You are on page 1of 15

1

Optimum Retrieval of Watermark from Wavelet Signicant Pixels


Jayalakshmi M*, S. N. Merchant, U. B. Desai SPANN Lab, Electrical Engineering Department Indian Institute of Technology, Bombay, Powai, Mumbai-76, India email: (jlakshmi*,merchant,ubdesai)@ee.iitb.ac.in

Abstract Watermark retrieval in wavelet domain has been proved to be more robust from signicant pixels than high absolute pixels when a single copy of the watermark is embedded in the original data. The watermark energy can be maximized at each of the selected pixels by quantizing them to the maximum allowable levels suggested by Human Visual System (HVS) model. But as the attacks become very severe, a single copy of the watermark is not sufcient for correct retrieval. Hence, we propose a method of optimum retrieval of watermark from multiple embedded copies under very severe attacks. We propose to use Chair-Varshney (CV) decision fusion rule to decide each bit in the watermark instead of majority rule for optimum watermark retrieval. Simulations are performed to show the superiority of the method by embedding different number of watermark copies under various attacks.

I. I NTRODUCTION Recent years have witnessed an outgrowth in the volume of digital data which can be easily manipulated and reproduced. Digital watermarking has been proposed as a means for copyright protection, owner verication, content authentication and broadcast monitoring. A number of watermarking algorithms in spatial domain [1], [2] as well as transform domain have been proposed. A major disadvantage of spatial domain techniques is the low robustness of the watermark. The robustness of the watermark could be improved if the properties of the cover image could be exploited. The most commonly
*Corresponding author

used transforms for digital watermarking are Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT) [3], [4], [5], [6], [7], [8]. Given its suitability to model the HVS behavior, the DWT has gained interest among watermarking researchers. In [9] a blind watermarking algorithm which embeds the watermark in the DWT domain by exploiting the characteristics of the human visual system is presented. Here, watermark strength modulation is accomplished through a mask giving a pixel by pixel measure of the sensitivity of the human eye to local image perturbations. Mask construction relies on a work by Lewis and Knowles [10]. Some modications to the method by Lewis and Knowles are proposed by Barni et al. [9] to make it suitable to the computation of the maximum visibly tolerable watermark energy that can be used for each DWT coefcient. When a watermark is embedded into an original data, the set of pixels which are selected for watermark embedding is to be chosen such that both transparency and robustness conditions are met. At the same time, the number of watermark copies to be embedded is also of concern since too many copies may distort the perceptual quality of the image. If the selection of pixels is such that the robustness is assured, then a single copy of the watermark is enough to meet the requirements. This is very well achieved with wavelet domain signicant pixels quantized to their maximum using HVS model. But, as the attacks become severe and noticeable distortions are produced on the watermarked image, the retrieval becomes extremely difcult. So, we propose to embed more number of watermarks under such stringent conditions. But, as there is lot of deterioration, normal detection and retrieval techniques fail to reproduce the original watermark. Hence we propose to use CV rule for optimum detection of each of the watermark bit. To use CV rule for detection, we need to know the probabilities of detection ( of false alarm (

) and probabilities

) for each bit in each copy of the watermark. Initially,

and

are xed for all bits in every copy of the watermark. These values are later modied to different values depending on the retrieved bit and the actually expected bit at that position. This is possible since it is a non-blind scheme. In this paper, we have selected only those attacks which very severely distort the image and affect the retrieval of the watermark. We test the proposed method for robust detection under high compression, noise addition, different types of ltering and histogram 2

