- Email: [email protected]
A new wavelet based logo-watermarking scheme
Pattern Recognition Letters 26 (2005) 1019–1027 www.elsevier.com/locate/patrec
A new wavelet based logo-watermarking scheme A. Adhipathi Reddy *, B.N. Chatterji Department of Electronics and Electrical Communication Engineering, Indian Institute of Technology, Kharagpur 721 302, West Bengal, India Received 2 September 2004 Available online 11 November 2004
Abstract A new wavelet based logo-watermarking scheme for copyright protection of digital image is presented. Instead of using a noise type Gaussian sequence, a visually meaningful gray scale logo is used as watermark. Watermark embedding process is carried out by transforming both the image and logo in wavelet domain. To embed the watermark robustly and imperceptibly, watermark bits are added to the significant coefficients of each subband selected by considering the human visual system (HVS) characteristics. A scheme is developed for reliable extraction of watermark from distorted images. From the experimental results it can be observed that proposed method is robust to wide variety of attacks. Comparison with the existing methods shows the superiority of the proposed method. 2004 Elsevier B.V. All rights reserved. Keywords: Watermarking; Logo; Wavelet transform; HVS characteristics
1. Introduction With the increasing use of Internet and effortless copying, tampering and distribution of digital data, copyright protection for multimedia data has become an important issue. Digital watermarking emerged as a tool for protecting the multimedia data from copyright infringement. In digital watermarking an imperceptible signal ‘‘mark’’ is embed* Corresponding author. Tel.: +91 3222 281476; fax: +91 3222 255303. E-mail addresses: [email protected] (A.A. Reddy), [email protected] (B.N. Chatterji).
ded into the host image, which uniquely identifies the ownership. After embedding the watermark, there should be no perceptual degradation. These watermarks should not be removable by unauthorized person and should be robust against intentional and unintentional attacks. Different watermarking techniques have already been published in the literature. Overviews on watermarking techniques can be found in (Langelaar et al., 2000). Watermarking techniques can be broadly classified into two categories: such as spatial domain methods and transform domain methods. Spatial domain methods are less complex as no transform
0167-8655/$ - see front matter 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2004.09.047
1020
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
is used, but are not robust against attacks. Transform domain watermarking techniques are more robust in comparison to spatial domain methods. This is due to the fact that when image is inverse wavelet transformed watermark is distributed irregularly over the image, making the attacker difficult to read or modify. Among the transform domain watermarking techniques discrete wavelet transform (DWT) based watermarking techniques are gaining more popularity because of superior modeling of HVS. A detail survey on wavelet based watermarking techniques can be found in (Meerwald and Uhl, 2001). Another advantage of the wavelet domainwatermarking algorithms is that security can be improved by selecting a key dependent wavelet transform as implemented in (Wang et al., 2002). They used randomly generated orthonormal filter bank as a major part of the private key. Besides selecting the filter bank randomly, to improve the private control over the watermark, middle frequency subbands are also selected based on the private key. Similarly, Dietl et al. (2003) proposed to use wavelet filter parameterization as a secret transform domain to improve the security of the watermarking method. To take the advantage of localization and multiresolution property of the wavelet transform Zhang et al. (2003) and Wang and Lin (2004) proposed wavelet tree based watermarking algorithms. Dawei et al. (2004) proposed a new type of technique in which wavelet transform applies locally, based on chaotic logistic map, and embeds the watermark. This method shows good robustness to geometric attacks like cropping and rotation but is sensitive to common signal processing attacks like lowpass filtering and image sharpening. As pointed out by Braudaway (1997) and Zeng and Lei (1999), by embedding visually meaningful marks like logo or seal, it can be easy for convincing non-technical arbitrators by showing the extracted logo or seal than presenting a numerical value detected using statistical watermark detection techniques. Another advantage of using logo as watermark is that HVS characteristics can be exploited in recognizing noisy visual mark since HVS filters out random noise for better recognition of meaningful pattern.
Many researchers have investigated watermarking methods by embedding binary logos. Ohnishi and Matsui (1996) showed a method for embedding seal in wavelet transform domain. In order to make the method robust to JPEG compression they quantized the wavelet coefficients depending on the quality factor before embedding watermark. A method of embedding logos in DCT domain was proposed by Hsu and Wu (1996). First they permuted the seal using pseudo-random number traversing method and added this to the middle frequency coefficient of the 8 · 8 DCT coefficient block. Later they (Hsu and Wu, 1998) proposed a method in which both the image and binary logo are hierarchically wavelet decomposed and detail bands of the logo are added to corresponding bands of the image. Kim et al. (1999) proposed a method for embedding binary logos in the Fourier domain. Before embedding the logo into the host data, they modulated the logo by adding pseudo-noise generated with a secret key. Zeng and Lei (1999) embedded binary logos by segmenting the image into small blocks and embedding one bit for each block. For improving the detector performance they exploited spatial correlation of the logo by embedding adjacent logo bits in adjacent image blocks and incorporating them in extraction procedure. All the methods discussed above refer to binary watermark embedding only. But in many practical applications logos are grayscale images and these methods cannot be directly used for embedding them. Some authors like Niu et al. (2000) have embedded gray scale logos by converting them into bit planes. But by converting the grayscale logos into bit planes, these methods are not exploiting the perceptual characteristics of the logos and the host data in embedding the watermark. As pointed out by Kundur and Hatzinakos (2004) use of grayscale logo as watermark facilitates the embedding of arbitrary commercial logos and increases the quality of and overall number of possible logos identifiable by human observers. Kundur and Hatzinakos (2004) proposed a multiresolution fusion based watermarking method for embedding grayscale logos into wavelet-transformed images. For watermarking,
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
the logo is 1-level wavelet decomposed. Each subband of host image is divided into blocks of size equal to the size of subband of the logo. Four subbands of the logo that corresponds to different orientation are added to the same orientation blocks of the image. In this paper, a novel robust wavelet based grayscale logo watermarking technique is presented. To embed the watermark robustly and imperceptibly, HVS characteristics are used in selecting the significant coefficients and adding the watermark to these coefficient. To recover the watermark from the distorted images, a method of reliable watermark extraction is presented, in which the watermark bits are extracted by taking into consideration the distortion caused by the attacks. To show the validity of the proposed method, the watermarked images are tested for different type of attacks and results are compared with the existing methods. The significant advantage of our proposed method over the method of Kundur and Hatzinakos (2004) is that the watermark can be added to each wavelet coefficient with maximum strength without any perceptual degradation. This is because the weight factors are calculated for individual coefficients instead of calculating for a block of wavelet coefficients as in the method of Kundur and Hatzinakos (2004). The rest of the paper is organized as follows. The proposed algorithms for watermark embedding and extraction are explained in Section 2. In Section 3, the experimental results are presented. Finally, the concluding remarks are given in Section 4.
