Robust Adaptive Image Watermarking using Visual Models in DWT and DCT Domain

INTRODUCTION

The explosive growth of Internet and communication has led to the tremendous use of multimedia data like image, audio and video. Furthermore, due to the availability of tools to manipulate digital multimedia especially digital images, tampering of such data has become very easy. In this context, it is important to ensure the integrity of images and protection against unauthorized duplication of images. A common technique for copyright protection is to embed a watermark in to the image or video data to be transmitted. The important requirements of such watermarks are imperceptibility, robustness and security. Watermark imperceptibility means that the watermark should be hidden in cover image in such a way that it can not been seen. So, it is necessary to exploit the characteristics of the Human Visual System (HVS) in the watermark embedding process. Robustness of a watermark is the ability to extract watermark correctly even if the intentional or unintentional attacks are made of the watermarked image. To ensure security, only the authorized user should be allowed to embed and extract the watermark. Several digital watermarking algorithms have been reported in the literature. Based on the domain in which the watermark is embedded, image watermarking techniques can be divided into two categories namely spatial domain techniques and frequency domain techniques. The watermark can be secret information or a content hash or another image such as a logo. The watermark is added either in the spatial domain or frequency domains (Mukherjee et al., 2004; Wong et al., 2003a; Dong et al., 2005; Kay and Izquierdo, 2001).

The theoretical models of the HVS have been vastly applied in image processing. A main purpose of exploiting the characteristics of the HVS is to effectively find the image information which can be removed without degrading the subjective image quality of the visual perception. A JND profile of an image is a key concept of the psycho-visual properties of the HVS. Although some proposed watermarking techniques employed the JND profile to enhance their transparency and robustness, they still suffer from some drawbacks. First, the JND a wavelet transformed image is not used in the design of their techniques during watermark embedding (Wong et al., 2003b; Joo et al., 2002). Second, the original images are required for the calculation of the JND profile of the images during watermark extraction (Zhang et al., 2003; Wang and Lin, 2004). Podilchuk and Zeng (1998) proposed two image-adaptive watermarking techniques. The first technique makes use of a DCT based visual mask and they employed Watson Just Noticeable Difference (JND) and the second method is based on a visual model using four-level wavelet decomposition. The original image decomposed into 4-level using a DWT. The algorithm then selects all the coefficients with magnitude larger than JND as the significant coefficients for all sub bands, except the base band and inserts the watermark into the selected coefficients. However, since this algorithm selects the significant coefficients using a fixed JND in each sub band, its robustness is decreased. Wong et al. (2003b) modified Watson’s model to estimate the JND profile for Discrete Wavelet Transform (DWT) coefficients and then used the JND profile to develop a watermarking method. Unfortunately, the method required original images while extracting watermarks. Seo et al. (2008) proposed robust image watermarking and the watermark is embedded in DC coefficients by using JND as watermark strength to improve the imperceptibility of watermark system but the method is not robust even for higher JPEG quality. Reddy and Varadarajan (2009) proposed wavelet based watermarking scheme using HVS model and they used entropy value as texture characteristic of HVS and the host image is decomposed by means of Haar wavelet transform with the lifting scheme to obtain the four sub-bands LL, LH, HL, HH. They employed entropy masking of HVS model in the selection of appropriate sub-band. Afterwards, the sub-bands except LL are considered and the one with maximum entropy is chosen for watermark embedding. But the quality of watermarked image is not very high, perhaps since they used a fixed HVS model and the entropy value alone is not good enough to be considered as characteristic of HVS which entropy value might differs for different host image. To achieve higher robustness and imperceptibility, we proposed two adaptive images watermarking algorithms based on DCT and DWT. In DCT based algorithm, we employed Watson’s model to estimate the JND value and improve robustness and the quality of watermarked image by controlling the power of JND strength. In DWT domain, the visual model is designed to generate a Just Noticeable Difference mask (JND) by analyzing image characteristics such as textures and luminance of the cover image.

HUMAN VISUAL SYSTEM MODELS (HVS)

HVS in dct domain: To balance the transparency and robustness, an effective watermarking method should exploit HVS masking characteristics. The JND models used in this work employed Watson’s JND model. The visibility threshold t^F_u,v as a function of spatial frequency response in specific viewing conditions is usually derived by the model presented by Peterson et al. (1993) The human visual system’s sensitivity to variations in luminance is dependent on the local mean luminance. Luminance sensitivity is estimated by the formula:

(1)

where, X_b is the DC coefficient of the DCT for block b, is the DC coefficient corresponding to the mean luminance of the display and a is the parameter which controls the degree of luminance sensitivity and its suggested value 0.649. Considering all the above parameters, the JND (a contrast masking threshold) is derived as:

(2)

where, w is a number between zero and one and can assume different value for each DCT basic function. A typical empirical derived value is 0.7. In our experiment we lower the value of w in order to increase the imperceptibility of watermarked image.

HVS in dwt domain: In order to design an effective robust watermarking, it is necessary to take into account the visual effect of embedding a watermark into a host image. Human eyes have different sensitivity to different luminance, most sensitive to middle level luminance usually, Weber ratio keeps const 0.02 within a large range of middle level and sensitivity declines nonlinearly within the low and high luminance range (Yang and Sun, 2007). We can use the Eq. 3 and use ω (u, v) as contrast sensitivity factor.

