized using QIMM and
the coefficients are modified according to the binary        image processing and its applications.
watermark logo. Watermark extraction does not require           Quantization Index Modulation (QIM) data
the original image. Experimental results show that, the      embedding methods [2-4] are proved to be attractive in
proposed watermarking scheme is robust to image              a practical and theoretical engineering perspective. In
compression and rotational attacks. It is superior to the    QIM schemes, the amplitude of a vector whose entries
method proposed by Lin et al. [10] in terms of Peak Signal   are pixels or frequency coefficients are quantized using
to Noise Ratio (PSNR) and Bit Error Rate (BER).              a quantization lattice. While this approach provides a
                                                             gain in the watermark capacity over other content-
                                                             dependent schemes, QIM based digital watermarking
Index Terms— Oblivious image watermarking, Copyright
                                                             offers less robustness compared to QIMM.
protection,   Authentication, Discrete  multiwavelet,
Quantization Index Modulus Modulation.
                                                                There are various methods which divide the DWT
                                                             coefficients into blocks. These methods use significant
                   I. INTRODUCTION                           coefficient from each block to embed watermark bit.
                                                             The position of the significant coefficient is important in
   The reproduction and distribution of unauthorized         the extraction process of oblivious watermarking.
copies of copyright information is the order of the day      Lin et al. [10] method maintained the maximum
in view of the development of Internet facilities. As a      significant coefficient to remain largest in the block
result, copyright protection and authentication has          even after embedding. Hence, the position of the
attracted many scientific and business communities in        maximum coefficient is identified in the watermark
national and international level. The applications of        extraction process. However, the embedding capacity is
watermarking are in secret communication, medical            very low as only one significant coefficient per block is
imaging, broadcast monitoring, content protection,           used. If the block size is reduced to increase the number
tamper proofing, and fingerprinting.                         of blocks, PSNR of the watermarked image reduces.
   The watermarking techniques can be divided into two          The scalar wavelets are generated by one scaling
categories: spatial domain and transform domain              function where as multiwavelets have multiple scaling
techniques. The spatial domain techniques are                functions. A novel oblivious digital image
conceptually simple and have low computational               watermarking algorithm in DMWT domain using
complexities.    However,      the    spatial   domain       QIMM is proposed which provides higher embedding
watermarking techniques are generally not robust to          capacity and robustness to various attacks.
intentional or unintentional attacks. The transform             This paper is organized as follows: Introduction to
domain techniques are generally considered to be robust      DMWT is given in Section II. QIMM is introduced in
against attacks. The transform domain techniques             Section III. The embedding and extraction algorithms
embed the watermark by modulating the magnitude of           are proposed in Section IV. The experimental results
the coefficients in the transform domain such as             and analysis are drawn in Section V. Finally, the
Discrete Fourier Transform (DFT) [1], Discrete Cosine        conclusions are given in Section VI.

360                                                              JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011