2. Proposed watermarking method The proposed method embeds watermark by decomposing the host image and the watermark using wavelet transform. A visual mask based on HVS characteristics is used for calculating the weight factor for each wavelet coefficient of the host image and the significant coefficients from each subband are selected based on these weight factors. To embed the watermark in all selected wavelet coefficients, the watermark bits are repeat-
1021
edly added to the selected coefficients with their corresponding weight factors. These weight factors give the maximum amount of modification that can be applied to a wavelet coefficient without any perceptual degradation. The watermark used for embedding is a gray scale logo image, which is very small compared to the size of the host image. The watermark needs to be very small in order to make it spatially localized and to make robust against attacks like cropping and filtering. During the watermark recovery, the watermark repetitions are extracted and are combined by the weights calculated by considering the distortions in the surrounding pixels, the level of the subband and the weight factor of corresponding pixel. The method presented here is an improvement of the method proposed in (Adhipathi and Chatterji, 2003). The main difference lies in the watermark extraction algorithm. 2.1. HVS characteristics A number of factors effects the noise sensitivity of the human eye like luminance, frequency band, texture and proximity to an edge. Human eye is less sensitive to the areas of the image where brightness is high or low. As observed by Watson et al. (1997), human eye is less sensitive to noise in high frequency subbands and bands having orientation of ±45. Sensitivity of human eye to noise in textured area is less and it is more near the edges. Based on these observations, Lewis and Knowles (1992) developed a model for adaptively quantizing the wavelet coefficients for image compression. With some modification, Barni et al. (2001) developed a masking function for calculating the weight factors to embed the pseudorandom binary sequence into high frequency subbands of the host image. In our proposed method, we are using the model of Barni et al. (2001) for calculating the weight factors for wavelet coefficients of the host image. 2.2. Watermark embedding algorithm The wavelet representation of 4-level transformed image is shown in Fig. 1. Let us represent each subband with S hl ði; jÞ where h 2 {LL, LH,
1022
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027 S33 S30 S32 S 31
S12
2.3. Watermark extraction algorithm
S20
S10
S21
S00 S12
1 1
S
S01
S02
Fig. 1. 4-Level wavelet decomposed image representation.
HL, HH} represents the orientation and l 2 {0, 1, 2, 3} gives resolution level of subband of image of size I · J. Let x(m, n) represent the watermark of size M · N. The algorithm for embedding gray scale logo is formulated as follows: Step 1: Decompose the host image by L-levels and logo image by 1-level using DWT. Find the weight factors whl ði; jÞ for wavelet coefficients as given in (Barni et al., 2001). Step 2: Significant coefficients in each band are found out based on their weight factors. Weight factors of each subband are sorted in descending order, to find the threshold weight as given below T hl
¼ Sðp GS h Þ l
ð1Þ
where S( ) are the sorted weight factors of the subband, p is percentage of wavelet coefficients in which watermark is embedded and GS h is the size of the subband. The l coefficients, which have weight factors more than threshold value T hl , are considered as significant coefficients and are used for embedding the watermark. Step 3: Add watermark bits to significant coefficients of all subbands using Eq. (2). h b S l ði; jÞ ¼ S hl ði; jÞ þ awhl ði; jÞxðm; nÞ
ð2Þ
Here the constant a, gives the watermark strength. Step 4: After embedding watermark bits, L-level inverse wavelet transform of the image is found out to get the watermarked image.
For watermark recovery from watermarked image, both the original and the watermark images are needed. Although assuming accessibility to original image may not be possible in some practical applications, we consider the applications where robustness is important and have access to the original image. The steps for watermark extraction are as follows: Step 1: Both original and watermark images are L-level wavelet decomposed. Weight factors are found out by considering the original image as given in (Barni et al., 2001). Step 2: Each repetition of watermark bits are extracted from watermark image as given below: x0 ðm; nÞ ¼
h b S l ði; jÞ S hl ði; jÞ whl ði; jÞ
ð3Þ
h
where b S l ði; jÞ is the suspicious image subband. Step 3: Corresponding extracted watermark bits are combined by multiplying with distortion weight as given below: ^xðm; nÞ ¼ ^xðm; nÞ þ x0 ðm; nÞ 2 32 l h 6wl ði; jÞ 2 7 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5 4 q Dhl ði; jÞ
ð4Þ
where whl is the weight factor of the corresponding pixel and Dhl ði; jÞ is the distortion calculated in the neighboring Nx · Ny window as follows: Dhl ði; jÞ PiþN2x PjþN2y h ¼
x¼iN2x
N y¼j 2y
h b S l ðx; yÞ S hl ðx; yÞ
i2
Nx Ny ð5Þ
Weights whl calculated for lower frequency subband coefficients are very small in comparison to higher frequency subband coefficients. During combining of watermark repetitions these weights
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
gives more weightage to extracted coefficients of high frequency subbands. Therefore, in order to give equal weights to the extracted coefficients of all subbands, the multiplication factor 2l is used. Step 4: The corresponding distortion weights are summed up as follows sumðm; nÞ ¼ sumðm; nÞ 2
32 l
h 6wl ði; jÞ 2 7 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5 þ4 q Dhl ði; jÞ
ð6Þ
Step 5: After extracting all the watermark bits and by combining using Eq. (4), they are normalized as given below ^xðm; nÞ ¼
^xðm; nÞ sumðm; nÞ
ð7Þ
Step 6: To form the extracted logo, take 1-level inverse wavelet transform of ^xðm; nÞ and scale it between 0 and 255. To find out whether logo is present or not, the extracted logo is visually verified and compared with the original logo.
