Pattern Recognition 36 (2003) 969 – 975
www.elsevier.com/locate/patcog
Combinational image watermarking in the spatial and frequency domains Frank Y. Shih ∗ , Scott Y.T. Wu Computer Vision Laboratory, Department of Computer Science, New Jersey Institute of Technology, Newark, NJ 07102, USA Received 15 February 2002; received in revised form 24 April 2002; accepted 30 May 2002
Abstract In order to provide more watermarks and to minimize the distortion of the watermarked image, a novel technique using the combinational spatial and frequency domains is presented in this paper. The splitting of the watermark image into two parts, respectively, for spatial and frequency insertion relies on the user’s preference and data importance. Experimental results provide the comparisons when di7erent sized watermarks are embedded into a grayscale image. The proposed combinational image watermarking possesses the following advantages. More watermark data can be inserted into the host image, so that the capacity is increased. The splitting of the watermark into two parts makes the degree of protection double. The splitting strategy can be designed even more complicated to be unable to compose. Furthermore, to enhance robustness, a random permutation of the watermark is used to defeat the attacks of signal processing, such as image crops. ? 2002 Published by Elsevier Science Ltd on behalf of Pattern Recognition Society. Keywords: Image processing; Image watermarking; Compression; Security; Encryption; Signal processing
1. Introduction Digital images, video, and audio have revolutionized in the way of largely stored, manipulated, and transmitted. It gives rise to a wide range of applications in electronics, entertainment and medial industry [1]. Creative ways of storing, accessing and distributing data have generated lots of bene=ts into the digital multimedia =eld. However, these bene=ts brought with concomitant risks of data piracy. One of the solutions to provide security in copyright protection is watermarking which embeds special marks in a host image. Digital watermarking has been proposed as a way to identify the source, creator, owner, distributor, or authorized consumer of a document or an image. It can also be used for tracing images that have been illegally distributed. Watermarking, when complimented by encryption, ∗ Corresponding author. Tel.: +1-973-596-5654; fax: +1-973596-5777. E-mail address:
[email protected] (F.Y. Shih).
can serve many purposes including copyright protection, broadcast monitoring, and data authentication. There are many aspects to be noticed in watermarking design, for example, imperceptibility, security, capacity and robustness. Many researchers have been focusing on security and robustness [2– 6], but rarely on the watermarking capacity [7,8]. Indeed, both security and robustness are important because the watermark images are expected to be irremovable and unperceivable. Nevertheless, if we can embed a large watermark image into a host image, the application becomes more Fexible in many areas. There are two methods of performing watermarking, one in spatial domain, and the other in frequency domain. Each technique has its own advantage and disadvantage. In the spatial domain [3,9], we can simply insert watermark into a host image by changing the gray levels of some pixels in the host image, but the inserted information may be easily detected using computer analysis. In the frequency domain [4,5], we can insert watermark into the coeGcients of a transformed image, for example, using the discrete Fourier
0031-3203/02/$30.00 ? 2002 Published by Elsevier Science Ltd on behalf of Pattern Recognition Society. PII: S 0 0 3 1 - 3 2 0 3 ( 0 2 ) 0 0 1 2 2 - X
970
F.Y. Shih, S.Y.T. Wu / Pattern Recognition 36 (2003) 969 – 975
transform (DFT), discrete cosine transform (DCT) and discrete wavelet transform (DWT). But we cannot embed too much data in the frequency domain because the quality of the host image will be distorted signi=cantly. That is, the size of watermark should be smaller than the host image. Generally, the size of watermark is 1=16 of the host image. In order to provide more watermark data and to minimize the distortion of the watermarked image, a novel technique using the combinational spatial and frequency domains is developed in this paper. This paper is organized as follows. In Section 2, we present the overview of combinational image watermarking. Sections 3 and 4 introduce the techniques while embedding watermark in the spatial and frequency domains, respectively. Experimental results are shown in Section 5. Section 6 discusses the further encryption of combinational watermarking. Finally, the conclusions are made in Section 7.