      II. DISCRETE MULTIWAVELET TRANSFORM                    This adds several degrees of freedom in multiwavelet
                                                             design and makes it possible to have several useful
   In the research field of image processing applications,
                                                             properties such as symmetry, orthogonality, short
wavelet transform has given many advantages over
                                                             support, and a higher number of vanishing moments
traditional DFT, DCT, Walsh, and HT methods in terms
                                                             simultaneously. The scaling function vector is
of reducing the blocking artifacts. In this regard, the                                                   T
design of filters for various transforms require many                                                         (1)
desirable properties such as, compact support,               satisfies a matrix dilation equation
orthogonality, symmetry, and vanishing moments.
However, the design of filters for scalar wavelets [15] is                                              .           (2)
limited.                                                     The coefficients       are         matrices instead of
   Multiwavelets can achieve better level of                 scalars. Associated with these scaling functions are
performance and higher degree of freedom than scalar         wavelets                       , satisfying the matrix
wavelets with approximately similar computational            wavelet equation
complexity. In the design of multiwavelet filters
the desirable properties [16, 17] are achieved                                                                      (3)
simultaneously.                                                                                             T
                                                             where,                                                 (4)
   In recent years, a very few papers have been reported
using DMWT in watermarking applications. Some of             is a vector and the        are        matrices.
the papers have reported moderate robustness than               When             i.e., one scaling function and one
compared to scalar wavelets. An invisible watermark or       wavelet function, the multiwavelet system reduces to
logo is fused into the host image [18, 19]. Kwon and         the scalar wavelet system.
Tewfik [33] proposed adaptive digital watermarking               In practice, multiscaling and wavelet functions are
that uses successive subband quantization and                concerned with multiplicity               . The scaling
perceptual modelling using DGHM multiwavelet. The            functions         and          are orthogonal multiscaling
watermark is Gaussian random sequence and it needs           functions and            is a          matrices. Similarly,
the original watermark sequence for detection.                      and          are two wavelet filters and        is a
Kim et al. [34] proposed adaptive watermarking scheme               matrices. An important example is constructed by
using the properties of the edge and texture region of the   GHM (Geronimo, Hardin and Massopust) system [25].
original image. Image dependent parameters are                  Some salient features to choose multiwavelets are
assumed which determines the robustness of the               summarized below :
algorithm. Unlike most of the perceptual models,                The extra degrees of freedom that is inherent in
Ghouti et al. [31, 32] proposed balanced multiwavelet        multiwavelets has reduced the restrictions on the filter
using JND profile of the image. Serdean et al. [28]          properties. Further, it is well known that a scalar
proposed much simpler perceptual model which uses            wavelet cannot simultaneously have both orthogonality
blind spread-spectrum approach for watermark                 and symmetric property [15]. Symmetric filters are
embedding. The data hiding capacity has improved over        necessary for symmetric signal extension, while
previous methods. Efforts have been made to take             orthogonality makes the transform easier to design and
advantage of machine learning techniques for                 implement. The short support and vanishing moments
watermark embedding and extraction [21-23]. These            are generally preferred to achieve a better localized
algorithms depend on the ability of learning machine         approximation of the input function. Multiwavelets are
and its learning algorithm. Many algorithms used             able to possess the best of all these properties
optimization techniques [35-38] to improve robustness        simultaneously and is not possible in scalar wavelets
of the watermark by dynamically adjusting digital            [17].
watermark embedding positions. Prayoth et al. [39]              Multiwavelets have good energy compaction
proposed embedding technique based on construction of        properties which can decorrelate the signal into a
multiwavelet tree to embed watermark. Day et al. [30]        smaller number of scaling coefficients containing most
has integrated QIMM into multiple description                of the energy. These coefficients can be used for
watermarking technique. Quantization methods have            watermarking which are better resistant to image
proven good in wavelet [20] and CT [12] domain. The          compression, rotation and other attacks.
maximum wavelet coefficients are quantized and                  Hence, multiwavelets can perform better than scalar
watermarked for copyright protection. These algorithms       wavelets with similar computational complexity in
are generally robust for geometrical attacks [10].           image processing.
   Multiwavelets are introduced as a more powerful
multiscale analysis tool. Multiwavelet is defined using        III. QUANTIZATION INDEX MODULUS MODULATION
several wavelets with several           scaling functions
[16, 25]. When a Multi-Resolution Analysis (MRA) is            Quantization is a process of approximating the
generated using multiple scaling functions and wavelet       continuous set of values in the image data with
functions, it gives rise to the notion of multiwavelets      preferably small finite discrete set of values. The
[24-28]. A multiwavelet with ‗r‘ scaling functions and       quantizer input is the original data, and the output is
‗r‘ wavelet functions is said to have multiplicity ‗r‘.      always one level among a finite number of levels.

JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011                                                                                  361