3. Experimental results The performance of the proposed algorithm is tested on various types of images. Here the results are presented for grayscale 8-bit Lena image of size 512 · 512. The logo used for watermarking is an 8bit grayscale image of size 64 · 64. Original Lena
1023
and logo images are shown is Fig. 2(a) and (b) respectively. For 4-level wavelet decomposition Daubechies 9/7 filter coefficients are used. Same filter coefficients are used for 1-level decomposition of logo. For embedding the watermark in approximate band a is taken as 0.4 and for all other bands a = 1.0 is used. Watermark is added to 75% coefficients (p = 0.75) of each subband having large weight factors. Watermarked Lena image having PSNR value of 44.7 is shown in Fig. 2(c). If the original and the watermarked Lena images are observed we cannot find any perceptual degradation. Extracted logo from the watermarked image is shown in Fig. 2(d). Eight times magnified absolute difference between the original and the watermarked Lena image is shown in Fig. 3. It can be observed from Fig. 3 that the watermark bits are stored near edges and high textured regions where HVS is less sensitive. Robustness of the proposed method is evaluated for various types of image distortions as discussed below. 3.1. Non-geometric attacks Watermarked Lena image is tested for non-geometric attacks such as averaging, median filtering and additive noise. Logos extracted after applying 11 · 11 average and median filtering are shown in Fig. 4(a) and (b). After applying these filters, images are very much degraded and lot of data is lost but extracted logos are still recognizable. To test the robustness against adding noise, the watermarked image is degraded by adding salt and pepper noise randomly. The logo extracted from 50%
Fig. 2. (a) Original Lena image; (b) original logo image; (c) watermarked Lena image; (d) extracted logo.
1024
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
tracted logo from 0.1 bits/pixel (bpp) SPIHT compressed image is shown in Fig. 4(d). The watermarked Lena image shown in Fig. 5(a) is 64 times compressed. At 64 times compression the image in Fig. 5(a) is showing blocking artifacts, but extracted logo shown in Fig. 5(b) is recognizable. 3.2. Geometric attacks
Fig. 3. Absolute difference of original and watermarked Lena image magnified by 8 times.
noise added watermarked image is shown in Fig. 4(c). The extracted logo is noisy, but it is recognizable. The watermarked Lena image is also tested for SPIHT and JPEG compression attacks. The ex-
Watermarked Lena image is tested for geometric attacks such as cropping, resizing and rotation attacks. Robustness against cropping attack decreases as the size of the watermark increases. This is because with increasing size of the watermark it is less spatially localized. Our proposed method is showing better robustness to cropping attack. This can be observed from Fig. 5(c). Extracted logo shown in Fig. 5(d) is still recognizable even when the remaining area is 1.37% of the total area. For watermark extraction from resized image, it is again brought to the original size. Watermark extracted from 4 times down sized image is shown
Fig. 4. Watermark extracted after (a) 11 · 11 average filtering; (b) 11 · 11 median filtering; (c) adding 50% noise; (d) 0.1 bpp SPIHT compression.
Fig. 5. (a) 64 times compressed watermarked Lena image; (b) extracted logo; (c) 1.37% remained watermarked Lena image after cropping; (d) extracted logo.
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
1025
Fig. 6. Extracted logos after (a) resizing image by a factor of 4; (b) 0.5 of rotation; (c) histogram equalization.
in Fig. 6(a). Even though the wavelet transform is not rotational invariant, our proposed method can extract the watermark for small rotations. Extracted logo from 0.5 rotated watermarked Lena image is shown in Fig. 6(b). Watermarked Lena image is also tested for histogram equalization and warping attacks. Histogram equalization has less effect on the proposed method. This can be observed from Fig. 6(c). Fig. 7(a) and (b) shows warped watermarked Lena image and corresponding extracted logo. It can be easy to conclude that distorting any part of image has no effect in extracting the logo. 3.3. Comparison with existing methods
and a randomly generated binary watermark of size 32 · 32. For embedding the watermark with our method, a = 0.1 for approximate band and a = 0.5 for all other bands is taken. To verify the presence of watermark, the normalized correlation q between original and extracted watermark is calculated using Eq. (8) and it is compared with the appropriate threshold T. For the watermark length of 1024, the threshold T is chosen to be 0.2 for a false positive probability of Pfp = 1010 (Kundur and Hatzinakos, 2004). M1 m 1 P NP
q¼
m¼0
For comparison of the proposed method with the existing methods, we have implemented the fusion based watermarking method of Kundur and Hatzinakos (2004). Comparison is carried out by using gray scale Lena image of size 512 · 512
Fig. 7. (a) Warped watermark Lena image: (b) extracted logo.
xðm; nÞ^xðm; nÞ
m¼0 n¼0 M1 m 1 P NP
ð8Þ x2 ðm; nÞ
n¼0
Graphs in Fig. 8(a) and (b) show the detector responses for average and median filtering attacks, respectively. Solid line corresponds to our proposed method and dashed line corresponds to the fusion-based method. With averaging attack in the proposed method, the watermark is detected upto a filter size of 9 · 9 and with fusion based method upto a filter size of 7 · 7. For median filtering attack, with our method, the watermark is detected upto a filter size of 11 · 11 and with fusion based method upto a filter size of 8 · 8. Fig. 9 shows the detector responses with JPEG and SPIHT compression attacks. With our method the watermark is detected upto a compression ratio of 60 and with the fusion based method upto a compression ratio of 40 as shown in Fig. 9(a). As shown in Fig. 9(b), for SPIHT compression, with both the methods watermark is detected upto a bit rate of 0.1. But performance of our method is better than the fusion based method.
