International Journal of Scientific & Engineering Research Volume 4, Issue 1, January-2013 1

ISSN 2229-5518

VLSI Architecture of Invisible Watermarking using Lsb and Lifting Scheme for Real time Copyright Protection

Kondeti shyam (M tech)

Department of Electronics and communication Engineering

Malla reddy institute of technology and science

Hyderabad, India

shyamk.4u@gmail.com

Index Terms - Blind watermarking, wavelet transform, quantization, geometric attacks

I. INTRODUCTION

With the rapid development of the digital multimedia technique, the need for the copyright protection of multimedia data becomes more crucial. In order to meet the requirement for multimedia data security, various kinds of watermarking techniques have been proposed.

Generally, we can classify current watermarking techniques as two classes: algorithms processed in spatial domain and algorithms processed in frequency domain, such as DCT and DWT. As the current image compression standard JPEG 2000 is based on DWT, much more attention has been paid to the wavelet-based watermarking algorithms.

There are three key requirements that should be satisfied by all effective watermarking methods: invisibilit y, robustness, and the ability of watermarking detection. Robustness includes the ability to resist the attacks of common image processing (filtering, compression, etc) and the ability to resist geometric attacks (rotation, scaling, cropping, etc). For most of the watermarking methods the robustness against geometric attacks is vital for the detection of watermark. Although geometric attacks do not cause serious visual distortions, they can severely affect the watermark detection by changing the detected position of the embedded watermark. So robustness against geometric attacks is a pressing necessity for the watermarking technique. Wavelet-based watermarking methods differ in many ways. In [1], the perceptually significant coefficients for each subband are selected to embed the watermark. These perceptually significant coefficients are chosen by using different scale fact ors depending on the

level of decomposition

————————————————

*K.shyam is currently pursuing masters degree program in embedded systems and vlsi dsign engineering inJntuh University, INDIA, PH-*

9848829577. E-mail: shyamk.4u@gmail.com

. In [2], a different number of watermarks are embedded into each band in proportion to the energy contained in the band. In [4], the wavelet coefficients are grouped into so-called wavelet trees. Quantizing one of two certain wavelet trees embeds one watermark bit. In [6], the watermark is embedded by combining the rotation-invariant log-polar mapping (LPM) with the wavelet transform. The method is resilient to geometric attacks.

This paper proposes an improved blind wavelet-based watermarking scheme, which integrates the wavelet super tree quantization with the pixel-wise masking. The wavelet coefficients are ordered into so-called wavelet super trees, similar to [7]. Each of the watermark bits is embedded by quantizing one of two certain wavelet super trees by means of quantization illustrated in [3]. The disparity between super tree pairs is used to extract the watermark, which is similar to [8]. In order to obtain a better tradeoff between the robustness and the imperceptibility, the pixel-wise masking in [5], which is later modified to adapt for watermarking embedding by [9], is combined with the wavelet super tree quantization. The pixel-wise masking used in this paper is adjusted according to the characteristics of the super tree quantization. Moreover, the watermark is embedded into the bands in the three different orientations redundantly to deduce the false positive probability. Comparing with the method in [6], the proposed algorithm is much easier and suitable for the real time application. From the experiment results shown in section 4, we can see that while the PSNR of the watermarked image is higher, the robustness is stronger than the method in [7].

This paper is organized as follows. SectionĊ illustrates the proposed watermarking scheme in detail. In sectionċ, the simulation results are presented, including geometric attacks and common image processing. Finally the conclusion is given in section Č.

II. THE PROPOSED WATERMARKING SCHEME

IJSER © 2013 http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 4, Issue 1, January-2013 2

ISSN 2229-5518

The original image is transformed into wavelet coefficients. We use 4-level wavelet transform of a 512x512 image as example. We have 13 frequency subbands as shown

in Fig 1.

Firstly, all the wavelet coefficients are quantized to integers, and then these coefficients are described in binary representation as shown in Fig 4.

