Chen-hui YI,Xiang-zhen ZHOU,Xiao-mao HOU

(1School of Electronic Information, Hunan Institute of Information Technology， Changsha 410151， China) (2School of Computer Science and Engineering, Beihang University, Beijing 100191, China)

Abstract: In order to effectively improve the imperceptibility, robustness and security of image watermarking technology, a robust blind watermarking scheme based on region of interest and discrete wavelet transform is proposed. This scheme uses the region of interest of the host image as the watermark image. The requirements are met by the watermark generated from the host image, in conjunction with the embedding strategy and Arnold scrambling. First, the first-order discrete wavelet transform is applied to the watermark, and the approximation coefficient is selected as the information to be embedded. Each approximation coefficient is embedded in the low frequency subband of the selected block of the host image in the wavelet domain; then, prior to embedding, the approximation coefficients of the watermark and the block of the host image are Arnold scrambled. This makes the solution more robust and secure. The test data show that compared with the current watermarking method, the proposed technique has better robustness and security, and the correlation coefficient NC of the watermark image is stable at around 0.98. Therefore, the proposed scheme could resist various attacks and achieve high security, imperceptibility and robustness.

Key words: Image watermarking, Discrete wavelet transform, Arnold scrambling, Wavelet domain, Low frequency coefficient, Region of interest, Blind watermark

With the rapid development of digital multimedia technology and the Internet, people can more easily copy, transmit, distribute and store information. Therefore, copyright protection and copy protection of multimedia data are becoming more and more important. Digital watermarking is an effective method for copyright protection, copy protection, proof of ownership, etc. [1]. Digital watermarking is a technique for embedding copyright or other information (called a watermark) into an image or audio or video [2]. Watermarks can be extracted from multimedia to prove ownership or to obtain some copyright-related information [3]. In some copyright protection applications, the watermark extraction algorithm can use the original image to find the watermark, called a non-blind watermark. In most other applications, such as copy protection, the watermark extraction algorithm cannot access the original image. This watermarking algorithm is called blind watermarking [4]. Therefore, we should maintain a balance between the robustness and imperceptibility of the watermark.

According to the watermark embedded domain, the watermarking technique can be divided into a spatial domain and a transform domain. Spatial domain technology embeds a watermark by directly changing the pixel values of the image. The transform domain technique implements the embedding process by changing the transform coefficients [5]. The airspace technique is simple, but not as robust as the transform domain technique for different attacks [6]. The three common transform domain methods are discrete cosine transform (DCT) [7-9], discrete Fourier transform (DFT) [10-14], and discrete wavelet transform (DWT) [15-20]. Among the transform domain methods, wavelet-based methods are more popular due to their superior frequency positioning characteristics. Hua et al [6] used the SIFT mechanism to quickly and accurately detect the feature points of the carrier image, and then based on the non-subsampled Contourlet transform mechanism, the experimental results verify the robustness and security of the algorithm. Saeid Fazli et al. [7] decomposed the carrier image based on the DWT mechanism to obtain four frequency domain sub-blocks. When the watermark information is embedded in the carrier image, the above algorithm could cause a large change in its RGB component, which reduces its imperceptibility. In addition, the imperceptibility and robustness of such technologies need to be further improved.

In order to solve the above problems, this paper proposes a blind watermarking algorithm based on region of interest (ROI) and discrete wavelet transform to simultaneously satisfy the imperceptibility, robustness and security of replication protection applications. Unlike the above watermarking technique in which the watermark is obtained from an external source, the technique uses the ROI of the host image as the watermark. The approximation coefficients of the watermark are embedded in the low frequency band of the image block to provide better robustness against most attacks. Prior to embedding, the approximation coefficients of the watermark and the block of the host image are subjected to Arnold transform to make the scheme more robust and secure. During the embedding process, the individual coefficients of the watermark are replaced by coefficients in the image block. The difference between the considered coefficients and the watermark coefficients is very small. The simulation results show that compared with the existing scheme, the proposed technique exhibits the greatest robustness against attacks such as JPEG compression and filtering.

1 Arnold transformation principle

The Arnold transform is a 2D chaotic map used to scramble digital images. The Arnold transform (also known as cat face transform) can be used to scramble the image so that the original meaningful image becomes a meaningless image. The Arnold transformation is intuitive, simple, and cyclical, making it easy to use. This transformation can pre-process the image before other image processing, such as scrambling the watermark before the digital blind watermark is embedded. Usually an Arnold transform does not achieve the desired effect, and the image needs to be transformed multiple times in succession. The Arnold transform has periodicity, that is, the Arnold transform could be continuously performed on the image, and finally the original image can be obtained.

Assume that the size of the original image is N×N. The Arnold transformation is defined as follows:

(1)

Where,x,y∈ {0, 1, 2, …,N-1}. Each pixel (x,y) in the image is converted to another pixel (x′,y′) by the Eq. (1). The scrambled image is obtained when all of the pixels in the image are transformed. Arnold scrambling is a periodic process, so the original position of the (x,y) coordinates is restored after thetthiterations. The factor t is called the transformation period. Because of its periodicity, it is used in many digital image scrambling processes. Arnold mapping is the easiest of the various scrambling methods. This mapping is reserved because the determinant of the transformation matrix is 1[18]. In addition, it can be extended to higher dimensions. This mapping provides better security due to the increased number of security keys.

2 The proposed algorithm

2.1 Watermark embedding

The ROI of the host image is extracted and a level 1 DWT is applied to obtainLL,LH,HL, andHHsub-bands. TheLLsub-band is transformed n times by the Arnold transform and the scrambling matrixLsis obtained. The individual components of the scrambling matrixLsare embedded in selected blocks of the host image. Fig.1 shows a block diagram of the steps of the proposed embedding algorithm.