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027 1
1
0.8
0.8
Detector Response
Detector Response
1026
0.6
0.4 0.2
0.6
0.4 0.2
0
0 3
(a)
5
7
9
11
Filter Size
3
(b)
5
7
9
11
Filter Size
1
1
0.8
0.8
Detector Response
Detector Response
Fig. 8. Detector responses after (a) average filtering attack; (b) median filter attack (solid line corresponds to the proposed method and dashed line corresponds to the fusion-based method).
0.6 0.4 0.2 0 10
20
(a)
30
40
50
0.6 0.4 0.2 0 0.1
60
Compression Ratio
0.3
0.5
0.7
0.9
Bits Rate (bpp)
(b)
1
1
0.8
0.8
Detector Response
Detector Response
Fig. 9. Detector responses after (a) JPEG compression attack; (b) SPIHT compression attack (solid line corresponds to our proposed method and dashed line corresponds to the fusion-based method).
0.6
0.4 0.2 0
0.4 0.2 0
10
(a)
0.6
20
30
40
Percent of Noise
0
50
(b)
20
40
60
80
100
Percent Remaining Image Area
Fig. 10. Detector responses after (a) adding noise attack; (b) cropping attack (solid line corresponds to our proposed method and dashed line corresponds to the fusion-based method).
Results in Fig. 10(a) show the detector responses with different percentage of salt and pepper noise attack. Watermark is detected upto 40% noise with our method and upto 25% noise with fusion based method. With cropping attack both the methods are showing similar robustness as shown in Fig. 10(b). In both the cases watermark is detected upto a remaining area of 2.5%.
There are many more attacks that can be tested for, but the results presented here give a good indication of the capabilities of the proposed method. One advantage of our proposed method is that as the security of the method lies in the original image, attacker cannot extract the data unless the attacker has access to the original image. However, if attacker tries to remove the watermark,
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
quality of the image degrades proportional to the extracted logo. This means that the watermark cannot be removed without degrading the image quality substantially. 4. Conclusion A wavelet-based grayscale logo watermarking technique for digital images is presented. Watermark is embedded into the coefficients selected based on the weight factors calculated by exploiting the HVS characteristics. A method for reliable watermark extraction based on the distortion in the surrounding pixels is presented. Robustness of the algorithm is tested against different type of attacks. Finally, to show the superiority of the proposed method we have compared our results with the results of Kundur and Hatzinakos (2004). From the results it can be observed that our proposed method is showing better robustness to attacks than the method of Kundur and Hatzinakos (2004).
References Adhipathi, R.A., Chatterji, B.N., 2003. Wavelet based robust logo watermarking scheme. In: Internat. Conf. on Advances in Pattern Recognition, Kolkata, India, December. Barni, M., Bartolini, F., Piva, A., 2001. Improved waveletbased watermarking through pixel-wise masking. IEEE Trans. Image Process. 10 (May), 783–791. Braudaway, G.W., 1997. Protecting publicly-available images with an invisible image watermark. In: Internat. Conf. on Image Process., vol. 1, October, pp. 524–527. Dawei, Z., Guanrong, C., Wenbo, L., 2004. A chaos-based robust wavelet-domain watermarking algorithm. J. Chaos Solitons Fractals (October), 47–54. Dietl, W., Meerwald, P., Uhl, A., 2003. Protection of waveletbased watermarking systems using filter parametrization. Signal Process. (October), 2095–2116.
1027
Hsu, C.T., Wu, J.-L., 1996. Hidden signatures in images. In Int. Conf. on Image Process., vol. 3, Lausanne, Switzerland, September, pp. 223–226. Hsu, C.T., Wu, J.-L., 1998. Multiresolution watermarking for digital images. IEEE Trans. Circuits Syst. 45 (August), 1097–1101. Kim, W.G., Lee, J.C., Lee, W.D., 1999. An image watermarking scheme with hidden signatures. In Int. Conf. on Image Process., Japan 2 (October), pp. 206–210. Kundur, D., Hatzinakos, D., 2004. Towards Robust logo watermarking using multiresolution image fusion. IEEE Trans. Multimedia 6 (February), 185–197. Langelaar, G.C., Setyawan, I., Lagendijk, R.I., 2000. Watermarking digital image and video data. IEEE Signal Process. Magazine (September), 20–46. Lewis, A.S., Knowles, G., 1992. Image compression using the 2D wavelet transform. IEEE Trans. Image Process. 1 (April), 244–250. Meerwald, P., Uhl, A., 2001. A survey of wavelet-domain watermarking algorithms. In Proc. of SPIE, Electronic Imaging, Security and Watermarking of Multimedia Contents III, CA, USA 4314 (January), pp. 505–516. Niu, X., Sun, S., Xiang, W., 2000. Multiresolution watermarking for video based on gray-level digital watermark. IEEE Trans. Consumber Electron. 46 (May), 375–384. Ohnishi, J., Matsui, K., 1996. Embedding a seal into a picture under orthogonal wavelet transform. In Proc. of IEEE Internat. Conf. on Multimedia and Computing System, Hiroshima, Japan pp. 514–521. Wang, S.H., Lin, Y.P., 2004. Wavelet tree quantization for copyright protection watermarking. IEEE Trans. Image Process. 13 (February), 154–165. Wang, Y., Doherty, J.F., Van Dyck, R.E., 2002. A waveletbased watermarking algorithm for ownership verification of digital images. IEEE Trans. Image Process. 11 (February), 77–88. Watson, A.B., Yang, G.Y., Solomon, J.A., Villasenor, J., 1997. Visibility of wavelet quantization noise. IEEE Trans. Image Process. 6 (August), 1164–1175. Zeng, B.L.W., Lei, S., 1999. Extraction of multiresolution watermark images for resolving rightful ownership. In Proc. of SPIE, Security and Watermarking of Multimedia Contents 3657 (January), pp. 404–414. Zhang, X.D., Feng, J., Lo, K.T., 2003. Image watermarking using tree-based spatial-frequency feature of wavelet transform. J. Visual Comm. Image Representation (December), 474–491.