   There are three common scalar quantization methods:         where,
     1. Dither Modulation (DM)
     2. Quantization Index Modulation (QIM)                            Δ                               3Δ
                                                                       8                        v1 <
     3. Quantization Index Modulus Modulation                                                            8
          (QIMM).                                                      
                                                                        5Δ                3Δ                  7Δ
   In the previously proposed watermarking algorithms,            z' =                          ≤ v1 <                            (7)
wavelets are efficiently used with the DM [20],
                                                                       8                   8                  8
QIM [3, 4] and QIMM [12, 30, 40]. QIMM
                                                                        9Δ                 7Δ
implementation is simpler than the dither quantization.
                                                                       8                        ≤ v1 < Δ
   In any watermarking algorithm there is always a                                         8
trade-off between robustness against attacks and quality
of the image. Quality is determined by PSNR of the             case (ii): If   wi            then,           x' = xi - v1 + z''
                                                                                                              i                    (8)
watermarked image. Generally, higher degree of
robustness is achieved by higher quantization step size.       where,
But, higher quantization step size reduces the PSNR of
the watermarked image. The embedding quantization                         Δ                              Δ
step size δQIM of QIM is almost equal to two times of                    - 8                    v1 <
δQIMM of QIMM, i.e., QIMM can achieve same mean                          
                                                                          3Δ               Δ                 5Δ
square error with half of the quantization step size in            z'' =                        ≤ v1 <                            (9)
QIM [30]. Therefore, a better robustness and PSNR is                     8                 8                  8
obtained in QIMM than compared to QIM keeping                             7Δ               5Δ
quantization step size constant. The significant                         8                       ≤ v1 < Δ
coefficients obtained by wavelet transform are of the                                       8
order of 100s. Hence, the quantization steps used in           At the detector, the watermarked signal is
wavelet transform are from 5 to few 10s [30]. The
significant coefficients of Multiwavelet transform are of                                X' = {x1',x2',...,xn'}
the order 1000s. Therefore, the quantization steps can be
higher values, with good PSNR and higher robustness            and is detected as follows:
against attacks. There are two methods in QIMM.
                                                                                                Δ             Δ            3Δ
   1) Method 1: The host signal (coefficients obtained
                                                                       0           0 ≤ v2 <         or            ≤ v2 <
after multiwavelet transform) is divided by the                                                 4             2            4
quantization step size (  ) and the nearest integer value        w' =                                                           (10)
                                                                       1           Δ            Δ             3Δ
is obtained. This value is then executed with modulo 2                                  ≤ v2 <          or          ≤ v2 < Δ
to obtain the remainder as 0 or 1. Then, if remainder is               
                                                                                   4            2              4
equal to the watermark bit value, then the reconstructed
value is the quantized host signal, else it is biased either   where,                   v2 = mod(xi',  )                         (11)
+1 or -1 [30].
   2) Method 2: Modulo quantization step (  ) is                       IV. PROPOSED WATERMARKING SCHEME
executed to the host signal and the remainder is
                                                                  A. Prefiltering of Host Image
subtracted from the host signal. The error signal is
added to the result which is near to the host signal              Multiwavelets differ from scalar wavelet systems as it
depending on the watermark bit 1 or 0 [12]. A                  needs several input streams rather than one. This
maximum error signal of {+  /4 or -  /4} is added so         necessitates a pre-filtering operation on the input
the noise immunity is  /4.                                    stream. This pre-filtering operation is also called multi-
   The method 2 is used in the paper as the coefficients       wavelet initialization and can be performed in a
obtained using multiwavelet transform are very large in        critically sampled or an over-sampled fashion.
size. QIMM used in this proposed algorithm is                  Prefiltering based on Strela‘s algorithm [16] is used. In
described as follows:                                          this algorithm, over-sampled procedure is used where
   In the embedding process, if the host signal is             two identical rows are taken as input to the multifilter
                                                               bank. This procedure is also called Repeated Rows
                    X = {x1 ,x2 ,....,xn   }                   (RR). It introduces over-sampling of data by a factor
                                                               of 2. RR prefiltering have proven useful in many of the
and the watermark logo bits,     wi            Let  be the    image processing applications like feature extraction,
                                                               denoising, etc. As the data redundancy can increase the
Quantization step. To modify the host signal X to X' ,
                                                               embedding capacity, the RR prefiltering is employed in
the following cases are considered.
                                                               this paper.
Suppose             v1 = mod(xi , )                    (5)      B. The Watermark Embedding Procedure
                                                                  Due to the versatility and accuracy of the GHM, it is
case (i): If   wi       then,     x' = xi - v1 + z'
                                   i                    (6)    used in this paper. Experimentation has been performed

362                                                                   JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011

in different levels. The multichannel nature of                   The flow chart for watermark embedding algorithm is
multiwavelets yields a different sub-band structure               shown in Fig. 3(a).
compared with scalar wavelets. Multiwavelet sub-bands
of 2-dimensional image using first level decomposition                  Host Image N × N
produce 16 sub-bands as shown in Fig. 1.

                  L1L1       L2L1   H1L1     H2L1                    Perform RR Prefiltering                      Secret Key K

                  L1L2       L2L2   H1L2     H2L2                                    2N × 2N

                                                                    Compute GHM (kth level)                 Watermark Logo p × m
                  L1H1       L2H1   H1H1     H2H1                    Multiwavelet Transform

                  L1H2       L2H2   H1H2     H2H2
           Figure 1. First Level multiwavelet decomposition         Select Low-Low Pass sub               Scrambled Watermark Logo
                                                                              band                                  p×m

                         a                   b

                                                                                 Modify the selected coefficients xi using
                                                                                   QIMM as per watermark logo bits
                                                                             1                                               0
                                                                                            Watermark Bit

       Figure 2. Three level DGHM multiwavelet decomposition          Generate z‘ using (9) and              Generate z‘ using (7) and
         a)Prefiltering b) 1st Level decomposition c) 2nd Level         Modify xi as in (8)                    Modify xi as in (6)
               decomposition d) 3rd Level decomposition