A Wavelet

A. Watermark Embedding Tree 1 2 21

1) Wavelet trees

As the high-frequency bands at the forth level contain little energy and HVS is much sensitive to the noise in the low-frequency band at the first level. So we only use the

xn(1) xn(j)

A Wavelet

Super Tree

**... **xn(L)

coefficients in subbands HL4, HH4, LH4, HL3, HH3, LH3,

HL2, HH2, LH2 to form wavelet trees.

(1,1) (1,512)

A Wavelet

Tree

1 2 21

Fig. 3 A wavelet super tree built up by t wo wavelet t rees .

LL4 HL4

LH4 HH4

HL3

HL2

Sign bit

p

1 2 3 ... bn 41 L

fficient s

LH3

LH2

HH3

HH2

HL1

MSB 2

2an+1

2an

LH1 HH1 LSB 20

(512,1) (512,5 12)

Fig. 1 Tree structure of wavelet coefficients. LH4 LH3 LH2

Bit: i

Fi g. 4 Quantization of t he nth wavelet super t ree respect to qn

The sign of every coefficient in the super tree is stored in the first row so called sign bit. The least significant bit (LSB) is 20 , and the most significant bit (MSB) is denoted by 2P

.

1 The total number of the bits is N p L( p 1) .

Quantization index *q*n is a factor to control the strength of quantizing the *n*th tree. The index *q*n decides the number of bits discarded after the quantization. The larger *q*n is, the more

Fi g. 2 Ordered coefficients of a wavelet t ree. bits will be discarded. The coordinate of *q*n in the array is

As shown in Fig 1, the subbands HL4, HH4, LH4 is used an , bn as shown in Fig 4. All the bits below the quantization as roots to form wavelet trees. Each wavelet tree consists of index *q*n which are shown as shaded area will be discarded.

1+4+16=21 coefficients, shown in Fig. 2. As the total number

of trees is equal to the number of coefficients in subbands

HL4 HH4 and LH4 so there are a total of 3x32^2=3072

Wav tree

2) Super trees

Two wavelet trees form one super tree with the length of

Thus, the coefficients will be changed by the following equation

Q[x (

¯ n

L=42 as shown in Fig. 3. So there are totally 3072/2=1536 Let

round ( xn

( *j*))

n

denote the rounding of th

wavelet super trees Tn , for n=1,…,1536. coefficient *x *( *j*)

xn(j)

to the *a*nth bit. The quantization of

with respect to qn is denoted by Q[xn(j)]qn

3) Quantization of super trees [3-4] : .

Ever y two super trees are used when one watermark bit

is embedded. So the watermark bit capacity of the image is

1536/2=768 bits. In order to improve the robustness of the watermarked image, the watermark is repeatedly embedded into the bands in the three different orientations for three times. Here we generate a PN sequence of length Nw=256 as

the watermark *W *with a certain key.

The quantization step size / n ( *j*) is denoted by

Let the *j*th coefficient of the nth super tree be denoted by The quantization error of the *j*th coefficient with respect to *q*n is given by

IJSER © 2013 http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 4, Issue 1, January-2013 3

ISSN 2229-5518

The total quantization error of the *n*th tree with respect to

qn is

L

Step 2: Achieve a PN sequence W of length Nw using a key K. To guarantee the security, the key K also can be used to order the wavelet super trees in pseudorandom manner. Set

[n (*q*n ) ¦ *e*n ( *j*) (5) qn 1,

n

1 , *m *0 ;

j 1

4) Watermark Insertion

Step 3: Set qn

qn 1 to find the largest qn

that satisfies

Two trees T2n-1 and T2n are used to embed one watermark (6), if there is no such qn , let qn qmax . And quantize the

bit *w*n . If wn 1 , *T*2n-1 is quantized according to *q*n . If *w*n 1 , super trees *T*n with respect to *q*n . Thus we get the quantized

T2n is quantized according to qn . During the quantization procedure we need find the largest quantization index qn to

super trees *T1*n;

Step 4: If wn = -1 , let T 2

n -1

=T12 *n *-1 , T=

22 n

T2 n ;

satisfy the following inequation If *w*n =1 , let

T 22 *n *-1 =T2 *n *-1

,

T 22 *n *=T12 *n *. Set n=n+1. If

[2n 1 (qn ) [ [2n (qn ) [ (6) n<Nw, go to step 3;

[ is the reference error which can control the strength of *Step 5: Set m m *1 , *if m *3 , *set n *256 u *m *1 , *and go to step 3. m help to embed the watermark bits into the*

the quantization. If there is no largest *q*n to satisfy (6), we

bands in the three different orientations for three times;

will let *q*n *q*max . *q*max is a pre-determined value. *Step 6: R*egroup the modified super trees *T *2n to form the

5) Design of the weighted factor

We use the method in [9] for reference to make the pixel-

wise masking of the proposed algorithm. In [9], the following consideration about human eye’s sensibility to noise is taken into account.