A new wavelet based logo-watermarking scheme A. Adhipathi Reddy *, B.N. Chatterji Department of Electronics and Electrical Communication Engineering, Indian Institute of Technology, Kharagpur 721 302, West Bengal, India Received 2 September 2004 Available online 11 November 2004
Abstract A new wavelet based logo-watermarking scheme for copyright protection of digital image is presented. Instead of using a noise type Gaussian sequence, a visually meaningful gray scale logo is used as watermark. Watermark embedding process is carried out by transforming both the image and logo in wavelet domain. To embed the watermark robustly and imperceptibly, watermark bits are added to the significant coefficients of each subband selected by considering the human visual system (HVS) characteristics. A scheme is developed for reliable extraction of watermark from distorted images. From the experimental results it can be observed that proposed method is robust to wide variety of attacks. Comparison with the existing methods shows the superiority of the proposed method. 2004 Elsevier B.V. All rights reserved. Keywords: Watermarking; Logo; Wavelet transform; HVS characteristics
1. Introduction With the increasing use of Internet and effortless copying, tampering and distribution of digital data, copyright protection for multimedia data has become an important issue. Digital watermarking emerged as a tool for protecting the multimedia data from copyright infringement. In digital watermarking an imperceptible signal ‘‘mark’’ is embed* Corresponding author. Tel.: +91 3222 281476; fax: +91 3222 255303. E-mail addresses: [email protected] (A.A. Reddy), [email protected] (B.N. Chatterji).
ded into the host image, which uniquely identifies the ownership. After embedding the watermark, there should be no perceptual degradation. These watermarks should not be removable by unauthorized person and should be robust against intentional and unintentional attacks. Different watermarking techniques have already been published in the literature. Overviews on watermarking techniques can be found in (Langelaar et al., 2000). Watermarking techniques can be broadly classified into two categories: such as spatial domain methods and transform domain methods. Spatial domain methods are less complex as no transform
0167-8655/$ - see front matter 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2004.09.047
1020
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
is used, but are not robust against attacks. Transform domain watermarking techniques are more robust in comparison to spatial domain methods. This is due to the fact that when image is inverse wavelet transformed watermark is distributed irregularly over the image, making the attacker difficult to read or modify. Among the transform domain watermarking techniques discrete wavelet transform (DWT) based watermarking techniques are gaining more popularity because of superior modeling of HVS. A detail survey on wavelet based watermarking techniques can be found in (Meerwald and Uhl, 2001). Another advantage of the wavelet domainwatermarking algorithms is that security can be improved by selecting a key dependent wavelet transform as implemented in (Wang et al., 2002). They used randomly generated orthonormal filter bank as a major part of the private key. Besides selecting the filter bank randomly, to improve the private control over the watermark, middle frequency subbands are also selected based on the private key. Similarly, Dietl et al. (2003) proposed to use wavelet filter parameterization as a secret transform domain to improve the security of the watermarking method. To take the advantage of localization and multiresolution property of the wavelet transform Zhang et al. (2003) and Wang and Lin (2004) proposed wavelet tree based watermarking algorithms. Dawei et al. (2004) proposed a new type of technique in which wavelet transform applies locally, based on chaotic logistic map, and embeds the watermark. This method shows good robustness to geometric attacks like cropping and rotation but is sensitive to common signal processing attacks like lowpass filtering and image sharpening. As pointed out by Braudaway (1997) and Zeng and Lei (1999), by embedding visually meaningful marks like logo or seal, it can be easy for convincing non-technical arbitrators by showing the extracted logo or seal than presenting a numerical value detected using statistical watermark detection techniques. Another advantage of using logo as watermark is that HVS characteristics can be exploited in recognizing noisy visual mark since HVS filters out random noise for better recognition of meaningful pattern.
Many researchers have investigated watermarking methods by embedding binary logos. Ohnishi and Matsui (1996) showed a method for embedding seal in wavelet transform domain. In order to make the method robust to JPEG compression they quantized the wavelet coefficients depending on the quality factor before embedding watermark. A method of embedding logos in DCT domain was proposed by Hsu and Wu (1996). First they permuted the seal using pseudo-random number traversing method and added this to the middle frequency coefficient of the 8 · 8 DCT coefficient block. Later they (Hsu and Wu, 1998) proposed a method in which both the image and binary logo are hierarchically wavelet decomposed and detail bands of the logo are added to corresponding bands of the image. Kim et al. (1999) proposed a method for embedding binary logos in the Fourier domain. Before embedding the logo into the host data, they modulated the logo by adding pseudo-noise generated with a secret key. Zeng and Lei (1999) embedded binary logos by segmenting the image into small blocks and embedding one bit for each block. For improving the detector performance they exploited spatial correlation of the logo by embedding adjacent logo bits in adjacent image blocks and incorporating them in extraction procedure. All the methods discussed above refer to binary watermark embedding only. But in many practical applications logos are grayscale images and these methods cannot be directly used for embedding them. Some authors like Niu et al. (2000) have embedded gray scale logos by converting them into bit planes. But by converting the grayscale logos into bit planes, these methods are not exploiting the perceptual characteristics of the logos and the host data in embedding the watermark. As pointed out by Kundur and Hatzinakos (2004) use of grayscale logo as watermark facilitates the embedding of arbitrary commercial logos and increases the quality of and overall number of possible logos identifiable by human observers. Kundur and Hatzinakos (2004) proposed a multiresolution fusion based watermarking method for embedding grayscale logos into wavelet-transformed images. For watermarking,
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
the logo is 1-level wavelet decomposed. Each subband of host image is divided into blocks of size equal to the size of subband of the logo. Four subbands of the logo that corresponds to different orientation are added to the same orientation blocks of the image. In this paper, a novel robust wavelet based grayscale logo watermarking technique is presented. To embed the watermark robustly and imperceptibly, HVS characteristics are used in selecting the significant coefficients and adding the watermark to these coefficient. To recover the watermark from the distorted images, a method of reliable watermark extraction is presented, in which the watermark bits are extracted by taking into consideration the distortion caused by the attacks. To show the validity of the proposed method, the watermarked images are tested for different type of attacks and results are compared with the existing methods. The significant advantage of our proposed method over the method of Kundur and Hatzinakos (2004) is that the watermark can be added to each wavelet coefficient with maximum strength without any perceptual degradation. This is because the weight factors are calculated for individual coefficients instead of calculating for a block of wavelet coefficients as in the method of Kundur and Hatzinakos (2004). The rest of the paper is organized as follows. The proposed algorithms for watermark embedding and extraction are explained in Section 2. In Section 3, the experimental results are presented. Finally, the concluding remarks are given in Section 4.