      The four ‗low-low-pass‘ sub-bands are further                                      Compute Inverse GHM
decomposed to            sub-bands in the next level of                                     Multiwavelet
decomposition. This process continues number of
times for kth level of decomposition. The number of sub
matrices will be equal to           where k is the number                                      RR Postfiltering
of levels of decomposition. This can be observed in
Fig. 2 for 3rd level GHM multiwavelet decomposition.                                       Watermarked Image
      The steps of watermark embedding algorithm are                                             N×N
as follows:
                                                                     Figure 3(a). Flow chart for watermark embedding algorithm
     1. The host image of size N × N is prefiltered
          using RR prefilter to produce two input                   C. The Watermark Extraction Procedure
          streams. The size of the image after RR
          prefiltering is 2N × 2N                                    The steps of watermark extraction algorithm are as
     2. GHM multiwavelet transform is applied to the
          host image after prefiltering.       A      level       
          decomposition is performed on the image. The                 1. The watermarked image of size N × N , is
          ‗low-low-pass‘ sub-bands of            level are                 prefiltered using RR prefilter. The size of the
          selected for watermark embedding where it                        image after RR prefiltering is 2N × 2N ,.
          concentrates its most of the energy.                         2. The GHM multiwavelet transform is applied to
     3. To increase the watermark security, the                            the resultant image after prefiltering. A kth
          watermark logo is scrambled with a secret key.                   level decomposition is performed on the image.
          It is impossible to obtain the original                      3. The bits of wi ' will be extracted from the
          watermark logo without the secret key.
                                                                           multiwavelet coefficients of selected blocks
     4. The bits of wi will be embedded into the                           using QIMM method as in (10) using (11)
           multiwavelet coefficients of selected blocks                    described in section III.
           using QIMM method proposed in section III.                 4.   The bits obtained are descrambled with the
      5.   The inverse GHM multiwavelet is applied to                      secret key.
           the modified sub-bands.                                    5.   After extracting the watermark, BER is used to
      6.   The RR post-filtering is performed and the                      quantify the correlation between the original
           final watermarked image is obtained.                            watermark and the extracted one.
                                                                                                                                        

JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011                                                                                  363

    The flow chart for watermark extraction algorithm
    is shown in Fig. 3(b).
                           Watermarked image N × N

                            Perform RR Prefiltering
                                              2N×2N                             (i)                                (ii)
                                                                     Figure 4(a). (i) Original Baboon image, (ii) Watermarked Baboon
                            Compute GHM( kth level)
                                                                                         image without attack
                             Multiwavelet Transform

                            Select Low-Low Pass sub

                            Extract w ' using QIMM
                                      i                                         (i)                 (ii)                  (iii)
                               as in (10) using (11)

    Secret Key K            Descrambled Watermark
                                  Logo p × m

                                                                                (iv)                 (v)                  (vi)
                              Extracted watermark
                                  Logo p × m
                                                                     Figure 4 (b). (i)-(iii) Original watermark logo of sizes        ,
                                                                           and            respectively & (iv)-(vi) Scrambled watermark
                             Compute BER between                      logo of sizes            ,          and          respectively
                              original and extracted                 Let the host signal be X and the watermarked signal
                                                                  be X' and are of size N × N , then PSNR in dB is given
     Figure 3(b). Flow chart for watermark extraction algorithm
                                                                                                                
        V. EXPERIMENTAL RESULTS AND ANALYSIS                                                N 2 max(max(X))2 
                                                                          PSNR = 10log10  N N                    (11)
       To evaluate the performance of the proposed                                                              
                                                                                             (X - X') 
watermarking scheme, experiments have been
conducted in which a GHM multiwavelet is used to                                            x=1 y=1             
decompose the original image using RR pre-processing.             where, X and X' indicate the pixel intensities of host
The number of transform coefficients are more in RR               and watermarked signals respectively.
pre-processing than compared to ‗critically sampled‘                 Various multiwavelets like GHM, c1, Sa4, Bih52s,
pre-processing, i.e., for an image of                    , the    Bih54n [15], and a scalar Haar wavelet are used. Haar
transform coefficients of RR are                     whereas      wavelet gave very inferior PSNR value as the  is very
in ‗critically sampled‘ pre-processing it is                      high. Haar wavelet coefficients are of the order of
only. Hence, embedding capacity is more in RR than                hundreds, where as multiwavelet coefficients are of the
compared ‗critically sampled‘ pre-processing.                     order of thousands in ‗low-low pass‘ band. Almost all
       The original image is a                    pixels gray     multiwavelets gave similar PSNR values but GHM
scale image. Images like Baboon, Lena, Boats and                  outperformed as shown in Figs. 5 & 6. The
Peppers are taken from Stirmark image database of                 experimental results also show that, RR pre-processing
Fabien Petitcolas. The original Baboon image and                  method is giving better PSNR than the ‗critically
watermarked Baboon image are shown in Fig. 4(a) (i) -             sampled‘ method for various multiwavelets. But,
(ii). The binary watermark logo of                ,               computational complexity of RR is higher than the
and             are chosen in the paper are as shown in           ‗critically sampled‘ pre- and post-processing.
Fig. 4(b) (i)-(iii) and Fig. 4(b) (iv)-(vi) shows the                For a given size of           watermark logo, as the
corresponding scrambled watermark logo.                           number of multiwavelet decomposition levels increased,
    BER is a very useful measure of the performance. In           computational complexity increases and also, the PSNR
this case, the bit error rate is calculated as the number of      of the image decreases. It means that, the embedding
incorrectly decoded bits divided by the total number of           capacity of the watermarked image has decreased for a
embedded bits in the watermarked image.                           given robustness towards attacks.