2. Overview of combinational image watermarking In order to insert more data into a host image, the simple way is to embed them in the spatial domain of the host image. However, the disadvantage is that the inserted data could be detectable by some simple extraction skills. How can we insert more signals but unperceivable? We address a new strategy of embedding large watermarks into a host image by splitting the watermark image into two parts. One is embedded in the spatial domain of a host image, and the other in the frequency domain. Let H be the original gray-level host image with size N × N and W be the binary watermark image with size M × M . W 1 and W 2 are the two separated watermarks from W: H S is the image combined from H and W 1 in the spatial domain. H DCT is the image where H s is the transformed into the frequency domain by DCT. H F is the image where H DCT and W 2 are combined in the frequency domain. Let ⊕ denote the operation that substitutes bits of watermark for least signi=cant bits (LSB) of the host image. The algorithm of the proposed combinational image watermarking is presented below. Its Fowchart is shown in Fig. 1.
Algorithm: 1. Separate watermark into two parts: W = {w(i; j); 0 6 i; j ¡ M }, where w(i; j) ∈ {0; 1}, W 1 = {w1 (i; j); 0 6 i; j ¡ M1 }, where w1 (i; j) ∈ {0; 1}, W 2 = {w2 (i; j); 0 6 i; j ¡ M2 }, where w2 (i; j) ∈ {0; 1}, M = M 1 + M2 2. Insert W 1 into the spatial domain of H to obtain H S asH s = {hs (i; j) = h(i; j) ⊕ w1 (i; j); 0 6 i; j ¡ N }, where h(i; j); hS (i; j) ∈ {0; 1; 2; : : : ; 2L − 1} and L is the number of bits used as in the gray level of pixels. 3. Transform H S by DCT to obtain H DCT . 4. Insert W 2 into the coeGcients of H DCT to obtain H F as H F = {hF (i; j) = hDCT (i; j) ⊕ w2 (i; j); 0 6 i; j ¡ N }, where hF (i; j) ∈ {0; 1; 2; : : : ; 2L − 1}. 5. Transform the embedded host image by Inverse DCT. The criteria of splitting the watermark image into two parts, which are individually inserted into the input image in the spatial and frequency domains, depend on users and applications. In principle, the most important data exist in the center of the image. Therefore, a simply way of splitting is to select the central window in the watermark image to be inserted into the frequency domain. With the user’s preference, we can crop the most private data to be inserted into the frequency domain.
3. The watermarking in the spatial domain There are many ways of embedding a watermark into the spatial domain of a host image, for example, substituting the less signi=cant bits of some pixels [10], changing the paired pixels [11], and coding by textured blocks [12]. As shown in Fig. 2, the watermarking can be implemented by modifying the bits of some pixels in the host image. Let H ∗ be the watermarked host image. The algorithm is shown below. Algorithm: 1. Obtain pixels from the host image. H = {h(i; j); 06i; j¡N }; h(i; j)∈{0; 1; 2; : : : ; 2L − 1}: 2. Obtain pixels from the watermark. W = {w(i; j); 0 6 i; j ¡ M }:
Fig. 1. The Fowchart in combinational spatial and frequency domains.
Fig. 2. The Fowchart in spatial domains.
F.Y. Shih, S.Y.T. Wu / Pattern Recognition 36 (2003) 969 – 975
971
3. Substitute the pixels of the watermark into the LSB pixels of the host image. H ∗ = {h∗ (i; j) = h(i; j) ⊕ w(i; j); 0 6 i; j ¡ N }; h∗ (i; j) ∈ {0; 1; 2; : : : ; 2L − 1}: Fig. 3. The Fowchart in frequency domains.
