D.-S. Huang, K. Li, and G.W. Irwin (Eds.): ICIC 2006, LNCS 4113, pp. 709 – 716, 2006. © Springer-Verlag Berlin Heidelberg 2006

Genetic Algorithm-Based Watermarking in Discrete Wavelet Transform Domain

Dongeun Lee1, Taekyung Kim1, Seongwon Lee2,*, and Joonki Paik1

1 Image Processing and Intelligent Systems Laboratory, Department of Image Engineering, Graduate School of Advanced Imaging Science, Multimedia, and Film,

Chung-Ang University, Seoul, Korea 2 Department of Computer Engineering, College of Electronics and Information,

Kwangwoon University, Seoul, Korea ehddms98@wm.cau.ac.kr

Abstract. This paper presents a watermarking algorithm in the discrete wavelet transform domain using evolutionary algorithm. The proposed algorithm con-sists of wavelet-domain watermark insertion and genetic algorithm-based watermark extraction. More specifically watermark is inserted to the low-frequency region of wavelet transform domain, and watermark extraction is efficiently performed by using the evolutionary algorithm. The proposed wa-termarking algorithm is robust against various attacks such as JPEG image compression and geometric transformations.

1 Introduction

Digital watermarking is a digital content copyright protection technique against unau-thorized uses such as illegal copy, distribution, and forgery. Digital watermarking inserts and extracts copyright information called watermark into the digital contents to prove the ownership of the copyright holder.

Watermark insertion can be done in either spatial domain or frequency domain. The spatial-domain watermark insertion manipulates image pixels, especially on least significant bits that have less perceptual effect on the image. Although the spatial-domain watermark insertion is simple and easy to implement, it is weak at various attacks and noise. On the other hand, the frequency-domain watermark insertion, which is robust at attacks, is made on the frequency coefficients of the image. The existing frequency transformation methods for watermark insertion include discrete Fourier transform (DFT) [1, 2], discrete cosine transform (DCT), and discrete wave-let transform (DWT) [3, 4].

Despite of recent progresses, existing watermarking algorithms are not sufficiently robust against JPEG compression and geometrical distortions such as translation, rotation, scaling, cropping, change of aspect ratio, and shearing. These geometrical distortions cause the loss of geometric synchronization that is necessary in watermark detection and decoding [5]. Two different types of solutions to resist geometrical * Corresponding author: swlee@kw.ac.kr

710 D. Lee et al.

attacks are: non-blind and blind methods [6]. In the non-blind approach, with the origi-nal image, the problem can be well resolved by effective search between the geometri-cally attacked images and the original ones [7, 8]. The blind solution that extracts watermark without the original image has wider applications but is obviously more challenging.

This paper proposes a more robust blind watermark extraction method using evolu-tionary algorithm against geometric attacks. The proposed algorithm inserts water-mark in the wavelet domain, and the genetic algorithm (GA) searches and automati-cally extracts watermark. Since the searching process generates lots of local matching, the correct alignment can be determined at the end of exhausted computation. The proposed algorithm can reduce the computational burden by using evolutionary searching method.

This paper is organized as follows. In section 2 we describe the watermark inser-tion method in the wavelet domain, and in section 3 a watermark extraction method is presented using GA. Experimental results with various types of attacks are given in section 4, and section 5 concludes the paper.

2 Watermark Embedding Algorithm

The proposed frequency-domain watermark insertion algorithm is performed in the DWT domain. Since the low frequency band remains robust to attacks, the watermark insertion in the proposed algorithm is carried out in the LL3 band for a 3-level DWT. It should be noted that watermark insertion should be carefully designed since the coefficients of LL3 band have strongest signal energy. Strong watermark could be visible at the LL3 band. Watermark is specially designed user information that is represented by images, text characters, sound data, and so on. In this paper, we used a watermark image because of the convenience of visual analysis and evaluation. Fig. 1 shows the block diagram of watermark insertion of the proposed algorithm. The reso-lution of the original image is 512×512. A 2-bit watermark image of size 64×64 is used in gray scale space.

Original Image

DWTDomain

Watermarkedin LL3 Subband

Logo Image

DWT

CoefficientsSetting

IDWT WatermarkedImage＋＋

Fig. 1. The block diagram of the watermark embedding procedure

The algorithm for embedding the binary watermark is; to apply a 3-level DWT with Daubechies D4 wavelet filters to an input original image f(x,y) of size 512×512×8bits, and then to check the magnitude of LL3 coefficients and find the

Genetic Algorithm-Based Watermarking in Discrete Wavelet Transform Domain 711

location of bits where the binary logo image will be inserted. Finally, the proposed method inserts bit patterns of watermark using a pre-specified threshold.