364                                                                                      JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011

                     Values of psnr in dB for various Multiwavelets                                             TABLE I. (a)
                    60                                                                     Comparison of the proposed method with Lin et al. method
                                                                                           Comparison             Lin et al. [10]     Proposed method
                                                                                                                              Wavelet              Multiwavelet
                                                                                 Transform Domain
                    40                                                                                                        (Haar )               ( GHM)
      PSNR in dB

                    30                                                          Decomposition Level                           Level : 3              Level : 3

                                                                                       PSNR in dB
                    10                                                               (logo:       )
                                                                                                                              40.3172               48.8454
                                                              bih52    bih54
                            haar    ghm       cl      sa4
                                                                s        n
                                                                                     Max. Embedding                                       .
                   ∆=160 6.7711 50.678 47.497 47.489 49.604 49.815                      Capacity
                   ∆=180 6.8018 49.666 46.207 45.847 48.585 48.794

Figure 5. PSNR of various multiwavelets at ∆=160 and         for
cover image baboon using ‗repeated rows‘ pre-processing and post-
                                                                                                        TABLE I. (b)
             processing.with a logo size of        .
                                                                                 Comparison of the proposed method with Lin et al. method with
                    Values of PSNR in dB for various Multiwavelets
                                                                                                    Attacks / BER                   Lin et al.       Proposed
                    60                                                                                                                [10]            method
                                                                               Low Pass Filtering (      )                           0.0742           0.0625
                    50                                                         Median Filtering (      )                             0.1035           0.0859
                    40                                                         Resizing ( 512-256-512)                               0.0234           0.0195
      PSNR in dB

                                                                               Rotation ( 1 Degree)                                  0.0352           0.0332
                    30                                                         JPEG ( QF=50)                                         0.0117              0
                                                                               JPEG ( QF=60)                                         0.0078              0
                    20                                                         Cropping ( ¼)                                         0.2383              0
                                                                               Bit Plane Removal :                                   0.0039              0
                    10                                                         (2nd LSB)
                                    ghma                      bih5a    bih54
                                             clap    sa4ap
                                                                p        n
                   ∆=160 6.7711 50.166 29.184 47.184 31.237 31.093                                  55
                                                                                      PSNR(in dB)

                   ∆=180 6.8018 48.883 29.399 45.878 31.106 30.848

Figure 6. PSNR of various multiwavelets at ∆=160 and ∆=180 for a                                    35
 cover image Baboon using ‗critically sampled‘ pre-processing and

            post-processing with a logo size of        .

                                   Multiwavelet : GHM                                                                        Quantization Step

                                                                                                              Figure 8. PSNR in dB vs Quantization
                    41.8                                                                 0.55
      PSNR in dB

                    41.6                                                                 0.45
                    41.4                                                                  0.4

                                                                                         0.25                                                    JPEG2000
                     41                                                                   0.2                                                    JPEG
                    40.8                                                                 0.15
                    40.6                                                                 0.05
                              Baboon       Peppers     Lena           Boats
                     PSNR    41.2328       41.3366    41.3731     42.1028                                10     20    30     40 50 60 75             90   100
                                                                                                                           Quality Factor (QF)
Figure 7. PSNR of various images for ‗GHM‘ multiwavelet at ∆=170
using ‗repeated rows‘ pre-processing and post-processing with a logo                 Figure 9. BER vs JPEG and JPEG2000 compression of recovered
                          size of                                                                        logo size of

JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011                                                                                               365

                                                                TABLE II.
               BER after attacked by JPEG comparison with various quality factors (QF) for Baboon image for different sized logos
 BER/ QF             10              20                      30                       40                        70                        100
                                   0.1074                  0.1270                      0                         0                         0

 (a) Logo :    0.4883
              NA (< 0.2)

                                   0.1863                  0.0398                  0.0046                       0                         0

 (b) Logo :    0.4844
              NA (< 0.2)

                                                                   TABLE III.
                                     BER after Rotational attacks for Baboon image for different sized logos
BER/ Rotation (Degree)         0.25ᵒ                   0.3ᵒ                    1ᵒ                     10ᵒ                  50ᵒ
                               0.0303                 0.0273                 0.0195                  0.0732               0.1592

(a) Logo :

                               0.0273                 0.0271                 0.0344                  0.0415               0.0732

(b) Logo :

                                                                TABLE IV.
                                                            BER for various attacks
BER/ Other attacks             Salt & Pepper Noise          Row Column Blanking              Row Column Copying              Cropping ( ¼)
                                     0.0117                         0.0850                           0.0381

 (a) Logo :

                                     0.0132                         0.0845                           0.0398                           0

 (b) Logo :

BER/ Other attacks           Bit Plane Removal              Gaussian Noise                 Low Pass Filtering        Resizing ( 512-256-512)
                                  ( 2nd LSB)

                                      0                          0.0361                         0.0654                       0.0225

(a) Logo :

                                      0                          0.0752                         0.1287                       0.0764

 (b) Logo :

366                                                                  JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011

