An Adaptive Video Watermarking Technique ... - Semantic Scholar

66 downloads 334 Views 606KB Size Report
An adaptive video watermarking technique based on DCT domain. Cheng-Han ... Seo et al. [9] pro- posed a quantization index modulation (QIM) watermark-.
An adaptive video watermarking technique based on DCT domain Cheng-Han Yang Department of Electrical Engineering National Tsing Hua University Hsinchu 300, Taiwan Hui-Yu Huang Department of Computer Science and Information Engineering National Formosa University Huwei, Yunlin, 632 Taiwan [email protected] Wen-Hsing Hsu Department of Electrical Engineering National Tsing Hua University Hsinchu 300,Taiwan Abstract

more effective method to protect the copyright. To solve this problem, many methods are developed [2, 3, 8]. Digital watermarking is a favorable method for copyright protection of the multimedia. It is a digital code embedded in the host data and typically contains information about origin, status, and/or destination of the data. Applications of watermarking technique include copyright protection, fingerprinting, authentication, copy control, tamper detection, and data hiding applications such as broadcast monitoring [8]. Most important of all applications is copyright protection, it is also the application most widely used nowadays. To aim at copyright protection, the watermarking must meet two main features: robustness and transparency. For robustness, it means that watermark should have ability to against various processing whether it is intentional or not. For transparency, it means that the watermark could not cause distortion seriously to decrease the visual quality of host data. Many kinds of watermarking techniques have been proposed which was worked in the spatial domain [5] and frequency domain [6, 10, 11, 12], etc. Lancini et al. [5] proposed a video watermarking technique in the spatial domain. In this approach, an important notion is mentioned: the compression algorithm will strongly decrease the chrominance quality. Many of techniques based on the frequency domain contain the discrete cosine transform (DCT), discrete fourier transform (DFT), discrete wavelet transform (DWT), quantization index method (QIM) [1, 4, 7, 9]. Chen and Wornell [1] proposed the quantization index modulation (QIM) and distortion-compensated (DC-QIM)

In this paper, we propose an effective video embedded watermarking technique based on DCT domain with high transparency and slight distortion. The watermark is mainly embedded into the uncompressed domain by adjusting the correlation between DCT coefficients of the selected blocks, and it can be detected without the original video data. The system contains the preprocessing, watermark embedding and watermark extraction. In the preprocessing, in order to improve the computation complexity and reduce the computation time, a pseudo 3D DCT by two times of the DCT transformation will firstly be obtained. In the embedding process, we embed the watermark into the successive raw frames transformed into the pseudo codes before compression, afterward a secret embedding key will be created. This secret embedding key will further use to the extraction processing. The experimental results show that the proposed scheme is extremely robust to against various attacks.

1. Introduction Owing to the rapid advance multimedia techniques and networks, People can arbitrarily and easily access or distribute any multimedia data by networks. Hence, this issue becomes more and more important, however, it has not

978-1-4244-2358-3/08/$20.00 © 2008 IEEE

589

Authorized licensed use limited to: Radu Preda. Downloaded on October 29, 2008 at 07:36 from IEEE Xplore. Restrictions apply.

CIT 2008

methods to embed the watermark in order to improve this program. By using of the (QIM) algorithm, it can achieve against arbitrary bounded and fully informed attacks and further arise to currently popular spread-spectrum method. Huang et al. [4] proposed an embedding algorithm using the quantization index modulation to resist the JPEG compression. This major steps of this approach had the optimal embedded position search, the embedded value decision, and the embedded/extracted processing for watermarks which worked in the DCT domain and adopted the quantization index modulation to achieve the purpose. Seo et al. [9] proposed a quantization index modulation (QIM) watermarking scheme for a digital image with two adaptive quantization step-sizes. Nguyen and Priemer [7] proposed a robust watermarking approach to embed a binary watermark sequence under the wavelet domain. This method is the robustness against attacks such as lowpass filtering, JPEG compression. In this paper, we propose an embedding watermark system based on the DCT domain. In our proposed approach, we will embed the watermark into the raw video data in order to adapt the various video compression standards and improve the performance. In addition, in order to avoid the distortion of the chrominance quality of video data, the watermark is only embedded into the luminance component of host data. Our proposed approach consists of preprocessing, watermark embedding and watermark detection. In embedding process, Firstly, each frame will be divided into a numbers of blocks which will be transformed into a DCT domain. Then, we will embed the watermark into a group of frames based on adjusting the correlation of the DCT coefficients by selected the blocks. While the watermark is embedded, some information can be recorded as a secret embedding key which can be used to enhance the security of the embedded system. In the detection process, the embedded watermark will be extracted by means of the secret embedding key. The rest of the paper is organized as follows. In Section 2, we describe watermark embedding and watermark extracting processes. In Section 3, we will present the experimental results. Finally, Section 4 gives the brief conclusions.

Figure 1. The system flowchart. order to achieve the security field, we apply the secret embedding key obtained by the embedding process to extract the watermark to improve the system robustness. As the same time, the detection process is used to this key to extract the embedded watermark without any original video. In the following, we will describe our method and the system flowchart is shown in Fig.1.

2.1. Preprocessing In this step, we take several successive frames as a group. Each frame within a group will be divided into a number of blocks which will be transformed into DCT domain by pseudo 3D DCT. By means of pseudo 3D DCT, our approach can reduce the computational complexity. Therefore, we embed the watermark only into luminance component of each frame in the uncompressed domain. In this way, the raw video data can be served as a sequence of still images. We embed the watermark into a group of frames by modifying the correlation of the successive frames. First, we take a number of consecutive frames as a group and every frame is divided into a number of blocks. Next, we pick the DC values of blocks located in the same position of consecutive frames of a group and transform those DC values into the DCT domain again. After transforming the second DCT process, we will obtain a new DC value and several AC values. This procedure is called a pseudo 3D DCT. The pseudo 3D DCT diagram is illustrated in Fig. 2. And the sum of all absolute AC values with weights is expressed as

2. Proposed method The proposed system consists of two steps watermark embedding and watermark detection. First of all, we will take the preprocessing to obtain the raw data from the videos. It is well-known that the characteristics of the DC coefficient possess the lowest frequency information, the stability, and the robustness in the DCT domain. For the embedding process, we embed the watermark into the DCT coefficients of some consecutive DC coefficients. Next, in

SU M (i) =

X

Wpre (i, k) × |AC(i, k)|,

(1)

j

where SU M (i), Wpre (i, k), and AC(i, k) denote the sum of all AC values,the corresponded weight, and AC value corresponding to the ith block,kth block respectively. Next, we arrange these sums and then achieve the embedding process.

590 Authorized licensed use limited to: Radu Preda. Downloaded on October 29, 2008 at 07:36 from IEEE Xplore. Restrictions apply.

Figure 3. method. Figure 2. Pseudo 3D DCT diagram.

Watermark

insertion by QIM

QIM method. The modification is expressed as:  2p, if WBit is 0, Matched, (a)    2p, if WBit is 1, Unmatched, (b) Q(i, k) = 2p + 1, if WBit is 1, Matched, (c)    2p + 1, if WBit is 0, Unmatched, (d)

For example, if we take four frames as a group, we will separate each frame into blocks and the size of every block is 8× 8 pixel, and further the block will be first transformed to DCT domain. Then we will pick the DC values of every block which locate the same position on the frames and transform these DC values to DCT domain again. After transforming process, we can obtain a new DC value and three AC values. According to Eq. (1), a sum of these AC values with their respective weight can be computed and obtained. By repeating above steps until all blocks of frames, we will acquire a sequence consisted of sums of every block. Finally, we arrange the sequence and start to embed the watermark.

(2) where p and WBit denote a random non-negative integer and the embedded watermark bit, respectively. If the relationship of Q(i, k) and the watermark bit conforms to Eq. (2) (a) or (c), Sum(i, k) will be not modified. Otherwise, Sum(i, k) will be changed. In order to increase the robustness of our proposed system, Sum(i, k) value will be changed to the center value corresponding to this section to gain the distortion tolerance. The insertion processing is illustrated as Fig. 3. After performing all blocks of frames within the group, we can derive the variation sequence Dif f (i) which consists of the D(i, k) of each block. It is reasonable that the blocks with the small variations will be selected to embed the watermark. Thus, the embedding positions can be determined according to the amount of D(i, k). Since Sum(i, k) is consisted of several AC values, the modification of Sum(i, k) equals to the change of AC values. Note that the low frequency component is more robust and visually sensitive than the high frequency component. That is, if the low frequency component is modulated, it will cause the distortions more seriously, but it has higher ability to resist attacks than the high frequency component does. Therefore, we use the weights to modulate the Sum(i, k) which is defined as: X Sum0 (i, k) = Ws (i, k, l)|AC(i, k, l)|

2.2. Watermark Embedding According to the preprocessing result, the Sum(i, k) values of all blocks can be obtained, and we try to find a threshold T (i) which can determine the robustness and the transparency of our proposed embedded system. This threshold is associated with the characteristic of video and the number of bits to be embedded. Because the different videos possess the distinct properties, based on this attribute, the threshold can be estimated, and it can help us to achieve the trade-off between the robustness and the transparency for the system. After determining the threshold, we divide each Sum(i, k) based on the threshold T (i), and the corresponding quotient Q(i, k) can be derived.

l

+We (i, k, l)D(i, k),

In order to embed the watermark bits, the Quantization Index Modulation (QIM) method is employed. Based on the QIM algorithm, the embedding domain is divided into several regions. The interval of every region is the same value which equals to the threshold T (i), and an index is assigned to each region. Every region represents a value of watermark. According to the Q(i, k) and the embedded bits, we will further modify the valuesof Sum(i, k) by means of the

(3)

where the D(i, k) represents the difference between Sum0 (i, k) and Sum(i, k). AC(i, k, l) and We (i, k, l) denote the AC value and the weights corresponding to the kth block within the ith group in the frames, respectively. We (i, k, l) can be adjusted by user and the sum of We (i, k, l) must equal to one. After determining the embedding position, we change the original value in the position into the center of the corresponding section by using

591 Authorized licensed use limited to: Radu Preda. Downloaded on October 29, 2008 at 07:36 from IEEE Xplore. Restrictions apply.

detection process is defined as  0, if Qex (i, k) = 2p, The embedded bit = 1, if Qex (i, k) = 2p + 1,

(4)

where p is a random non-negative integer. If Qex (i, k) is odd value, the embedded bit is 1. If Qex (i, k) is even value, the embedded bit is 0. By repeating above steps, we can exactly determine the embedded bits gradually until all watermark bits are extracted. Finally, using the secret seed S recorded in the secret embedding key, the watermark can be effectively recovered.

3. Experiments Based on our proposed method, we can effectively perform the embedding and detection the watermark to protect the copyright for multimedia data. Some results are displayed. In addition, we also compare our proposed method with Y. Wang’s method [11] to present the system performance. Here we use the 720×480 raw video sequence with 80 frames as the experimental data. Here, the watermark denotes a binary image with the size of 36×20. In experiments, we take four frames as a group, and the block size is 8×8. The threshold is computed the median of the searching region. The raw frame and the watermark are shown in Fig. 5. For the transparency measurement, we use the PSNR (peak-signal-to-noise ratio) value as a criterion expressed as

Figure 4. Watermark embedding process. Eq. (3). By repeating above procedures until all watermark bits are inserted, the embedding process will be achieved. Finally, all embedding positions, the secret seed S, weights Ws (i, k, l), and the threshold T (i) will be recorded as the secret embedding key. The embedding process is shown in Fig. 4.

PSNR = 10 log MSE =

2 Smax , MSE

h w 1 XX |S1 (x, y) − S2 (x, y)|2 , hw y x

(5) (6)

where S1 and S2 denote the corrupted and original images, respectively. The variables h and w represent the height and width of the image. For a gray level image, Smax represents 255 gray value. The PSNR values of 80 frames are displayed in Fig. 6. From this result, it is shown that the PSNR value of our proposed method is superior to the Y. Wang’s result. However, it can be obvious that the transparency of our proposed method is higher than that of Wang’s method.

2.3. Watermark Extraction The extraction process is the inverse of the embedding process. First of all, the raw video sequence is separated into several groups of frames and each frame is divided into blocks. Based on the embedding process, we can obtain the secret embedding key which is an important information to extract the embedded watermark. After determining the embedding blocks, we only transform the selected blocks into the DCT domain rather than all blocks of the frame. Thus, the computational complexity will be efficiently reduced. By applying pseudo 3D DCT to the selected blocks, we can further obtain Sumex (i, k), which is the sum of AC values of the kth blocks in the watermarked frames within the ith group by using Eq. (1). Then we divide Sumex (i, k) by the threshold T (i) which is recorded in the secret embedding key to calculate the quotient Qex (i, k). After obtaining Qex (i, k), we can determine which bit is embedded and the

To estimate the robustness of our watermarking process, we use the different attacks to demonstrate this performance. Because that the watermark is embedded into the uncompressed domain, we consider the influence for the compression schemes. The similarity measurement computed the difference between the extracted and the original watermark is expressed as Pw Ph ˆ i=0 j=0 W (i, j)W (i, j) NC = , (7) Pw Ph 2 i=0 j=0 [W (i, j)]

592 Authorized licensed use limited to: Radu Preda. Downloaded on October 29, 2008 at 07:36 from IEEE Xplore. Restrictions apply.

(a)

(b)

Figure 5. (a) Original frame, (b) watermark image.

Figure 7. NC values for the different MPEG compression bit-rate compared with our proposed method and Y. Wang’s method.

ˆ (i, j) present the original watermark where W (i, j) and W and extracted watermark, respectively. The variables h and w denote the height and width of the watermark. The results compared with our proposed method and Y. Wang’s method for the MPEG-2 compression shown in Fig. 7. For Fig. 7, the NC value of our proposed method is superior to Y. Wang’s method. Other results for the different attacks are illustrated in Figs. 8 and 9. According to above experiments, it is obvious that our proposed method is more robust for many attacks and compression process than Y. Wang’s method and the NC value can keep up 0.74.

4. Conclusions In this paper, we have proposed a robust video watermarking technique based on DCT domain to achieve the copyright protection. The experimental results show that our proposed method can obtain a good transparency and high robustness to against various attacks like noise, luminance modification and compression in the DCT domain. That is, our proposed approach can provide a trade-off mechanism between transparency and robustness. Furthermore, depending on the high resistance for the compression process, we can effectively inhibit the piracy via the net-

Figure 6. The PSNR value of video frames compared with our proposed method and Y. Wang’s method.

593 Authorized licensed use limited to: Radu Preda. Downloaded on October 29, 2008 at 07:36 from IEEE Xplore. Restrictions apply.

works in which video might be compressed with the low bit-rates. The other attacks will be the major study of our future work.

5. Acknowledgement This work was supported in part by the National Science Council of Republic of China under Grant No. NSC 952221-E-007-187-MY3.

References (a)

(b)

[1] B. Chen and G. W. Wornell. Quantization index modulation: a class of provably good methods for digital watermarking and information embeeding. IEEE Trans. on Information Theory, 47(4):1423–1443, May 2001. [2] I. J. Cox, F. J. Kilian, and T. Shamoon. Secure spread spectrum watermarking for multimedia. IEEE Trans. on Image Processing, 6(12):1673–1687, Dec. 1997. [3] C. T. Hsu and J. L. Wu. Hidden digital watermarks in images. IEEE Trans. on Image Processing, 8(1):58–68, Jan. 1999. [4] H. Y. Huang, C. H. Fan, and W. H. Hsu. An effective watermark embedding algorithm for high jpeg compression. In Proc. of Machine Vision Applications, pages 256–259, May. 2007. [5] R. Lancini, F. Mapelli, and S. Tubaro. A robust video watermarking technique in the spatial domain. In Proc. of the 8th IEEE Int. Symposium on Video/Image Processing and Multimedia Communications, volume 17, pages 251–256, June 2002. [6] A. S. Lewis and G. Knowles. Image compression using the 2-D wavelet transform. IEEE Trans. on Image Processing, 1(2):244–250, April 1992. [7] H. Nguyen and R. Priemer. Multiresoluation quantizationbased image watermarking. In Proc. of IEEE Int. Conf. on Eletro-Information Technology, pages 401–406, May. 2007. [8] C. I. Podilchuk and E. J. Delp. Digital watermarking: algorithms and applications. IEEE Signal Processing Magazine, 18(4):33–46, July 2001. [9] Y. S. Seo, W. G. Kim, Y. H. Suh, W. G. Oh, and C. J. Hwang. QIM watermarking for image with two adaptive quantization step-sizes. In Proc. of the 9th Int. Conf. on Advanced Communication Technology, volume 1, pages 797– 800, Aug. 2007. [10] V. Solachidis and I. Pitas. Circulary symmetric watermark embedding in 2-D DFT domain. IEEE Trans. on Image Processing, 10(11):1741–1753, Nov. 2001. [11] Y. Wang and A. Pearmain. Blind video MPEG-2 watermarking robust against scaling. In Proc. of IEEE Int. Conf. on Image Processing, volume 4, pages 2159–2162, Oct. 2004. [12] R. B. Wolfgang, C. I. Podilchuk, and E. J. Delp. Perceptual watermarks for digital images and video. Proc. of the IEEE, 87(7):1108–1126, July 1999.

(c)

Figure 8. (a) Watermarked frame with pepper and salt noise. (b) Extracted watermark of the proposed method, NC=0.9875. (c) Extracted watermark of Y. Wang’s method, NC=0.91111.

(a)

(b)

(c)

Figure 9. (a) Inverted the luminance component. (b) Extracted watermark of the proposed method, NC=1. (c) Extracted watermark of Y. Wang’s method, NC=0.74722.

594 Authorized licensed use limited to: Radu Preda. Downloaded on October 29, 2008 at 07:36 from IEEE Xplore. Restrictions apply.