4. The watermarking in the frequency domain Several approaches can be used in the frequency domain, for example, JPEG-based [13], spread spectrum [2,14], and content-based approaches [6]. How can we embed data into the frequency domain of a host image and it appears unperceivable? The transformation functions often-used are DCT, DWT, and DFT. Generally, we can insert data into the coeGcients of a transformed image. As shown in Figs. 3 and 4, we embed watermark into the coeGcients of a transformed host image. The important consideration is what locations are best to place for embedding watermark in the frequency domain to avoid distortion [15]. Let H m and W n be the subdivided images from H and W , respectively, H m DCT be the image transformed from H m by DCT, and H m F be the image combined by H m DCT and W n in the frequency domain. The algorithm is described as follows. Algorithm: 1. Divide the host image into sets of 8 × 8 blocks. H = {h(i; j); 0 6 i; j ¡ N }, H m = {hm (i; j); 0 6 i; j ¡ 8}, where hm (i; j) ∈ {0; 1; 2; : : : ; 2L − 1} and m is the total number of the 8 × 8 blocks. 2. Divide the watermark image into sets of 2 × 2 blocks. W = {w(i; j); 0 6 i; j ¡ M }, W n ={wn (i; j); 0 6 i; j ¡ 2}, where wn (i; j) ∈ {0; 1} and n is the total number of the 2 × 2 blocks. 3. Transform H m to H m DCT by DCT. 4. Insert W m into the coeGcients of H m DCT . H m F ={hm F (i; j) = hm hm
DCT
DCT
(i; j)⊕wm (i; j); 06i; j¡8};
(i; j) ∈ {0; 1; 2; : : : ; 2L − 1}:
5. Transform the embedded host image, H m F , by Inverse DCT. The criterion for embedding the watermark image into the frequency domain of a host image is that the total number of 8 × 8 blocks must be larger than the total number of 2 × 2 blocks.
5. Experimental results Fig. 5 illustrates that some parts of a watermark are important to be split for security purpose. For example, people
Fig. 4. The embedding skill in frequency domain.
cannot view whom the writer is in the images. Therefore, it is embedded into the frequency domain and the rest into the spatial domain. In this way, we not only enlarge the capacity, but also secure the information that we are concerned with. Fig. 6 shows the original Lena picture with size 256×256. Fig. 7 is the traditional technique of embedding a 64 × 64 watermark into the frequency domain of a host image. Fig. 7(a) is the original 64 × 64 watermark image and Fig. 7(b) is the watermarked Lena image by embedding Fig. 7(a) into the frequency domain of Fig. 6. Fig. 7(c) is the extracted watermark image from Fig. 7(b). Fig. 8 demonstrates the embedding of a large watermark, a 128×128 image, into a host image. Fig. 8(a) is the original 128 × 128 watermark image. Fig. 8(b) and (c) are the two divided images from the original watermark, respectively. We obtain Fig. 8(e) by embedding Fig. 8(b) into the spatial domain of Lena image and obtain Fig. 8(f) by embedding Fig. 8(c) into the frequency domain of the watermarked Lena image in Fig. 8(e). Fig. 8(d) is the extracted watermark from Fig. 8(f). Fig. 9(a) shows a larger watermark, a 256 × 256 image, which is split into two parts as in Fig. 5. Fig. 9(c) and (d) is the watermarked images after embedding watermark in the spatial and frequency domains of a host image, respectively. Fig. 9(b) is the extracted watermark from Fig. 9(d). The error measures used in this paper are normalized correlation (NC) and power signal-to-noise ratio (PSNR). They are de=ned as follows, where the symbols are de=ned in the
972
F.Y. Shih, S.Y.T. Wu / Pattern Recognition 36 (2003) 969 – 975
Fig. 5. A 64 × 64 image is cut from a 256 × 256 image.
Fig. 7. The traditional technique embedding a 64 × 64 watermark into Lena image. Fig. 6. Lena image.
6. Further encryption of combinational watermarking
previous sections: N N NC =
F j=1 H (i; j)H (i; j) ; N N 2 i=1 j=1 [H (i; j)]
i=1
PSNR = 10 log10
N N
i=1 j=1 [h N N i=1 j=1 [h(i; j)
∗
(i; j)]2
− h∗ (i; j)]2
:
Table 1 presents the results of embedding di7erent sizes of watermark into a host image. PSNR in the =rst step means comparing both the original and the embedded Lena images in the spatial domain. PSNR in the second step means comparing both images in the frequency domain. NC means the normalized correlation by comparing both the original and the extracted watermarks.