   The quantization step level chosen is              and the    cropping, and bit plane removal. The proposed
cover image is baboon. The PSNR is calculated taking             algorithm can be applied to image copyright protection
different multiwavelet decomposition levels, i.e., from          and authentication.
level 1 to level 8. Taking into consideration of
embedding capacity and robustness against attacks,                                       REFERENCES
multiwavelet decomposition level 3 is found to be                [1] O‘ Ruanaidh. J.J.K., Dowling W.J., and Boland F.M.,
appropriate. The PSNR values of various images using                  ―Phase watermarking of digital images,‖ Proc. IEEE
GHM multiwavelet for third level of decomposition is                  International conference Image Processing, vol. 3, pp.
shown in Fig. 7.                                                      239-242, 1996.
   In this proposed method, with a minimum value of              [2] Chen B., ―Design and analysis of digital watermarking,
quantization step equal to , the retrieved watermark                  information embedding, and data hiding systems,‖ Ph.D.
logo with BER equal to zero is obtained, i.e., the                    thesis, Massachusetts Institute of Technology,
                                                                      Cambridge, MA, 2000.
watermark is retrieved exactly for a Δ value of 22.
                                                                 [3] Chen B. and Wornell G.W., ―Quantization Index
Fig. 8 shows the PSNR (in dB) with Δ for a given level                Modulation: A class of provable good methods for digital
of decomposition from              to       . As the Δ value          watermarking and information embedding,‖ IEEE Trans.
increases, the PSNR value decreases and the resistance                On Information Theory, vol. 47, pp. 1423-1443, 2001.
to various attacks increases considerably.                       [4] Chen B. and Wornell G.W., ―Quantization Index
   The results of the proposed work are compared with                 Modulation methods for digital watermarking and
Lin et al. [10] using Baboon image as shown in                        information embedding of multimedia,‖ Journal of VLSI
Table I(a). Lin et al. [10] proposed a blind                          Signal Processing Systems for Signal, Image, and Video
                                                                      Technology, Special issue on Multimedia Signal
watermarking algorithm and its method is based on
                                                                      Processing, vol. 27, pp. 7-33, 2001.
maximum wavelet coefficient quantization.                        [5] Cox I.J., Kilian J., Leighton F.T., and Shamoon T.,
   The proposed work gives better PSNR value for a                    ―Secure spread spectrum watermarking for multimedia,‖
given logo size of                 i.e., 48.8454 where as the         IEEE Transaction Image Processing, vol. 6, pp. 1673-
PSNR for the logo size                      in Lin et al. [10]        1687, 1997.
methods is 40.3172 dB. In the proposed method, the               [6] Chu W.C, ―DCT-based image watermarking using
PSNR for the logo size                  is 41.2328 dB. Hence,         subsampling,‖ IEEE Trans. Multimedia, vol. 5,
the embedding capacity has greatly increased with                     pp. 34-38, 2003.
better PSNR value.                                               [7] Hsu C.T and Wu J.L, ―Hidden signatures in images,‖
                                                                      IEEE International conference Image Processing, pp.
   Different attacks were applied using tools of Stirmark             743-746, 1996.
benchmark, Anything 3D Corp. JPEG2000 compression                [8] Nikolaidis A and Pitas I, ―Asymptotically optimal
and Jasper toolkit with different quality factors. Fig 9              detection for additive watermarking in the DCT and
shows the BER vs JPEG and JPEG2000 compression of                     DWT domains,‖ IEEE Transaction Image Processing,
recovered logo size of                    Experimental results        vol. 12, pp. 563-571, 2003.
show that the BER is better for various attacks than             [9] Taweel G S El, Onsi H M, Samy M, and Darwish M G,
compared to Lin et al. [10] method as shown in                        ―Secure and Non-blind watermarking scheme for color
Table I (b). The Table II, Table III and Table IV shows               images based on DWT,‖ GVIP special issue on
                                                                      watermarking, 2007.
the BER for various attacks with two different
                                                                 [10] Wei-Hung Lin, Yuh-Rau Wang, Shi-Jinn Horng, Tzong-
watermark logo sizes                     and             along        Wann Kao, Yi Pan: ―A blind watermarking method using
with the retrieved watermark logos from watermarked                   maximum wavelet coefficient quantization,‖ Experts
image. The retrieved watermark logo against                           systems with Applications, Elsevier, 2009.
compression attack with a quality factor of 40 and               [11] Wang Y, Doherty J F and Van Dyck R E, ―A wavelet
above is exactly same as that of the original watermark               based watermarking algorithm for ownership verification
logo, i.e., BER is zero. This proposed algorithm has also             of digital images,‖ IEEE Transaction on Image
contributed better results for rotational attacks up to 50ᵒ           Processing, vol. 11, pp. 77-88, 2002.
compared to earlier works. The watermark logo of                 [12] Dongfang Chen, and Bing Zhou, ―An oblivious
                                                                      watermarking scheme based on Contourlet transform and
            is retrieved after a rotational attack of 50ᵒ with
                                                                      quantization index modulus modulation,‖ International
a BER of 0.0732.                                                      Journal of Intelligent Engineering & Systems, pp. 21-28,
                       VI. CONCLUSION                            [13] Do M N and Vetterli M, ―The Contourlet transform: an
                                                                      efficient      directional      multiresolution   image
    In this paper, an oblivious image watermarking in                 representation,‖ IEEE Transaction Image Processing,
DMWT using QIMM is proposed. The proposed                             vol. 14, pp. 2091-2106, 2005.
method is superior to Lin et al. [10] in terms of PSNR,          [14] Zhao Xu, Ke Wang and Xiaohua Qiao, ― A novel
BER and embedding capacity. The proposed method is                    watermarking scheme in Contourlet domain based on
robust against JPEG and JPEG2000 compression,                         independent component analysis,‖ Proc. of International
rotational attacks, and common attacks like low pass                  conference on intelligent information hiding and
filtering, Gaussian filtering, resizing, salt & pepper                multimedia signal processing, Califonia, USA, pp. 59-62,
noise, row column blanking, row column copying,                       2006.

JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011                                                                            367

[15] Strang G. and Nguyen T., Wavelets and Filter Banks,                security and watermarking of multimedia contents IV,
     Wellesley, MA: Wellesley-Cambridge Press, 1995.                    vol. 4675, pp 334-348, 29th April, 2002.
[16] Strela V., Heller P.N., Strang G., Topiwala P., and Heil    [34]   Ji-Hong Kim, Hee-Jung Kim, and Ki-Ryong Kwon,
     C : ‖The application of multiwavelet filterbanks to image          ―Multiwavelet image watermarking using perceptually
     processing,‖ IEEE Trans. Image Processing, vol. 8, no. 4,          tuned model,‖ IJCSNS International Journal of Computer
     pp. 548–563, 1999.                                                 Science and Network Security, vol. 6, No. 12, Dec. 2006.
[17] Keinert F.: Wavelets and multiwavelets, Chapman and         [35]   Kumsawat, P., Attakitmongcol, K., and Srikaew, A.: ―A
     Hall/CRC: 2004, ISBN: 1-58488-304-9.                               new approach for optimization in image watermarking by
[18] Zhao J., Liu Z., and Laganiere R.: ―Digital watermarking           using genetic algorithms,‖ IEEE Trans. Signal Process.,
     by using a feature-based multiwavelet fusion approach,‖            53, (12), pp. 4707–4719, 2005.
     Canadian Conf. on Electrical and Computer Engineering,      [36]   Prayoth Kumsavath, Kitti Attakitmongcol and Arthit
     vol. 1, pp. 563–566 , 2–5 May 2004.                                Srikaew, ―An optimal robust digital image watermarking
[19] Kurugollu F., Bouridane A., Roula M., and Boussakta                based on Genetic algorithms in multiwavelet Domain,‖
     S.: ―Comparison of different wavelet transforms for                WSEAS Transactions on Signal Processing, issue 1,
     fusion based watermarking applications,‖ Proc.10th IEEE            volume 5, Jan. 2009.
     Int. Conf. on Electronics, Circuits and Systems, 14–17      [37]   Jian Luo, Yinghui Pan, and Liang Tan, ―Digital
     December 2003, vol. 3, pp. 1188–1191.                              watermarking multi-objective optimization based on
[20] Xinshan Zhu, and Zhi Tang, ―Improved Quantization                  multi-wavelet,‖ IEEE International conference on control
     Index Modulation watermarking robust against amplitude             and automation, Dec., 2009.
     scaling distortions, IEEE International conference on       [38]   Pan Yinghui, ‖Digital watermarking particle swarm
     Multimedia and Expo, 2008.                                         optimization based on multi-wavelet‖, Journal of
[21] Zhang, J., Wang, N.-C., Xiong, F., ―Hiding a logo                  Convergence Information Technology, Vol. 5 , No. 3,
     watermark into the multiwavelet domain using neural                2010.
     networks,‖ In: Proceedings of 14th IEEE International       [39]   Prayoth K, Attakitmongcol K and Srikaew A, ―A robust
     Conference on Tools with Artificial Intelligence, ICTAI -          image watermarking scheme using multiwavelet tree,‖
     2002.                                                              Proceedings of the World Congress on Engineering,
[22] Li C.-H., Lu Z.-D., and Zhou K., ―An image                         vol. I, July, 2007.
     watermarking techniques based on support vector             [40]   Pierre Moulin, and Anil Kumar G, ―Block QIMM
     regression,‖ Proceedings of ISCIT, pp. 177–180, 2005.              Watermarking games,‖ IEEE Transaction on Information
[23] Hong Peng, Jun Wang C, and Weixing Wang, "Image                    Forensics and Security, Vol 1, Issue 3, Sep., 2006.
     watermarking method in multiwavelet domain based on
     support vector machine,‖ The Journal of systems and
     Software, pp. 1470-1477, 2010.                                                     N. Leelavathy, is currently working as
[24] Shen L.X, Tan H.H, and Tham J.Y, ―Symmetric-                                       Professor and HOD, Computer Science
     antisymmetric orthogonal multiwavelets and related                                 and Engineering Department, Krishna‘s
     scalar wavelets,‖ Appl. Comput. Harmonic Anal. 8 (3),                              Pragati Institute of Technology,
     pp. 258–279, 2000.                                                                 Rajahmundry, India. She is working
[25] Geronimo J., Hardin D., and Massopust P., ―Fractal                                 towards her Ph.D. at Jawaharlal Nehru
     Functions and Wavelet Expansions Based on Several                                  Technological University, Kakinada,
     Functions,‖ Journal of Approximation Theory, vol. 78,                              India. She received her M.Tech. in
     pp. 373-401, 1994.                                                                 Computer Science and Engineering from
[26] Tham J., Shen L., Lee S., and Tan H., ―A General            Jawaharlal Nehru Technological University, Hyderabad, India
     Approach for Analysis and Application of Discrete           in 2003. She did her B.E. in Electronics & Communication
     Multiwavelet Transforms,‖ IEEE Transactions Signal          Engineering from Vasavi College of Engineering, Hyderabad,
     Processing, vol. 48, pp. 457- 464, 2000.                    India in 1992. She has eleven years experience of teaching
[27] Md. Khalil I, ― Digital image watermarking : Scalar         undergraduate students and post graduate students. Her
     Wavelet versus Multiwavelets,‖ IEEE Trans., 2006.           research interests are in the areas of Digital Image Processing,
[28] Serdean C.V, Ibrahim M.K, Moemeni A, and Al-Akaidi          Image Watermarking, and Cryptography and Network
     M.M, ―Wavelet and Multiwavelet watermarking,‖ The           Security.
     Institution of Engineering and Technology, 2007.
[29] Soman K. and Ramachandran K., Insight into Wavelets                               E. V. Prasad, is working as Director,
     from Theory to Practice, Prentice Hall of India: New                              Institute of Science and Technology,
     Delhi, 2004.                                                                      Jawaharlal       Nehru      Technological
[30] Miin- Luen Day, Suh-Yin Lee and I-Chang Journal,                                  University, Kakinada, India. He received
     ―Multiple Description Watermarking Based on                                       his Ph.D. in Computer Science &
     Quantization Index Modulus Modulation,‖ Journal of                                Engineering from University of Roorkee
     Information Science and Engineering, 1785-1800, 2007.                             in 1990. He did his M.E. in Control
[31] Ghouti L, Bouridane A, and Boussakta S, ―High capacity                            Systems from Madras University in 1978.
     watermarking using balanced multiwavelet transforms,‖                             He received his B.E. in Electronics &
     IEEE, 2005.                                                 Communication Engineering from S V University in 1975. He
[32] Ghouti L, Bouridane A, Md. Ibrahim K, and Boussakta S,      has thirty three years of experience in teaching undergraduate
     ―Digital     image     watermarking     using    balanced   students and post graduate students. He has taught over 15
     multiwavelets,‖ IEEE transactions on signal processing,     courses in CSE and supervised 4 Ph.D students. He held
     Vol. 54, No. 4, April, 2006.                                different positions in his carrier like Head of the Department,
[33] Kwon Ki R and Tewfik A H, ―Adaptive watermarking            Vice Principal, Principal and BOS Chairman. He is the Co
     using successive subband quantization and perceptual        author of four books and published more than twenty papers in
     model based on multiwavelet transform,‖ Proc. SPIE-         National and International journals and 53 in various