3 Proposed GA-Based Watermark Detection Algorithm

Watermark detection algorithm consists of two stages; the formation of populations using GA and watermark extraction in the LL3 band in the DWT domain. The water-mark is then extracted in the LL3 band. GA realizes efficient watermark extraction by selecting superior genes and mating them to generate mutation. Fig. 2 shows the block diagram of the proposed watermark detection algorithm. The following subsection describes the watermark extraction process in the DWT domain.

PopulationFormation

DWTWatermarkDetection NC > T

yes

no

ImageTransform

WatermarkedImage

Watermark

Fig. 2. The block diagram of watermark detection procedure

GA is a computational model to solve a real-world problem by simulating a natural evolution process in a computer. Application areas of GA includes nonlinear, non-differentially problems with multiple extremes. GA is particularly suitable for adaptive searches and optimization. The proposed evolutionary algorithm adopts GA to extract the watermark out of the image deformed by geometry attacks such as rota-tion and translation. Initially, 20 genes (parents) are generated and applied to the at-tacked image to reverse attacks. Then, the embedded watermark is extracted and the fitness of the extracted watermark is measured. The genes with best fitness are se-lected and used for the next generation. We generate 20 genes by crossover and muta-tion with a predetermined rate. The process is repeated until the best gene is found. In the proposed algorithm the chromosome consists of a 16-bit string of 0 and 1. In the chromosome structure, two 4 bits are used for translations in both horizontal and vertical directions respectively and 8 bits are used for rotation attack.

Table 1. Parameters of genetic algorithm

Parameters Values Population size 20

Chromosome length 16 Maximum number of generation 20

Selection rate 0.8 Crossover rate 0.8 Mutation rate 0.2

Table 1 represents the parameters used in the experiments. The parameters are re-ferred to Ali’s paper [9] and fixed by experiments. The probability of crossover is empirically set to 0.8%, and that of mutation to 0.5%.

712 D. Lee et al.

In order to evaluate the feasibility of extracted watermark, normalized correlation (NC) is used. The number of mismatched data between the inserted and the extracted watermarks is used to represent the similarity of watermarks. NC for valid water-marks, which represents the characteristics of the extracted watermark, is defined as

∑

∑ ′=

yxyx

yxyxyx

w

ww

NC

,

2,

,,,

, (1)

where w represents the inserted watermark, w´ the extracted watermark. Experi-mental results are rounded to the fourth decimal place. The NC for random noise is about 0.5 and possibility of distinguishing extracted logo is higher than 0.7~0.8 NC.

Fig. 3 (a) shows the variation of the NC of the extracted watermark when the wa-termarked image is horizontal translation. As shown in the figure, watermark can be extracted every 23 pixels because 3-level DWT was used. Fig. 3 (b) shows the varia-tion of the NC of extracted watermark with rotation. In case of rotation, mutation is generated every 90 degree. Each population is decided to have high feasibility if it has high NC. As the selected genes of high NC are transferred to the next generation, the populations in the new generation also have high NC. Finally, high NC watermark can be extracted.

(a) Horizontal translation (b) Rotation

Fig. 3. Variation of NC with (a) translation and (b) rotation

The proposed watermark extraction algorithm uses GA to make the watermark ex-traction process robust against various types of attacks. GA generates populations, and transforms the image for watermark search. The extraction procedure is:

Step 1: GA generates populations and their feasibilities are evaluated. Highly feasi-ble genes are selected and mated to generate mutation.

Step 2: Watermark-inserted populations are transformed and applied the 3-level DWT to the restored image.

Step 3: In order to evaluate the feasibility of extracted watermark, normalized corre-lation (NC) is measured. If NC is larger than a predetermined value, the ex-tracted watermark is decided to be feasible, and the algorithm terminates. Otherwise go back to step 1.

Genetic Algorithm-Based Watermarking in Discrete Wavelet Transform Domain 713

4 Experimental Results

The performance of the proposed algorithm is tested with various types of images. A test image is an 8-bit grayscale Lena image of size 512×512. The watermark is a 64×64 binary image. The original Lena and the watermark image are respectively shown in Fig. 4 (a) and (b). Daubechies D4 filter coefficients are used for 3-level wavelet decomposition. Performance evaluation of the proposed algorithm on test images with various characteristics was studied. Peak signal-to-noise ratio (PSNR) is used to analyze the quality of the watermarked image, which is defined as

,)(

1255

log10

,

2,,

2

10

∑ ′−=

yxyxyx II

XY

PSNR

(2)

where I represents the original image, I´ the modified image, and X and Y represent the horizontal and vertical image sizes respectively.

(a)

(b) (c)

(d)

Fig. 4. (a) The original image, (b) the watermark, (c) the watermarked image, and (d) the ex-tracted watermark image

The watermarked Lena image with PSNR 42.11db is shown in Fig. 4(c). There is no perceptual degradation in the watermarked image. The extracted watermark from the watermarked image is shown in Fig. 4(d).

4.1 Non-geometric Attacks

We classify the attack patterns into non-geometric attacks and geometric attacks. Watermarked images are first tested for non-geometric attacks such as Gaussian filter-ing, median filtering and compression. The proposed algorithm can easily cope with wavelet-based JPEG2000 and DCT-based JPEG as shown in Table 2 and 3. Table 2 shows the performance of the proposed algorithm against compression attacks from 0.5 to 0.1 of JPEG2000 compression rates. The extracted watermarks are seriously damaged below the rate 0.1. JPEG compression is also evaluated over various com-pression rates. Table 3 presents the PSNRs and NCs according to JPEG quality fac-tors. The PSNR values shown value between watermarked image and attacked image. In other words, PSNRs express the extent of degradation by attacks.

714 D. Lee et al.

We also include non-geometric attacks such as Gaussian, sharpening, and median filter. The PSNR of the attacked images are shown in Table 4. The attack patterns used in the experiment are from the Korean Watermarking Certification [10].

Table 2. PSNRs and NCs of watermark embedded images at JPEG2000 compression

Lena Lake Boat JPEG2000 Rate PSNR NC PSNR NC PSNR NC 0.50 49.87 1 48.13 0.999 48.57 0.999 0.45 48.54 0.998 48.13 0.974 48.13 0.999 0.40 48.13 0.997 45.12 0.972 48.13 0.973 0.35 48.13 0.968 43.36 0.961 45.12 0.978 0.30 45.12 0.969 41.14 0.973 42.11 0.894 0.25 43.36 0.968 39.68 0.858 40.35 0.892 0.20 41.14 0.903 37.34 0.773 38.59 0.857 0.15 39.68 0.854 35.58 0.692 36.99 0.731 0.10 37.72 0.773 33.36 0.601 34.91 0.667

Table 3. PSNRs and NCs of watermark embedded images at JPEG compression

Lena Lake Boat JPEG Quality PSNR NC PSNR NC PSNR NC

10 43.36 1 42.11 1 42.11 1 8 39.68 0.998 36.67 0.999 37.72 0.992 6 39.10 0.988 35.58 0.989 36.99 0.991 4 36.67 0.915 33.51 0.917 34.71 0.917 2 34.15 0.832 30.97 0.830 31.80 0.823 0 32.11 0.622 29.27 0.614 30.00 0.622

Table 4. PSNRs and NCs of watermark embedded images at another attack method

Lena Lake Boat Attack

PSNR NC PSNR NC PSNR NC Gaussian 26.52 0.808 26.99 0.854 25.80 0.892

sharpnening 18.67 0.735 15.13 0.639 16.23 0.671 Median(3×3) 26.23 0.873 26.28 0.811 25.19 0.854 Median(5×5) 22.40 0.738 21.81 0.710 21.27 0.696

4.2 Geometric Attacks

The geometric attacks are further divided into translation, cropping, rotation, flip, and scale. Fig. 5(a) shows a graph for finding the best fitness using proposed algorithm in the watermarked image with an attack of 15 degree rotation, and (b) shows the ex-tracted watermark with NC value of 7.842. As the generations pass, GA finds the optimal fitness by generating better populations. The geometric parameters are evalu-ated to determine the geometric attack patterns, and the performance of the watermark extraction is listed in Table 5.

The proposed watermarking algorithm can endure translation attacks if the transla-tion is more than 20%, horizontal or vertical flip attacks, and all degree rotation at-tacks. The watermark can be identified with cropping attacks up to 50%.

Genetic Algorithm-Based Watermarking in Discrete Wavelet Transform Domain 715

(a)

(b)

Fig. 5. (a) A graph of finding the best fitness with an attack of 15 degree rotation, (b) and the extracted watermark

Table 5. Experimental results for the geometric attacks

Geometric attacks Lena (NC)

Lake (NC)

Boat (NC)

Goldhill (NC)

Drop (NC)

Peppers (NC)

Translation (0.5 pixel) 0.978 0.993 0.963 0.981 0.987 0.994 Translation (10 pixel) 0.977 0.981 0.976 0.979 0.986 0.980 Translation (30 pixel) 0.941 0.954 0.937 0.942 0.961 0.953

Rotation(30°) 0.711 0.719 0.704 0.730 0.778 0.718

Rotation(45°) 0.699 0.689 0.693 0.697 0.762 0.691

Rotation(90°) 0.988 0.965 0.968 0.979 0.986 0.963

Rotation(180°) 0.970 0.975 0.984 0.973 0.984 0.986 Cropping(10%) 0.903 0.891 0.908 0.846 0.829 0.834 Cropping(20%) 0.839 0.827 0.837 0.768 0.752 0.760 Cropping(30%) 0.791 0.656 0.802 0.669 0.654 0.659 Cropping(40%) 0.746 0.567 0.751 0.578 0.573 0.572 Cropping(50%) 0.727 0.512 0.723 0.524 0.518 0.517

5 Conclusions

In this paper we present a novel watermark extraction algorithm based on DWT and GA. In order to insert the watermark strongly, the proposed algorithm transforms the image into wavelet domain and inserts the watermark into the lowest frequency band. The proposed GA-based watermarking algorithm can effectively extract the water-mark with various attacks such as translation and rotation. We also evaluated the performance of the proposed watermarking technique against the attack of image compression such as JPEG and JPEG2000. Further research on other types of geomet-rical attacks including scaling is in progress in order to ensure more robust watermark extraction.

716 D. Lee et al.

Acknowledgments

This research was supported by Korean Ministry of Information and Communication under the HNRC-ITRC program supervised by the IITA.

References

1. Kang, X., Huang, J., Shi, Y., Lin, Y.: A DWT-DFT Composite Watermarking Scheme Robust to Both Affine Transform and JPEG Compression. IEEE Trans., Vol. 13. Circuits, Systems for Video Technology (2003) 776-786

2. Pereira, S., Pun. T.: “Robust Template Matching for Affine Resistant Image Watermark-ing. IEEE Trans., Vol. 9. Image Processing (2000) 1123-1129

3. Reddy, A. A., Chatterji, B. N.: A New Wavelet-Based Logo-Watermarking Scheme. Pat-tern Recognition Letters, Vol. 26. Elsevier (2005) 1019-1027

4. Dawei, Z. L., Guanrong, C., Wenbo L.: A Chaos-Based Robust Wavelet-Domain Water-marking Algorithm. Chaos, Solitons & Fractals, Vol. 22 (2004) 47-54

5. Deguillaume, F., Voloshynovskiy, S., Pun, T.: A Method for the Estimation and Recover-ing from General Affine Transform in Digital Watermarking Applications. Security and Watermarking of Multimedia Contents IV, Vol. 4675. SPIE (2002) 313-322

6. Dugelay, J. L., Petitcolas, F. A. P.: Possible Counter-Attackers Against Random Geomet-ric Distortions. Security and Watermarking of Multimedia Contents II, Vol. 3971. SPIE (2002) 338-345

7. Braudaway, G. W., Minter, F.: Automatic Recovery of Invisible Image Watermarks from Geometrically Distorted Images. Security and watermarking of multimedia contents I, Vol. 3971. SPIE (2000) 74-81

8. Kang, X., Huang, J., Shi, Y.: An Image Watermarking Algorithm Robust to Geometric Distortion. Lecture Notes in Computer Science, Vol. 2613. Springer-Verlag, Digital Wa-termarking (2002) 212-223

9. Ali, F. F., Nakao, Z., Chen, Y. W.: An Evolutionary Approach for Graph Coloring. in IEEE, Vol. 5. SMC Conf. (1999) 524-532

10. Oh, W., Kim, H.: The Watermarking Evaluation and Certification Technique of Image. Telecommunications Technology Association, Vol. 90 (2003) 95-103