2. Proposed watermarking method The proposed method embeds watermark by decomposing the host image and the watermark using wavelet transform. A visual mask based on HVS characteristics is used for calculating the weight factor for each wavelet coefficient of the host image and the significant coefficients from each subband are selected based on these weight factors. To embed the watermark in all selected wavelet coefficients, the watermark bits are repeat-
1021
edly added to the selected coefficients with their corresponding weight factors. These weight factors give the maximum amount of modification that can be applied to a wavelet coefficient without any perceptual degradation. The watermark used for embedding is a gray scale logo image, which is very small compared to the size of the host image. The watermark needs to be very small in order to make it spatially localized and to make robust against attacks like cropping and filtering. During the watermark recovery, the watermark repetitions are extracted and are combined by the weights calculated by considering the distortions in the surrounding pixels, the level of the subband and the weight factor of corresponding pixel. The method presented here is an improvement of the method proposed in (Adhipathi and Chatterji, 2003). The main difference lies in the watermark extraction algorithm. 2.1. HVS characteristics A number of factors effects the noise sensitivity of the human eye like luminance, frequency band, texture and proximity to an edge. Human eye is less sensitive to the areas of the image where brightness is high or low. As observed by Watson et al. (1997), human eye is less sensitive to noise in high frequency subbands and bands having orientation of ±45. Sensitivity of human eye to noise in textured area is less and it is more near the edges. Based on these observations, Lewis and Knowles (1992) developed a model for adaptively quantizing the wavelet coefficients for image compression. With some modification, Barni et al. (2001) developed a masking function for calculating the weight factors to embed the pseudorandom binary sequence into high frequency subbands of the host image. In our proposed method, we are using the model of Barni et al. (2001) for calculating the weight factors for wavelet coefficients of the host image. 2.2. Watermark embedding algorithm The wavelet representation of 4-level transformed image is shown in Fig. 1. Let us represent each subband with S hl ði; jÞ where h 2 {LL, LH,
1022
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027 S33 S30 S32 S 31
S12
2.3. Watermark extraction algorithm
S20
S10
S21
S00 S12
1 1
S
S01
S02
Fig. 1. 4-Level wavelet decomposed image representation.
HL, HH} represents the orientation and l 2 {0, 1, 2, 3} gives resolution level of subband of image of size I · J. Let x(m, n) represent the watermark of size M · N. The algorithm for embedding gray scale logo is formulated as follows: Step 1: Decompose the host image by L-levels and logo image by 1-level using DWT. Find the weight factors whl ði; jÞ for wavelet coefficients as given in (Barni et al., 2001). Step 2: Significant coefficients in each band are found out based on their weight factors. Weight factors of each subband are sorted in descending order, to find the threshold weight as given below T hl
¼ Sðp GS h Þ l
ð1Þ
where S( ) are the sorted weight factors of the subband, p is percentage of wavelet coefficients in which watermark is embedded and GS h is the size of the subband. The l coefficients, which have weight factors more than threshold value T hl , are considered as significant coefficients and are used for embedding the watermark. Step 3: Add watermark bits to significant coefficients of all subbands using Eq. (2). h b S l ði; jÞ ¼ S hl ði; jÞ þ awhl ði; jÞxðm; nÞ
ð2Þ
Here the constant a, gives the watermark strength. Step 4: After embedding watermark bits, L-level inverse wavelet transform of the image is found out to get the watermarked image.
For watermark recovery from watermarked image, both the original and the watermark images are needed. Although assuming accessibility to original image may not be possible in some practical applications, we consider the applications where robustness is important and have access to the original image. The steps for watermark extraction are as follows: Step 1: Both original and watermark images are L-level wavelet decomposed. Weight factors are found out by considering the original image as given in (Barni et al., 2001). Step 2: Each repetition of watermark bits are extracted from watermark image as given below: x0 ðm; nÞ ¼
h b S l ði; jÞ S hl ði; jÞ whl ði; jÞ
ð3Þ
h
where b S l ði; jÞ is the suspicious image subband. Step 3: Corresponding extracted watermark bits are combined by multiplying with distortion weight as given below: ^xðm; nÞ ¼ ^xðm; nÞ þ x0 ðm; nÞ 2 32 l h 6wl ði; jÞ 2 7 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5 4 q Dhl ði; jÞ
ð4Þ
where whl is the weight factor of the corresponding pixel and Dhl ði; jÞ is the distortion calculated in the neighboring Nx · Ny window as follows: Dhl ði; jÞ PiþN2x PjþN2y h ¼
x¼iN2x
N y¼j 2y
h b S l ðx; yÞ S hl ðx; yÞ
i2
Nx Ny ð5Þ
Weights whl calculated for lower frequency subband coefficients are very small in comparison to higher frequency subband coefficients. During combining of watermark repetitions these weights
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
gives more weightage to extracted coefficients of high frequency subbands. Therefore, in order to give equal weights to the extracted coefficients of all subbands, the multiplication factor 2l is used. Step 4: The corresponding distortion weights are summed up as follows sumðm; nÞ ¼ sumðm; nÞ 2
32 l
h 6wl ði; jÞ 2 7 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5 þ4 q Dhl ði; jÞ
ð6Þ
Step 5: After extracting all the watermark bits and by combining using Eq. (4), they are normalized as given below ^xðm; nÞ ¼
^xðm; nÞ sumðm; nÞ
ð7Þ
Step 6: To form the extracted logo, take 1-level inverse wavelet transform of ^xðm; nÞ and scale it between 0 and 255. To find out whether logo is present or not, the extracted logo is visually verified and compared with the original logo.