368                                                                JOURNAL OF MULTIMEDIA, VOL. 6, NO. 4, AUGUST 2011

Conferences. His research interests include             Parallel
Computing, Data Mining, and Information Security.

                            S. Srinivas Kumar is currently
                            working as Professor and Director,
                            Sponsored Research, Jawaharlal
                            Nehru Technological University,
                            Kakinada, India. He received his
                            Ph.D. from E&ECE Department
                            IIT, Kharagpur. He received
                            his M.Tech. from Jawaharlal
                            Nehru Technological University,
                            Hyderabad, India. He has twenty
                            one years experience of teaching
undergraduate and post-graduate students and guided number
of post-graduate thesis. He has published 30 research papers in
National and International journals. Presently he is guiding
ten Ph.D. students in the area of Image processing. His
research interests are in the areas of digital image processing,
computer vision, and application of artificial neural networks
and fuzzy logic to engineering problems.

                           Chandra Mohan B is currently
                           working as Professor & Dean,
                           academics, Bapatla Engineering
                           College, Bapatla, India. He
                           received his Ph.D. from JNTU
                           College of Engineering, Kakinada,
                           India. He received his M.Tech.
                           from Cochin University of Science
                           & Technology, Cochin, India. He
                           has seventeen years experience of
                           teaching undergraduate students
and post graduate students. He has published seven research
papers in National and International journals. His research
interests are in the areas of image watermarking, and image


Shared By: