Watermarking 7

							 A DWT-BASED ROBUST SEMI-BLIND
IMAGE WATERMARKING ALGORITHM
        USING TWO BANDS

         Ersin Elbasiα and Ahmet M. Eskiciogluβ
         αThe  Graduate Center, The City University of New York
                   365 Fifth Avenue, New York, NY 10016
  βDepartment of Computer and Information Science, Brooklyn College

The City University of New York, 2900 Bedford Avenue, Brooklyn, NY 11210


  IS&T/SPIE’s 18th Annual Symposium on Electronic Imaging, Security,
Steganography, and Watermarking of Multimedia Contents VIII Conference
                  San Jose, CA, January 15–19, 2006.



                                                                           1
CLASSIFICATION OF IMAGE WATERMARKING SYSTEMS



    Criterion          Class                           Brief description

    Domain type        Pixel                           Pixels values are modified to embed
                                                       the watermark.

                       Transform                       Transform coefficients are modified
                                                       to embed the watermark. Recent
                                                       popular transforms are Discrete
                                                       Cosine Transform (DCT), Discrete
                                                       Wavelet Transform (DWT), and
                                                       Discrete Fourier Transform (DFT).
    Watermark type     Pseudo random number            Allows the detector to statistically
                       (PRN) sequence (having a        check the presence or absence of a
                       normal distribution with zero   watermark. A PRN sequence is
                       mean and unity variance)        generated by feeding the generator
                                                       with a secret seed.
                       Visual watermark                The watermark is actually
                                                       reconstructed, and its visual quality
                                                       is evaluated.
    Information type   Non-blind [7,12,18]             Both the original image and the
                                                       secret key(s)
                       Semi-blind                      The watermark and the secret key(s)

                       Blind                           Only the secret key(s)


                                                                                               2
A DWT-BASED SEMI-BLIND IMAGE WATERMARKING SCHEME


   “A New Wavelet-Based Scheme for Watermarking Images” by Dugad,
    Ratakonda and Ahuja (ICIP 1998, October 4-7, 1998, Chicago, IL).
        The LL band is left out.
        In the other bands (HL, LH, and HH), the watermark is embedded into the
         coefficients that are higher than a given threshold T1.
        During watermark detection, all the high pass coefficients above another threshold
         T2 (T2  T1) are used in correlation with the original watermark.
        Watermark embedding
          Compute the NxN DWT of an NxN gray scale image I.
          Exclude the low pass DWT coefficients.
          Embed the watermark into the DWT coefficients > T1:
                        ,               , where i runs over all DWT coefficients > T1.
          Replace         with       in the DWT domain.
          Compute the inverse DWT to obtain the watermarked image I’.
        Watermark detection
          Compute the DWT of the watermarked and possibly attacked image I*.
          Exclude the low pass DWT coefficients.
          Select all the DWT coefficients higher than T2.
          Compute the sum z =          , where i runs over all DWT coefficients > T2, where yi
          represents either the real watermark or a fake watermark,      represents the watermarked and
          possibly attacked DWT coefficients.
          Choose a predefined threshold Tz =           .
          If z exceeds Tz, the conclusion is the watermark is present.                                3
EXPERIMENTS



   In this paper, we extend the idea to embed the same watermark in
    two bands (LL and HH).
   In our experiments, we obtained the first level decomposition using
    the Haar filter.
   The values of the scaling factor and the threshold for each band are
    given in Table 1.


     Table 1. Scaling factor α and threshold T

                                 LL               HH
      Parameters/Bands
             α                  0.01              0.4
             T1                  90               45
             T2                  100              55




                                                                      4
EMBEDDING TWO WATERMARKS INTO AN IMAGE




     Original Lena   Watermarked Lena    The difference
                      (PSNR=41.17)




                                                          5
ATTACKS




       JPEG compression              Resizing              Gaussian noise
            (Q=25)              (256 → 128 → 256)     (mean = 0, variance = 0.001)




          Low pass filtering                            Histogram equalization
          (window size=3x3)                                  (automatic)
                                   Rotation   (200)




                                                                                     6
       Contrast adjustment      Gamma correction       Cropping on both sides
      ([l=0 h=0.8],[b=0 t=1])       (1.5)
DETECTOR RESPONSE FOR UNATTACKED WATERMARKED LENA




          LL band (T=0.831)   HH band (T=5.118)




                                                    7
DETECTOR RESPONSE FOR JPEG COMPRESSION: Q=25




          LL band (T=1.134)   HH band (T=5.904)




                                                  8
DETECTOR RESPONSE FOR RESIZING




          LL band (T=1.281)      HH band (T=5.563)




                                                     9
DETECTOR RESPONSE FOR GAUSSIAN NOISE




          LL band (T=1.056)   HH band (T=6.741)




                                                  10
DETECTOR RESPONSE FOR LOW PASS FILTERING




          LL band (T=1.065)    HH band (T=2.869)




                                                   11
DETECTOR RESPONSE FOR ROTATION (200)




           LL band (T=1.039)    HH band (T=5.027)




                                                    12
DETECTOR RESPONSE FOR HISTOGRAM EQUALIZATION




          LL band (T=1.309)   HH band (T=9.357)




                                                  13
DETECTOR RESPONSE FOR CONTRACT ADJUSTMENT




          LL band (T=1.252)   HH band (T=6.731)




                                                  14
DETECTOR RESPONSE FOR GAMMA CORRECTION




          LL band (T=0.892)   HH band (T=5.120)




                                                  15
DETECTOR RESPONSE FOR CROPPING




          LL band (T=1.025)      HH band (T=6.150)




                                                     16
CONCLUSIONS



   A DWT-based semi-blind image watermarking paper
        The LL band is left out.
        In the other bands (HL, LH, and HH), the watermark is embedded into the
         coefficients that are higher than a given threshold T1.
        During watermark detection, all the high pass coefficients higher than another
         threshold T2 (T2  T1) are chosen for correlation with the original watermark.
   In this paper, we have extended the idea by embedding the same
    watermark in two bands (LL and HH) using different scaling factors
    and thresholds for each band.
   For one group of attacks (JPEG compression, resizing, adding
    Gaussian noise, low pass filtering, and rotation), the correlation
    with the real watermark is higher than the threshold in the LL band.
   For another group of attacks (histogram equalization, contrast
    adjustment, gamma correction, and cropping), the correlation with
    the real watermark is higher than the threshold in the HH band.



                                                                                    17