3. Experimental results The performance of the proposed algorithm is tested on various types of images. Here the results are presented for grayscale 8-bit Lena image of size 512 · 512. The logo used for watermarking is an 8bit grayscale image of size 64 · 64. Original Lena
1023
and logo images are shown is Fig. 2(a) and (b) respectively. For 4-level wavelet decomposition Daubechies 9/7 filter coefficients are used. Same filter coefficients are used for 1-level decomposition of logo. For embedding the watermark in approximate band a is taken as 0.4 and for all other bands a = 1.0 is used. Watermark is added to 75% coefficients (p = 0.75) of each subband having large weight factors. Watermarked Lena image having PSNR value of 44.7 is shown in Fig. 2(c). If the original and the watermarked Lena images are observed we cannot find any perceptual degradation. Extracted logo from the watermarked image is shown in Fig. 2(d). Eight times magnified absolute difference between the original and the watermarked Lena image is shown in Fig. 3. It can be observed from Fig. 3 that the watermark bits are stored near edges and high textured regions where HVS is less sensitive. Robustness of the proposed method is evaluated for various types of image distortions as discussed below. 3.1. Non-geometric attacks Watermarked Lena image is tested for non-geometric attacks such as averaging, median filtering and additive noise. Logos extracted after applying 11 · 11 average and median filtering are shown in Fig. 4(a) and (b). After applying these filters, images are very much degraded and lot of data is lost but extracted logos are still recognizable. To test the robustness against adding noise, the watermarked image is degraded by adding salt and pepper noise randomly. The logo extracted from 50%
Fig. 2. (a) Original Lena image; (b) original logo image; (c) watermarked Lena image; (d) extracted logo.
1024
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
tracted logo from 0.1 bits/pixel (bpp) SPIHT compressed image is shown in Fig. 4(d). The watermarked Lena image shown in Fig. 5(a) is 64 times compressed. At 64 times compression the image in Fig. 5(a) is showing blocking artifacts, but extracted logo shown in Fig. 5(b) is recognizable. 3.2. Geometric attacks
Fig. 3. Absolute difference of original and watermarked Lena image magnified by 8 times.
noise added watermarked image is shown in Fig. 4(c). The extracted logo is noisy, but it is recognizable. The watermarked Lena image is also tested for SPIHT and JPEG compression attacks. The ex-
Watermarked Lena image is tested for geometric attacks such as cropping, resizing and rotation attacks. Robustness against cropping attack decreases as the size of the watermark increases. This is because with increasing size of the watermark it is less spatially localized. Our proposed method is showing better robustness to cropping attack. This can be observed from Fig. 5(c). Extracted logo shown in Fig. 5(d) is still recognizable even when the remaining area is 1.37% of the total area. For watermark extraction from resized image, it is again brought to the original size. Watermark extracted from 4 times down sized image is shown
Fig. 4. Watermark extracted after (a) 11 · 11 average filtering; (b) 11 · 11 median filtering; (c) adding 50% noise; (d) 0.1 bpp SPIHT compression.
Fig. 5. (a) 64 times compressed watermarked Lena image; (b) extracted logo; (c) 1.37% remained watermarked Lena image after cropping; (d) extracted logo.
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
1025
Fig. 6. Extracted logos after (a) resizing image by a factor of 4; (b) 0.5 of rotation; (c) histogram equalization.
in Fig. 6(a). Even though the wavelet transform is not rotational invariant, our proposed method can extract the watermark for small rotations. Extracted logo from 0.5 rotated watermarked Lena image is shown in Fig. 6(b). Watermarked Lena image is also tested for histogram equalization and warping attacks. Histogram equalization has less effect on the proposed method. This can be observed from Fig. 6(c). Fig. 7(a) and (b) shows warped watermarked Lena image and corresponding extracted logo. It can be easy to conclude that distorting any part of image has no effect in extracting the logo. 3.3. Comparison with existing methods
and a randomly generated binary watermark of size 32 · 32. For embedding the watermark with our method, a = 0.1 for approximate band and a = 0.5 for all other bands is taken. To verify the presence of watermark, the normalized correlation q between original and extracted watermark is calculated using Eq. (8) and it is compared with the appropriate threshold T. For the watermark length of 1024, the threshold T is chosen to be 0.2 for a false positive probability of Pfp = 1010 (Kundur and Hatzinakos, 2004). M1 m 1 P NP
q¼
m¼0
For comparison of the proposed method with the existing methods, we have implemented the fusion based watermarking method of Kundur and Hatzinakos (2004). Comparison is carried out by using gray scale Lena image of size 512 · 512
Fig. 7. (a) Warped watermark Lena image: (b) extracted logo.
xðm; nÞ^xðm; nÞ
m¼0 n¼0 M1 m 1 P NP
ð8Þ x2 ðm; nÞ
n¼0
Graphs in Fig. 8(a) and (b) show the detector responses for average and median filtering attacks, respectively. Solid line corresponds to our proposed method and dashed line corresponds to the fusion-based method. With averaging attack in the proposed method, the watermark is detected upto a filter size of 9 · 9 and with fusion based method upto a filter size of 7 · 7. For median filtering attack, with our method, the watermark is detected upto a filter size of 11 · 11 and with fusion based method upto a filter size of 8 · 8. Fig. 9 shows the detector responses with JPEG and SPIHT compression attacks. With our method the watermark is detected upto a compression ratio of 60 and with the fusion based method upto a compression ratio of 40 as shown in Fig. 9(a). As shown in Fig. 9(b), for SPIHT compression, with both the methods watermark is detected upto a bit rate of 0.1. But performance of our method is better than the fusion based method.