equalization. Also, we can notice improvement in detection as the number of watermark copies are increased. We have simulated results with majority rule and CV rule with watermark copies varying from 1 to 8. It is observed that just 3 copies are enough for correct retrieval when CV rule is used for deciding the watermark bits. The rest of the paper is organized as follows. Section II briey explains the idea of signicant pixels and introduces HVS model used in our method. Section III describes how CV rule can be efciently utilized for this particular application. Section IV gives the simulation results with Lena image under various embedding conditions and attacks. Section V draws conclusions. II. P ROPOSED A LGORITHM Wavelet representation of any data gives information about the variations of the data at different scales. Wavelet detail image is obtained as the convolution of the image with the wavelet function dilated at different scales. We know from which signal points each wavelet coefcient at a particular scale is computed. We can further study the wavelet coefcients for the same points at a ner scale. These coefcients are called children coefcients. We have used three level wavelet decomposition using Haar wavelet to locate signicant pixels [11]. The three level wavelet decomposition is shown in Fig. 1. In the proposed

Fig. 1.

Three level wavelet decomposition with different bands

algorithm watermark is embedded into the signicant pixels in bands and  . But for calculating the signicance factor  we have considered all the bands except
 .

A. Watermark Embedding Let us denote the bands by

where B can be replaced by V, H or D as varies.

The sufx i denotes the level of wavelet decomposition in which that particular band is present. To locate the signicant pixels, choose every pixel in third level, say its corresponding children coefcients at all ner resolutions namely

signicance factor (S) of every pixel in band   




and

and . The

 !"$# #

6)98A@ B6"C8EDF" 021%' 34 &) ( 1  5 7 QRS8UT VQWC8UXY" 4 &)1 ( 1 1 P   G 1 HI%' `  !bab

is dened as follows.

(1)

After calculating signicance factor(S) at every pixel in bands dc and , these values are sorted. Pixels with higher signicance factor are watermarked rst. This essentially means that the pixels with maximum signicance factor are watermarked when a single copy is inserted. Moreover we expect higher the rst set of signicant pixels.

for the rst copy recovered from

fhgpi , obtained after randomizing. The watermark is embedded at every signicant pixel in band R  as
Let the watermark be represented by a column vector of size follows.

Here

3 w  C q   !rs   S#utve )  !


is the watermarked pixel and

@'xi B 6 " yyy

(2)

is the multiplication factor to keep the water-

mark below the level of perception. The value of

is unity if the watermark is binary.

The value of q, which is the maximum quantization at every pixel below the level of perception, is calculated using HVS model as given in [9]. The model presented in this paper is with reference to the four level decomposed image, where band
 in Fig. 1 is further decomposed into

S d c

and

. According to

this model maximum allowable distortion at every pixel is estimated as the weighted product of three different parameters.

w 0  !D !bD V D V ! d


Each term in the above equation is explained below. Here

w 0  !rw 0  !VI6

(3) (4) and

qDq

q q

denote the level of

decomposition and the orientation of the selected band respectively. The rst term 4

'D 

takes into account the sensitivity to noise depending on the band. Eyes are less sensitive to noise in high resolution bands and bands having orientation of Q 1.00, if l=0   0.32, if l=1 6 i  xi . D !  0.16, if l=2 i otherwise  0.10, if l=3 . 

The second term takes into account the local brightness based on the gray-level values of the low pass version of the image. Also it considers the fact that eyes are less sensitive to very dark and very bright regions of the image. In [10], this factor is computed in the following way.

bD V r ir# D V 


where

(5)

 P P! # 0 " i # P! # 0" D V   6  i # 6 6

(6)

Since eye is less sensitive to very dark regions as in the case of bright regions, in [9], this factor is modied as in the following equation.

q7D V ! 

i8$bD V !  D V  

 '&)( y0 o1325476Ie 9  8@4


i %

(7)

The third term takes care of the fact that eye is less sensitive to noise in highly textured areas but more sensitive near edges. P 0 ! A i A

D V ! 