Fig.1 Watermark embedding step block diagram of the algorithm

2.2 Watermark extraction

The extraction process is the opposite of the embedding process. The watermark can be extracted from the distorted watermarked image without using the host and the watermark image. This scheme requires the use of secret keysK1andK2, as well as iteration factorsnandmfor watermark extraction. The block diagram of the watermark extraction algorithm is shown in Fig.2.

3 Results and discussion

This section will test the proposed algorithm on different grayscale images of size 512 × 512 pixels, namely Lena, Boat, Elain and Peppers. A region of interest having a size of 64 × 64 pixels is extracted from the image as a watermark image. The Fig.3 shows all the test images and watermarks. The size of the block used for watermark embedding is 8×8 pixels.

Fig.2 Block diagram of the watermark extraction step of the proposed algorithm

Fig.3 Test image and watermark

3.1 Unperceptibility analysis

The peak signal-to-noise ratio (PSNR) is used to measure the quality of the watermarked image, which is defined as follows:

(2)

Where,N1andN2represent the number of pixels in each row and column of the image, respectively;XandYrepresent the original and watermarked images, respectively; max represents the maximum value of the pixel values in imageY. The PSNR of all watermarked images is given in Table 1. Furthermore, since the watermark is part of the image and is closely related to the host image, the replacement factor does not degrade the visual quality of the host image.

Table 1 PSNR values of watermarked images

3.2 Robustness analysis

In order to evaluate the robustness of the proposed algorithm, geometric and non-geometric attacks are applied to the watermarked image, and then the watermark is extracted from the attacked watermarked image and compared with the existing DWT-SVD-based method [20]. The relevant parameter pairs of the two methods are shown in Table 2. The adopted attacks include salt and pepper noise, Gaussian noise, speckle noise, JPEG compression, median filtering, averaging filtering, and Wiener filtering. All attacks were simulated by using MATLAB 7.8. In order to compare the similarity between the original watermark and the extracted watermark, the normalized correlation (NC) is calculated as follows:

(3)

Where,WandWErepresent the original watermark and the extracted watermark, respectively;M1andM2represent the size of the watermark image, respectively. The evaluation results of the test images are given below. For the sake of brevity, we show the visual effects of Lena and Boat image simulation.

Table 2 Comparison of relevant parameters of proposed scheme and scheme in [20]

3.2.1 Noise adding analysis

To analyze the effects of noise, three different variances of noise are added to the watermarked image. The result of adding a noise attack indicates that the extracted watermark is highly correlated with the original watermark. Therefore, the robustness of the algorithm under this attack is verified. Table 3 compares the proposed scheme for the Lena image after noise attack with the NC of the scheme in literature [20]. As can be seen from Table 3, the proposed scheme exhibits better performance in resisting attacks.

3.2.2 JPEG compression analysis

The watermarked image is compressed to have different quality factors from 1 to 90. When the quality factor is lowered, the quality of the watermarked image is reduced, but the NC value is still high. As shown in Fig.4, it shows the NC value of the extracted watermark and different quality factors. As can be seen from the this figure, the NC values of all the test images are higher than 0.99. This shows that the scheme is very robust to JPEG attacks. Fig.5 compares the robustness of the JPEG compression attack of Lena images with the proposed scheme and the scheme of literature [20]. As shown in Fig.5, both solutions have the same robustness for high quality factors. However, for lower quality factors, the proposed scheme performs better than that of the scheme in literature [20].

Table 3 NC comparison between the proposed scheme and the method in [20]

Fig.4 NC value of the image extracted watermark during JPEG compression attack

Fig.5 NC comparison between the proposed scheme and the method in [20] for Lena image against rotation attack

3.2.3 Filter analysis

Median filtering, averaging filtering, and Wiener filtering with different template sizes are performed on the watermarked image. The watermark image after these attacks is blurred at the edges. Fig.6(a)-(c) show the NC values of the watermarks extracted for all test images after S×S filtering, where S ranges from 3 to 50. Fig.6(a) shows the NC value of the watermark extracted after the filter size S is changed under median filtering. Fig.6(b) shows the NC value of the watermark extracted after the filter size S is changed under the average filtering. Fig.6(c) shows the NC value of the watermark extracted after the filter size S is changed under Wiener filtering. As can be seen from Fig.8, a higher NC value is achieved with filtering with different template sizes. Even when the watermarked image is severely degraded by a 50x50 filtering attack, the proposed scheme can effectively detect the watermark. This shows that the proposed scheme is robust to filtering attacks. Fig.6(d) compares the simulation results of the Lena image under the median filtering attack with the proposed scheme in literature [20]. It is obvious that the proposed scheme is superior to the scheme in literature [20].

Fig.6 Comparison of NC values of image extracted watermark after filtering

4 Conclusion

This paper proposes a blind watermarking scheme for Arnold scrambling images based on ROI and DWT. This scheme satisfies its imperceptibility, robustness, and security by using watermarks generated by host images, embedded strategies and Arnold scrambling. This scheme uses the ROI of the host image as a watermark to enhance robustness and imperceptibility. In the absence of a host image, the watermark extraction can be performed using the key and iteration factor defined in the embedding process. By hiding this information, a high level of security is achieved. The robustness of the scheme was tested by various attacks and the results were compared with the existing methods. Watermarks can be efficiently extracted from significantly distorted images when subjected to various attacks, such as JPEG compression, noise addition, median filtering, averaging filtering, Wiener filtering, resizing, and rotation. The results show that this method is superior to the existing methods in resisting most conventional attacks.