• The eye is less sensitive to noise in high resolution bands and in those bands having orientation of 450.

Based on this consideration, we modify the weighted product 4 for each wavelet coefficient in [9] to combine with

The wavelet tree quantation as(7)

(7)

modified wavelet coefficients. It is an adverse process of

grouping wavelet coefficients to super trees;

Step 7: The IDWT is applied on the modified wavelet coefficients. Quantize the output of IDWT to integers. Then the watermarked image is obtained.

. Watermarking Extraction

1) Designs of the watermark decoder

Let *N*2n-1 is the number of coefficients in *T *'2 n 1 which satisfy the following equation

This term refers from the sensibility to the noise *T *'n is the super trees of the watermarked image,

according to the change of bands and particularly according to

the orientation. The parameter *l *denote the resolution level of

the band and the parameter Ӱ denote the orientation of the*e *'n ( *j *) and / 'n ( *j*) is respectively the quantization error and the quantization step size of the *j*th coefficient in super tree

band (Ӱ=0 means vertical orientation;Ӱ=1 means diagonal orientation;Ӱ=2 means horizontal orientation). Equation (7)

T 'n

let H

. H is a constant. Based on the literature [7], we

0.1 . And we get the watermark decoder corresponding

means the biggest of 4 is achieved by the diagonal high- frequency bands in the second, third and fourth resolutation levels(o=1;,l=2,l=3,l=-4).

Then we design the reference error [ renewedly as (8)

with the watermark insertion method as below

2) Detection of watermark bit

(8) To extract the watermark *W *' , the DWT is realized Therefore, the energy of the watermark embedded in the watermarked image. Just as what we have discussed above, diagonal high-frequency bands is stronger than the watermark we obtain three versions *W*1 ' , *W*2 ' , *W*3 ' through extracting

embedded in other two high-frequency bands. In this way, we can strengthen the energy of the embedded watermark with

watermark from the bands in the three different orientations.

By (12), then the final watermark *W *' is achieved.

respect to the invisibility of the watermark. According to the *W *' = *W*1 ' + *W*2 ' + *W*3 (12)

IJSER © 2013

International Journal of Scientific & Engineering Research Volume 4, Issue 1, January-2013 4

ISSN 2229-5518

experiment results shown later we can see that the robustness of the watermarked image is improved apparently.

6) Embedding procedure

The result is compared with a threshold UT . If U (W

,W ') t UT , the image is marked by the watermark W ;

otherwise, the image is not marked by the watermark W .

The probability of the false positive error *P*fp can be computed as follows based on [3]

N w § *N *·

A. Imperceptibility

The watermarked image has PSNR=40.6db which means that the watermark is almost imperceptible. Also from the original image and the watermarked image shown in Fig.5, we can see that the watermark is imperceptible.

For better comparison, we respectively detect the

watermark from the original image and the watermarked

P f p ¦ ¨ k ¸ E

k (1 P E )k

(14) image without any attack. The U for original image Uo

0.03

k E © ¹ and the U for watermarked image U

0.80 .

E (( UT 1) / 2)*N*w , *P*E Pr *ob*(*w*n z *w *'n ) , it is

reasonable

B. Robustness

to assume PE 0.5 . We can choose the value of UT according *1) Rotation*

to the requirement of the false positive probability.

3) Extraction procedure

Step 1: Apply a 4-level DWT decomposition on the

The watermarked image is rotated by a small angle with the step of 0.250. TABLE I reports the obtained results against

rotation attacks. It can be seen that the proposed algorithm can

512 u 512 *watermarked image, and form super trees *resist a rotation of up to 2.250

T 'n (n=1,…,6Nw )using the wavelet coefficients except the TABLE I.

coefficients of bands LL4, HL1, HH1, LH1;

Step 2: Use the key K to order the wavelet super trees in

WATERMARK EXISTENCE IN APPLYING ROTATION ATTACKS TO THE

WATERMARKE D IMAGE “LENA” (ROTATION UNIT: DEGREE ; +: CL OCKWI SE; -

pseudorandom manner. Set qn

' 1 , *n *1 , *m *0 *;*

:COUNTE RCL OCKWISE)

Step 3: Set qn ' qn ' 1 to find the largest qn ' that satisfies

(6), *if there is no such q*n

' *, let q*n ' *q*

max

. And quantize the

super trees T 'n with respect to qn ' , thus we get the quantized super trees T1'n ;

Step 4: Compute N2n and N2n-1 using (9) and (10). Set

n=n+1. If n<Nw, go to step 3;

Step 5: Set m m 1 , if m 3 , set n 256 u m 1 , and go to step 3. m help to embed the watermark bits into the

bands in the three different orientations for three times;

Step 6: Then use (12) to get the extracted watermark W ' . Use (13) to compute U (W , W ') . If U(W, W ') t UT , the image is marked by the watermark W; if U (W ,W ') UT , the image is not marked by the watermark W.

III. SIMULATION RESULTS

The 512 u 512 image “LENA” is used for our experiments. The original image and the watermarked image are shown in Fig.5. A 4-level DWT decomposition is performed using Daubechies-8 filters. Here we choose the value of the pre-determined parameters as follows

TAB LE Ċ.

WATERMARK EXISTENCE IN APPLYING CROP PING ATTACKS TO THE WATERMARKED IMAGE “LE NA” (CROPPING FACTOR IS THE RATIO OF THE RE MAI NE D CENTER PART TO THE WHOLE WATERMARKE D IMAGE)

Cropping Factor | 0.75 | 0.50 | 0.25 |

U | 0.61 | 0.48 | 0.26 |

Exi stence | Y | Y | Y |

TABLE ċ.

WATERMARK EXISTENCE IN APPLYING SCALING ATTACKS TO THE

H 0.1 , [ 215 , qmax

336 , N w

256 , UT

5

0.26 . So the

WATERMARKED IMAGE “LENA” (SCALING FACTOR IS THE RATIO OF THE

SCALED UP/DOWN WATERMARKED IMAGE SIZE TO T HE ORIGINAL

false positive probability Pfp 2.22 u10

. WATERMARKED IMAGE )

Scaling Factor | 0.75 | 0.50 | 0.25 | 10 | 20. | 30. 40. |

U | 0.48 | 0.35 | 0.32 | 0.80 | 0.80 | 0.80 0.80 |

Exist ence | Y | Y | Y | Y | Y | Y Y |

Ori ginal I mage Water marked I mage

(PSNR= 40.6.)

Fig. 5 Origi nal Image and Watermarked I mage

TABLE Ҋ.

WATERMARK EXISTENCE OF AP PLYING JPEG COMPRE SSION ATTACKS TO THE

WATERMARKED IMAGE “LE NA”

IJSER © 2013

International Journal of Scientific & Engineering Research Volume 4, Issue 1, January-2013 5

ISSN 2229-5518

TABLE V

WATERMARK EXISTENCE OF APPLYING COMMON IMAGE PROCE SSING

ATTACKS TO THE WATE RMARKED IMAGE “LE NA”

Image Pro cessi ng Att acks | U | Existence |

Median Filter (size 3x3) | 0.35 | Y |

Median Filter (size 4x4) | 0.26 | Y |

Sharpeni n g | 0.3 8 | Y |

Gaussian Filter (size 3x3) | 0.70 | Y |

TABLE IV

WATERMARK EXISTENCE COMPARISON OF THE PROP OSED ME THOD AND THE

METHOD IN [7] WITH THE SAME ATT ACKS

Different Attacks | Watermark Existence | |

Different Attacks | Ref [7] (PSNR= 38.2 db) | Proposed (PSNR= 40.6db) |

Rotation (degree : 1.00) | No | Yes |

Rotation (degree : 2.25) | No | Yes |

JPEG ( QF=3 0) | No | Yes |

JPEG ( QF=2 0) | No | Yes |

Median Filter (4x4) | Yes | Yes |

Gaussian Filter | Yes | Yes |

Sharpening | Yes | Yes |

2) Cropping

The circumference of the original watermarked image is cropped and the remained center part of the watermarked image is used to extract the watermark. The results are shown in TABLE Ċ. The watermark can still be detected when the cropping factor is 0.25.

3) Scaling

The watermarked image is scaled up or down to examine the algorithm robustness. The results given in TABLE ċ show that the proposed method is still effective when the scaling factor is down to 0.25 or up to 40.

4) Compression

For JPEG compression, the quality factor from 90 to 10 is used, and the results are shown in TABLE Ҋ. From Tabled 4 we can see that even with quality factor low to 20, the proposed scheme can detect the existence of the watermark.

5) Common image processing

Median filter, Gaussian filter, and sharpening are applied to the watermarked image to test the algorithm robustness against common image processing. The results are shown in

TABLE V. The proposed algorithm can resist median filter of window size 4x4, Gaussian filter of window size 3x3 and also sharpening.

Additionally, we compare the proposed scheme with the method in [7] using the image “LENA”. The results are shown in TABLE Ď. We can see that as to the common image processing attacks the performance of the two methods is similar. However, with respect to rotation and JPEG compression, the proposed method is notably more robust.

IV. CONCLUSIONS

In this paper, a blind wavelet-based watermarking method using the HVS and the quantization of wavelet super trees is presented. The wavelet coefficients are ordered in a certain manner to form the wavelet super trees. The energy of the inserted watermark is modified by the adjusted pixel-wise masking depending on the human eye’s sensibility to the noise. The watermark bits are embedded repeatedly into bands in the three different orientations. The proposed method apparently improves the performance of robustness against the common image processing, especially against the geometric attacks and JPEG compression.

REFERENCES

[1] J. R. Kim, Y. S. Moon, “A Robust Wavelet-Based Di gital Watermarking

Using Level-Adaptive Thresholdi ng”, in Proc. ICIP, vol. 2, pp. 226– 230,

1999

[2] Y. S. Kim, O. H. Kwon, R. H. Park, “W avelet Based Watermarking Method for Di gital Ima ges Using the Human Visual System” , in Proc. IEEE ISCAS, vol. 4, pp. 80-83, 1999.

[3] D. Kundur and D. Hatzinakos, “Digital watermarkin g, using

multiresolution wavelet decomposition,” in Proc. IEEE ICASSP, vol . 5,

1998, pp. 2969– 2972.

[4] S.-H.Wang and Y.-P. Lin, “Blind watermarking using wavelet tree quantization,” in Proc. IEEE ICME, vol. 1, 2002, pp. 589– 5 92.

[5] A. S. Lewis, G. Knowles, Ϙ Image Co mpressio n Using the 2-D Wavelet

Transform ϙ, IEEE Transactions on Image Processin g, Vol. 1, No. 2, p p

244 – 250, April 1992

[6] G. Xie, H. She n,ϘRob ust Wavelet-Based Blind Image Watermarking

Against Geometric Att acksϙ, in Proc. IEEE ICME, vol . 3, pp. 2051-2054,

2004.

[7] S.-H.Wang and Y.-P. Lin, “Wavelet Tree Quantization for Copyright Prote ction W atermarking ” , IEEE Transacti ons on Image Processing, VOL. 13, NO. 2, pp. 154-165, February 200 4.

[8] G. C. Langelaar and R. L. Langendijk, “ Optimal differential energy

watermarking of DCT encode d i mages and video,” IEEE Trans. Image

Processing, vol. 10, pp. 148–158, Jan. 2001.

[9] Mauro Barni, Franco Bartolini, Alessandro Pi va, ϘImproved Wavelet- Based Watermarking Trou gh Pixel-Wise M as kingϙ IEEE Transactions on Image Processing, Vol. 10, No. 5, pp. 783-791, May 2001

IJSER © 2013