3 F 0 IH#  3 G 3 3"4 i7 4 B 4 C 4 E 6 D yTSVU6XWE P ir#H # 6 P! 0 ir#YP#


A A

# 6  3 RQ P 0 a ! 6 `
3P

(8)

where the rst term gives the local mean square value and Var gives the variance around the 2x2 neighborhood of each pixel. After embedding the watermark, inverse transformation is performed to get the watermarked image. We have used Peak Signalto-Noise Ratio (PSNR) as measure of perceptual quality of the watermarked image. B. Watermark Detection and Evaluation For extracting the watermark from a possibly tampered image we need to use the original image and hence our algorithm is non-blind. Since watermark bits are embedded 5

only into the signicant pixels, we need to locate these pixels on the possibly attacked image and then get the quantization at those pixels. Thus we could get all inserted copies of the watermark. From these copies retrieved, we decide each watermark bit using CV rule. Normalized correlation coefcient( ) is dened as a measure of similarity between the original watermark (w) and retrieved watermark (w). Suppose there are N bits in the watermark, then normalized correlation coefcient is dened as follows. (9)

4 e e q  4 4 e e

III. O PTIMUM D ECISION F USION

USING

CV RULE

The embedded watermark is of binary nature and hence we need to decide whether a 1 or a 0 has been embedded in the selected pixels. When we consider very severe attacks, the retrieval becomes poor with a single copy of the watermark. Hence multiple copies are embedded and it is necessary to know how the corresponding bit from different copies can be fused together to form a good estimate while each copy is prone to large amount of error.

Fig. 2.

Bit aggregation from different watermarks

Let us assume that assume that

e 0

is the

D

bit of the



copy of the watermark. Let us also

watermark copies are embedded into the original data. While retrieving

the watermark, corresponding to each bit of the watermark, we receive

bits. Our aim

is to make an optimum decision on each of the watermark bit, based on these Let

be the probability of detection for any bit

e 0 . Then i 8

bits.

is the corresponding

probability of miss detection. Also, if a bit is not detected correctly, we surely detect it incorrectly. This makes the false alarm probability same as the probability of miss 6

detection. Also, let the a priori probabilities of detecting watermark bit 1 and 0 be and

respectively. Moreover, we assume them to be equal. Thus,

 G xi 8

for

any bit. The data fusion problem can be viewed as a binary hypothesis testing problem with corresponding bit from each watermark copy being the observations. The optimum decision rule is given by the likelihood ratio test.

e e 

0 Ve 0 Ve

0 Ve 0 Ve

PV0 "yyyy e PV0 "yyyy e

30 x  iW 8 V W 30  (  d 8 V d

(10)

The quantity on heft hand side is the likelihood ratio and Bayes optimum threshold is on the right hand side. Assuming the minimum probability of error criterion,

V 

( and

  d xi .

V
(11)

 e x  W i  e  (

Using Bayes rule to express the conditional probabilities, xiW e  ( e i The corresponding likelihood ratio test is
x  W i e  D   (  e 

(12)

(13)

But the left hand side of the equation is dened as follows [12].
x  iW e  sD D    ( e 

G A  A G  i 8  D D   #  #  i 8
and

(14)

We assume that the content owner has the copy of the original image and watermark for proving the ownership. We assume initial values for

according to the

correlation coefcient obtained between the original watermark and retrieved watermark. Thus all bits from one particular copy will be having the same

initially. Then we

compare each of the extracted bit with the corresponding bit in the original watermark. If both these are same we increase the

value by some amount  and decrease the

value by the same amount. This actually accounts for false positives. We have simulated results with 200 random watermarks, to check if any other watermark is detected or any false positive is produced by this method.

IV. E XPERIMENTAL R ESULTS We have chosen Lena image of size

i 6g i 6

and a binary watermark of size 7 i

g i7

for experimentation. Three level wavelet transform of the image is performed and the signicant pixels are sorted. Pixels with higher signicance factor are selected at rst for embedding watermark. The pixels are quantized as per HVS model. We have embedded a single watermark at rst on the highest signicant pixels and then increased the number of watermarks from one to eight. The Peak Signal to Noise Ratio (PSNR) variation with different number of embedded copies is shown in Table I.
TABLE I PSNR( B) VARIATION OF

WATERMARKED IMAGES WITH NUMBER OF COPIES

No of copies PSNR

47.46

45.00

43.58

42.61

41.83

41.11

40.58

40.20

Figure 1a shows the binary watermark chosen and Figure 1b and 1c show the watermarked images with a single copy and three copies of watermark embedded respectively.

(a) Watermark
Fig. 3.

(b) Watermarked(1copy)

(c) Watermarked(3copies)

Original watermark and watermarked image

In the following part, we show how the decision from different retrieved watermarks could be efciently fused together to form the best estimate of the watermark. We perform retrieval using CV rule and majority rule with different number of watermark copies. A. Robustness against attacks We have shown that a single copy of the watermark embedded gives very good robustness when embedded into wavelet signicant pixels retaining high PSNR [13]. 8

In this section, we show the results of some of the attacks which severely distort the watermarked image and the retrieval becomes very poor from such images. The improvement in detection with increase in number of watermark copies is obvious. Furthermore, substantial enhancement in detection with CV rule is also noticeable. This section includes the results of retrieval after different signal processing operations. The rst attack that is being considered is JPEG compression with very low quality factors. We present the results with quality factors 3, 5 and 10. From Table II, it is obvious that with 3 copies of watermark, we get a very good performance as far as the watermark retrieval is considered. It should also be noted that with 3 copies of watermark, majority rule could only give a correlation of 0.7185 which is not very good with a watermark of size 16x16. Figure 4 shows the JPEG compressed image with quality factor 3 and the retrieved watermarks from it using CV rule and majority rule when 3 copies of watermark are embedded. In this paper, for all attacks considered, we have included images with 3 copies of watermark embedded.
TABLE II C ORRELATION COEFFICIENTS AFTER JPEG C OMPRESSION
No of copies Majority rule Q=3 1 2 3 4 5 6 7 8 CV Rule Q=3 Majority rule Q=5 CV Rule Q=5 Majority rule Q=10 CV Rule Q=10

0.7172 0.7768 0.7185 0.7637 0.7763 0.8266 0.7895 0.8390

0.7172 0.8896 0.9673 0.9446 0.9739 0.9738 0.9837 0.9869

0.7860 0.8258 0.8202 0.8807 0.8649 0.8953 0.8506 0.9039

0.7860 0.9251 0.9902 0.9837 0.9967 0.9935 1 1

0.9412 0.9490 0.9737 0.9744 0.9804 0.9840 0.9967 0.9936

0.9412 1 1 1 1 1 1 1

Another attack that was performed on the watermarked image was the compression of the mean ltered image. The quality factor of compression was chosen as 10. The results of mean ltering alone and mean ltering with compression are shown in Table III. The attacked image with mean ltering and compression and the retrieved watermarks using CV rule and majority rule are shown in Figure 5. Histogram equalization with just 2 intensity levels was another attack that was performed on the watermarked image. The results of this attack is shown in Table IV and 9

(a) Compressed image(Q=3)


Fig. 4.

(b) CV Rule

(c) Majority rule

JPEG compressed image and retrieved watermarks using CV rule and majority rule TABLE III C ORRELATION COEFFICIENTS AFTER MEAN FILTER + JPEG (Q=10)
No of copies Mean lter Majority rule 1 2 3 4 5 6 7 8 Mean lter CV Rule Mean lter and JPEG Majority Rule Mean lter and JPEG CV Rule

0.9075 0.9326 0.9671 0.9627 1 1 1 1

0.9075 0.9869 1 0.9967 1 1 1 1

0.8253 0.8945 0.8664 0.8935 0.9203 0.9370 0.9438 0.9657

0.8253 0.9606 0.9902 0.9869 0.9902 0.9968 1 1

(a) Mean lter+Compression(Q=10)


Fig. 5.

(b) CV Rule

(c) Majority rule

Mean ltered and compressed image and retrieved watermark using CV rule and majority rule

3 copies of embedded watermark give a good retrieval with CV rule. At the same time, majority rule gives a very low correlation coefcient with retrieved watermark as shown

10

in Figure 6.
TABLE IV C ORRELATION COEFFICIENTS AFTER HISTOGRAM EQUALIZATION
No of copies 1 2 3 4 5 6 7 8 Majority rule CV Rule

0.6266 0.7427 0.6538 0.7731 0.6866 0.7686 0.7709 0.8111

0.6266 0.8452 0.9340 0.9484 0.9360 0.9620 0.9778 0.9811

(a) Histogram equalized image

(b) CV Rule

(c) Majority rule

Fig. 6. Histogram equalized image(only 2 levels) and retrieved watermark using CV rule and majority rule

We have tried other signal processing operations using GIMP software. They include Gaussian blur with different window sizes, linear blur and pixelization. The results with Gaussian blurred image with different window sizes are included in Table V. Moreover the attacked image with

window in which 3 watermark copies are inserted is shown in

Figure 7a. The retrieved watermarks from this image with correlation coefcients 0.9540 and 0.6435 by CV rule and majority rule respectively are also shown in Figure 7b and 7c. Linear blur is performed on the watermarked image and the retrieval is veried with both majority rule and CV rule. The results are included in Table VI. Three copies of the watermark could give a correlation coefcient of 0.9440 with CV rule and 0.6797 with majority rule as shown in Figure 8. 11

TABLE V C ORRELATION COEFFICIENTS AFTER G AUSSIAN BLUR


No of copies 3x3 Majority rule 1 2 3 4 5 6 7 8 3x3 CV Rule 5x5 Majority Rule 5x5 CV Rule 7x7 Majority Rule 7x7 CV Rule

0.7621 0.8783 0.8656 0.8833 0.9115 0.9439 0.9545 0.9687

0.7621 0.9509 0.9902 0.9838 0.9805 0.9935 0.9967 0.9967

0.6509 0.7799 0.7252 0.7808 0.7844 0.8566 0.8545 0.8962

0.6509 0.8958 0.9738 0.9574 0.9543 0.9707 0.9804 0.9837

0.6038 0.7522 0.6435 0.7248 0.6984 0.7758 0.7501 0.7957

0.6038 0.8693 0.9540 0.9416 0.9211 0.9577 0.9673 0.9737

(a) Gaussian blur (7x7)


Fig. 7.

(b) CV Rule

(c) Majority rule

Gaussian blurred image and retrieved watermark using CV rule and majority rule TABLE VI C ORRELATION COEFFICIENTS AFTER LINEAR BLUR
No of copies 1 2 3 4 5 6 7 8 Majority rule CV Rule

0.6312 0.7640 0.6797 0.7602 0.7045 0.7593 0.7193 0.7647

0.6312 0.8730 0.9440 0.9481 0.9281 0.9543 0.9707 0.9772

Pixelization was another attack that was considered. The results with

pixelization

are included in Table VII. The attacked image where 3 watermark copies are inserted 12

(a) Linear blur


Fig. 8.

(b) CV Rule

(c) Majority rule

Linear blurred image and retrieved watermark using CV rule and majority rule TABLE VII C ORRELATION COEFFICIENTS AFTER PIXELIZATION
No of copies 1 2 3 4 5 6 7 8 Majority rule CV Rule

0.5618 0.7000 0.5577 0.6523 0.5618 0.6224 0.5632 0.6394

0.5618 0.8015 0.9181 0.8968 0.8603 0.8912 0.9176 0.9347

(a) Pixelized (8x8)


Fig. 9.

(b) CV Rule

(c) Majority rule

Pixelized image and retrieved watermark using CV rule and majority rule

and the retrieved watermarks using CV rule and majority rule are shown in Figure 9.