For the purpose of enhancing robustness, a random permutation of the watermark is used to defeat the attacks of signal processing, such as image crops. The procedures are shown in Fig. 10. A random sequence generator is used in order to relocate the order of sequential numbers [5]. For example, the 12-bit random sequence generator is used to relocate the order of a watermark with size 64 × 64, as shown in Fig. 11. We rearrange the bits 9 and 6 to the rear of the whole sequence and the result is shown in Fig. 11(b). Fig. 12 shows the result when a half of the Lena image is cropped. Fig. 12(a) is the original 128 × 128 watermark, Fig. 12(b) is the cropped Lena image, and Fig. 12(c) is the extracted watermark.
F.Y. Shih, S.Y.T. Wu / Pattern Recognition 36 (2003) 969 – 975
Fig. 8. The results when embedding a 128 × 128 watermark into Lena image.
Fig. 9. The results when embedding a 256 × 256 watermark into Lena image.
973
974
F.Y. Shih, S.Y.T. Wu / Pattern Recognition 36 (2003) 969 – 975
Fig. 10. The procedures of randomly permuting the watermark. Table 1 The comparisons when embedding di7erent sized watermarks into a Lena image with size 256 × 256
PSNR in =rst step PSNR in second step NC
64 × 64
128 × 128
256 × 256
None 64.57 1
56.58 55.93 0.9813
51.14 50.98 0.9644
Fig. 11. A 12-bit random sequence generator.
Fig. 12. The results when cropping half of a Lena image.
Table 2 shows the results when parts of an embedded host image with a watermark of 128 × 128 are cropped. For example, if the size of 64 × 64 is cropped from a 256 × 256 embedded host image, the NC is 0.92. If a half of the embedded host image is cropped (i.e., eight of 64 × 64) as shown in Fig. 12(b), the NC is 0.52.
Table 2 The comparisons when cutting di7erent sizes of a Lena image
7. Conclusions
spatial domain, we mainly substitute the LSB bits in a host image with the bits in the watermark image. In the frequency domain, we insert data into the low frequency of a host image. Simulation results demonstrate the embedding a large watermark into a host image.
In this paper, we have presented the progression of embedding a large watermark into a host image. The technique is based on both spatial and frequency domains. In the
Times of 1 64 × 64 NC 0.92
2
3
4
5
6
7
8
0.88
0.81
0.73
0.68
0.61
0.58
0.52
F.Y. Shih, S.Y.T. Wu / Pattern Recognition 36 (2003) 969 – 975
Mostly, the important part of an image is not enormous. We can cut the important part and embed it into the frequency domain and the rest into the spatial domain of a host image. Therefore, we cannot only enlarge the size of watermark, but also remain the property of security and imperceptibility. The proposed combinational image watermarking possesses the following advantages. More watermark data can be inserted into the host image, so that the capacity is increased. The splitting of the watermark into two parts makes the degree of protection double. The splitting strategy can be designed even more complicated to be unable to compose. References [1] H. Berghel, L. O’Gorman, Protecting ownership rights through digital watermarking, IEEE Comput. Mag. 29 (7) (1996) 101–103. [2] I. Cox, J. Kilian, T. Leighton, T. Shamoon, Secure spread spectrum watermarking for images audio and video, Proceedings of the 1996 IEEE International Conference on Image Processing (ICIP’96), Lausanne, Switzerland, Vol. III, pp. 243–246. [3] N. Nikolaidis, I. Pitas, Robust image watermarking in the spatial domain, Signal Process. 66 (3) (1998) 385–403. [4] Jiwu Huang, Yun Q. Shi, Yi Shi, Embedding image watermarks in DC components, IEEE Trans. CSVT 10 (6) (2000) 974–979. [5] Shinfeng D. Lin, Chin-Feng Chen, A robust DCT-based watermarking for copyright protection, IEEE Trans. Consumer Electron. 46 (3) (2000) 415–421.