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027 1
1
0.8
0.8
Detector Response
Detector Response
1026
0.6
0.4 0.2
0.6
0.4 0.2
0
0 3
(a)
5
7
9
11
Filter Size
3
(b)
5
7
9
11
Filter Size
1
1
0.8
0.8
Detector Response
Detector Response
Fig. 8. Detector responses after (a) average filtering attack; (b) median filter attack (solid line corresponds to the proposed method and dashed line corresponds to the fusion-based method).
0.6 0.4 0.2 0 10
20
(a)
30
40
50
0.6 0.4 0.2 0 0.1
60
Compression Ratio
0.3
0.5
0.7
0.9
Bits Rate (bpp)
(b)
1
1
0.8
0.8
Detector Response
Detector Response
Fig. 9. Detector responses after (a) JPEG compression attack; (b) SPIHT compression attack (solid line corresponds to our proposed method and dashed line corresponds to the fusion-based method).
0.6
0.4 0.2 0
0.4 0.2 0
10
(a)
0.6
20
30
40
Percent of Noise
0
50
(b)
20
40
60
80
100
Percent Remaining Image Area
Fig. 10. Detector responses after (a) adding noise attack; (b) cropping attack (solid line corresponds to our proposed method and dashed line corresponds to the fusion-based method).
Results in Fig. 10(a) show the detector responses with different percentage of salt and pepper noise attack. Watermark is detected upto 40% noise with our method and upto 25% noise with fusion based method. With cropping attack both the methods are showing similar robustness as shown in Fig. 10(b). In both the cases watermark is detected upto a remaining area of 2.5%.
There are many more attacks that can be tested for, but the results presented here give a good indication of the capabilities of the proposed method. One advantage of our proposed method is that as the security of the method lies in the original image, attacker cannot extract the data unless the attacker has access to the original image. However, if attacker tries to remove the watermark,
A.A. Reddy, B.N. Chatterji / Pattern Recognition Letters 26 (2005) 1019–1027
quality of the image degrades proportional to the extracted logo. This means that the watermark cannot be removed without degrading the image quality substantially. 4. Conclusion A wavelet-based grayscale logo watermarking technique for digital images is presented. Watermark is embedded into the coefficients selected based on the weight factors calculated by exploiting the HVS characteristics. A method for reliable watermark extraction based on the distortion in the surrounding pixels is presented. Robustness of the algorithm is tested against different type of attacks. Finally, to show the superiority of the proposed method we have compared our results with the results of Kundur and Hatzinakos (2004). From the results it can be observed that our proposed method is showing better robustness to attacks than the method of Kundur and Hatzinakos (2004).
References Adhipathi, R.A., Chatterji, B.N., 2003. Wavelet based robust logo watermarking scheme. In: Internat. Conf. on Advances in Pattern Recognition, Kolkata, India, December. Barni, M., Bartolini, F., Piva, A., 2001. Improved waveletbased watermarking through pixel-wise masking. IEEE Trans. Image Process. 10 (May), 783–791. Braudaway, G.W., 1997. Protecting publicly-available images with an invisible image watermark. In: Internat. Conf. on Image Process., vol. 1, October, pp. 524–527. Dawei, Z., Guanrong, C., Wenbo, L., 2004. A chaos-based robust wavelet-domain watermarking algorithm. J. Chaos Solitons Fractals (October), 47–54. Dietl, W., Meerwald, P., Uhl, A., 2003. Protection of waveletbased watermarking systems using filter parametrization. Signal Process. (October), 2095–2116.
1027
Hsu, C.T., Wu, J.-L., 1996. Hidden signatures in images. In Int. Conf. on Image Process., vol. 3, Lausanne, Switzerland, September, pp. 223–226. Hsu, C.T., Wu, J.-L., 1998. Multiresolution watermarking for digital images. IEEE Trans. Circuits Syst. 45 (August), 1097–1101. Kim, W.G., Lee, J.C., Lee, W.D., 1999. An image watermarking scheme with hidden signatures. In Int. Conf. on Image Process., Japan 2 (October), pp. 206–210. Kundur, D., Hatzinakos, D., 2004. Towards Robust logo watermarking using multiresolution image fusion. IEEE Trans. Multimedia 6 (February), 185–197. Langelaar, G.C., Setyawan, I., Lagendijk, R.I., 2000. Watermarking digital image and video data. IEEE Signal Process. Magazine (September), 20–46. Lewis, A.S., Knowles, G., 1992. Image compression using the 2D wavelet transform. IEEE Trans. Image Process. 1 (April), 244–250. Meerwald, P., Uhl, A., 2001. A survey of wavelet-domain watermarking algorithms. In Proc. of SPIE, Electronic Imaging, Security and Watermarking of Multimedia Contents III, CA, USA 4314 (January), pp. 505–516. Niu, X., Sun, S., Xiang, W., 2000. Multiresolution watermarking for video based on gray-level digital watermark. IEEE Trans. Consumber Electron. 46 (May), 375–384. Ohnishi, J., Matsui, K., 1996. Embedding a seal into a picture under orthogonal wavelet transform. In Proc. of IEEE Internat. Conf. on Multimedia and Computing System, Hiroshima, Japan pp. 514–521. Wang, S.H., Lin, Y.P., 2004. Wavelet tree quantization for copyright protection watermarking. IEEE Trans. Image Process. 13 (February), 154–165. Wang, Y., Doherty, J.F., Van Dyck, R.E., 2002. A waveletbased watermarking algorithm for ownership verification of digital images. IEEE Trans. Image Process. 11 (February), 77–88. Watson, A.B., Yang, G.Y., Solomon, J.A., Villasenor, J., 1997. Visibility of wavelet quantization noise. IEEE Trans. Image Process. 6 (August), 1164–1175. Zeng, B.L.W., Lei, S., 1999. Extraction of multiresolution watermark images for resolving rightful ownership. In Proc. of SPIE, Security and Watermarking of Multimedia Contents 3657 (January), pp. 404–414. Zhang, X.D., Feng, J., Lo, K.T., 2003. Image watermarking using tree-based spatial-frequency feature of wavelet transform. J. Visual Comm. Image Representation (December), 474–491.