13

B. False Positives We have initially mentioned that,

and

for each bit for any given copy of the

watermark would be either increased or decreased after comparing it with the actual bit in that position. This may result in false positives. To check for any false positives, we have simulated our proposed method with 200 random watermarks. However, not a single false positive was detected. We have selected 0.9 as the threshold value of correlation coefcient for watermark detection. The correlation coefcients with both CV rule and majority rule with Figure 10.

and

decided using random watermark samples are shown in

Fig. 10.

Correlation coefcients with random watermarks

V. C ONCLUSION In this paper, we have introduced optimum data fusion for perfect reconstruction of watermarks from the retrieved watermark copies under severe attacks. The detection/retrieval was proved to be robust under normal attacks from wavelet signicant pixels. But as the attacks become severe and distort the image to a high degree it becomes difcult to detect it. The detection in such a case could be improved by embedding multiple copies of watermark. Majority rule is usually considered for efcient detection from multiple copies. We have shown in this work that under such conditions, CV rule outperforms 14

majority rule. From experimentation, we could see that just three copies of watermark are sufcient for retrieval even when the attack severely distorts the image. R EFERENCES
[1] R. B. Wolfgang and E. J. Delp, A watermark for digital images, Proceedings of IEEE Int. Conf. Image proc., vol. 3, pp. 219222, Sept. 1996. [2] R.B. Wolfgang and E. J. Delp, Fragile watermarking using the vw2d watermark, Proceedings of SPIE, vol. 3228, pp. 297308, Nov. 1997. [3] J. J. K. O. Ruanaidh, W. J. Dowling, and F. M. Boland, Phase watermarking of digital images, Proc. Int. Conf. Image Processing, vol. 3, pp. 239242, Sept. 1996. [4] C. T. Hsu and J. L. Wu, Hidden signatures in images, Proc. Int. Conf. Image Processing, vol. 3, pp. 223226, Sept. 1996. [5] D. Kundur and D. Hatzinakos, A robust digital image watermarking scheme using wavelet based fusion, Proceedings of IEEE Int. Conf. Image proc., pp. 544547, Oct. 1997. [6] H. Inoue, A. Miyazaki, A. Yamamoto, and T. Katsura, A digital watermark based on wavelet transform and its robustness on image compression and transformation, IEICE Trans. Fund. Electron., Commun., Comput. Sci.,, vol. E82-A, pp. 210, Jan. 1999. [7] X. Xia, C. J. Boncelet, and G. R. Arce, A multiresolution watermark for digital images, Proceedings of IEEE Int. Conf. Image Proc., pp. 548551, Oct. 1997. [8] Masoud Alghoniemy and Ahmed H. Tewk, Geometric invariance in image watermarking, IEEE Trans. Image Proc., vol. 13, pp. 145153, Feb. 2004. [9] Mauro Barni, Franco Bartolini, and Alessandro Piva, Improved wavelet based watermarking through pixel-wise masking, IEEE Trans. Image Proc., vol. 10, no. 5, pp. 470477, May 2001. [10] A. S. Lewis and G. Knowles, Image compression using 2-d wavelet transform, IEEE Trans. Image Proc., vol. 1, pp. 244250, Apr. 1992. [11] Q. Tian, N. Sebe, M. S. Lew, E. Loupias, and T. S. Huang, Content-based image retrieval using wavelet-based salient points, Proceedings of SPIE, vol. 4315, pp. 425436, Dec. 2000. [12] Z. Chair and P. K. Varshney, Optimal data fusion in multiple sensor detection systems, IEEE Trans. Aerospace and Electronic Systems, vol. AES-22, pp. 98101, Jan. 1986. [13] M Jayalakshmi, S. N. Merchant, and U. B. Desai, Signicant pixel watermarking using human visual system model in wavelet domain, Lecture Notes in Computer Science, vol. 4338, pp. 206215, Dec. 2006.

15

You might also like