975
[6] P. Bas, J.-M. Chassery, B. Macq, Image watermarking: an evolution to content based approaches Pattern Recognition 35 (2002) 545–561. [7] M. Barni, F. Bartolini, A. De Rosa, A. Piva, Capacity of the watermark channel: how many bits can be hidden within a digital image? Proc. SPIE 3657 (1999) 437–448. [8] P. Moulin, M.K. Mihcak, The data-hiding capacity of image sources, IEEE Trans. Image Process. (2002), in press. [9] O. Bruyndonckx, J.-J. Quisquater, B. Macq, Spatial method for copyright labeling of digital images, Proceeding of IEEE Workshop on Nonlinear Signal and Image processing, Neos Marmaras, Greece, 20 –22 June 1995, pp. 456 – 459. [10] R. Wolfgang, E. Delp, A watermarking technique for digital imagery: further studies, International Conference on Imaging Science, Systems and Technology, Los Vegas, Nevada, Juillet 1997. [11] I. Pitas, T. Kaskalis, Applying signatures on digital images, Workshop on Nonlinear Signal and Image Processing, IEEE, Neos Marmaras, June 1995, pp. 460 – 463. [12] G. Caronni, Assuring ownership rights for digital images, Proceedings of Reliable IT Systems, VIS’95, Viewveg Publishing Company, Germany, 1995. [13] K.E. Zhao J., Embedding robust labels into images for copyright protection, Technical Report, Fraunhofer Institute for Computer Graphics, Darmatadt, Germany, 1994. [14] I. Cox, J. Killian, T. Leighton, T. Shamoon, Secure spread spectrum watermarking for multimedia, IEEE Trans. Image Process. 6 (12) (1997) 1673–1687. [15] C.-T. Hsu, J.-L. Wu, Hidden signature in images, IEEE Trans. Image Process. 8 (1999) 58–68.
About the Author—FRANK Y. SHIH received the B.S. degree from National Cheng-Kung University, Taiwan, in 1980, the M.S. degree from the State University of New York at Stony Brook, in 1984, and the Ph.D. degree from Purdue University, West Lafayette, Indiana, in 1987, all in Electrical Engineering. He is presently a professor jointly appointed in the Department of Computer Science (CS) nad the Department of Electrical and Computer Engineering (ECE) at New Jersey Institute of Technology, Newark, NJ. He currently serves as the Associate Chair of the CS Department and the Director of Computer Vision Laboratory. Dr. Shin is on the Editorial Board of the International Journal of Systems Integration. He is also an associate editor of the International Journal of Information Sciences, and of the International Journal of Pattern Recognition. He has served as a member of several organizing committees for technical conferences and workshops. He was the recipient of the Research Initiation Award from the National Science Foundation in 1991. He was the recipient of the Winner of the International Pattern Recognition Society Award for Outstanding Paper Contribution. He has received several awards for distinguished research at New Jersey Institute of Technology.He has served several times in the Proposal Review Panel of the National Science Foundation on Computer Vision and Machine Intelligence. He holds the IEEE senior membership. Dr. Shih has published over 110 technical papers in well-known prestigious journals and conferences. His current research interests include image processing, computer vision, computer graphics, arti=cial intelligence, expert systems, robotics, computer architecture, fuzzy logic, and neural networks. About the Author—YI-TA WU was born in Taipei, Taiwan, on March 22 1971. He received the B.S. degree in Physics from Tamkang University, Taipei, Taiwan, in 1995, and the M.S. degree in Computer Science from National Dong-Hwa University, Hualien, Taiwan, in 1997. He is current a Ph.D. student in the Computer Science Department, New Jersey Institute of Technology, Newark, NJ. His research interest include image processing, current vision, pattern recognition, and database.