(3)

where, β denotes the maximum of contrast sensitivity, ave (u, v) is the average luminance of B_u,v, I₁ and I₂ the predetermined threshold value.

Human eyes are more sensitive to the noise in image smooth areas and less sensitive to the one in image texture areas. As for texture masking, we can use the local variance of wavelet sub block band because the variance is bigger at the textures and edges than at the smooth region so we use the variance of wavelet blocks v (u, v) as texture masking:

(4)

where, B (u, v) is wavelet sub blocks b and α is mean value of wavelet sub blocks. The effect of HVS masking characteristics is incorporated into the JND threshold value based on all above considerations as follows: Let ω (u, v) and v (u, v) be λ and δ, respectively.

(5)

where, Ψ and γ are predetermined threshold values and ρ is parameter which is used to control texture masking and Γ is the JND threshold value of wavelet sub blocks, max ( ) and min ( ) denotes the maximum and minimum set value respectively.

PROPOSED IMAGE ADAPTIVE WATERMARKING EMBEDDING IN DCT AND DWT DOMAIN

Watermarking embedding in dct domain: The proposed watermarking algorithm starts by partitioning the cover image I of size (MxN) into n by n non-overlapping blocks. The watermark image (logo or trademark) w is of size m xn. The steps involved in the proposed DCT domain embedding is presented as follows:

(6)

WATERMARK EMBEDDING IN DWT DOMAIN

The process of embedding the watermark image into the host image is presented in this sub-section. The host image I is first partitioned into 8 x 8 blocks and then each block is decomposed into four sub-bands (LL, LH, HL and HH) by applying DWT in each block. The watermark is embedded into host image by the altering the coefficients of the vertical and horizontal detail sub bands of image blocks. The watermark embedding technique is given as follows:

W = {w(i, j)| w(i, j) ε{0,1}, 0 = i = m-1, 0 = j = n-1}, PN = {pn(i, j)| pn(i, j) ε{0, 1}, 0 = i = m-1, 0 = j = n-1}. We employed (7) to modulate the watermark W and then obtain the final watermark, W ₀ which will be embedded:

(7)

(8)

where, σ is the watermark strength, φ is the average value of detail sub-bands and x is the detail sub-bands coefficients in each block

If W₀ = 1 and if X>Y and if X-Y<Γ and else if Y+X<Γ, we adopt Eq. 9 and 10 to hide the watermark:

(9)

(10)

PROPOSED WATERMARKING EXTRACTION IN DCT AND DWT DOMAIN

Watermarking extraction in dct domain: To extract the watermark we first partition the original image OI into blocks of size (NxN). Then each block are computed by applying the DCT transformation on the original image and the watermarked image and the difference between the two are computed to extract the watermark. The correlation coefficient between the extracted watermark and the original watermark is computed to extract the watermark. The watermark extraction algorithm presented below:

(11)

Watermarking extraction in dwt domain: In this section we presented watermark extraction and it is the inverse procedure of watermark embedding. The steps are as follows:

(12)

(13)

EXPERIMENTAL RESULTS

The simulated experiments are carried out to demonstrate the effect of the proposed schemes. In experiments, we adopt gray-scale images with size 512x512 as the original image and use binary images with size 64x64 as the watermark image. Figure 1a-h displays different host images. Let β, I_1, I₂ be 0.40, 60 and 120, ρ set to be 0.5 and Ψ and γ are set to 60 and 20, respectively.

Invisibility: To evaluate the invisibility of the watermark, we test the two proposed algorithms on series of standard images. Table 1 shows the PSNR values of the watermarked images. And we lower the value of w and can been that the PSNR value is higher and the quality of watermarked image is good and can be observed from Fig. 2b. And we take Lena image as an example; Fig. 2a and d show the original image, watermarked image and watermark images.

From Table 1 and Fig. 2, we can find that the watermarked images have high PSNR values and the watermark is invisible to human eyes. Even any visual difference was hardly noticeable on the watermarked image with lower PSNR value, because of taking advantage of HVS characteristics

Table 1:	PSNR values of watermarked images

Fig. 1:	(a-h) Host images

Fig. 2:	(a and b) Original image and watermarked image and (c and d) original watermark image, respectively

Robustness: This section examines the robustness against attacks, including JPEG compression with different quality factors, low pass filtering (LPF), sharpening and Histogram equalization, salt and pepper noise, Gaussian noise, resizing and so on. Figure 3a-g and 4a-g show extracted watermark image for DCT based method.

As shown in Fig. 3 and 4 that the proposed DCT method is robust for various image distortions and we can observe that the extract watermark is very good and has a good similarity with the original watermark.

The proposed DWT based algorithm is also tested for various types of image distortions. The distorted Lena image by noise and histogram equalization and corresponding extracted watermarks are shown in Fig. 5a-d and Fig. 6c-d, respectively.

Fig. 3:	Extracted watermark (a-c) JPEG compression with quality factor 85, 90 and 95% and (d-g) Gaussian noise (0.1, 0.2, 0.3 and 0.5%), respectively

It can be seen that both the proposed methods are robust against JPEG compressions, salt and pepper noise and low pass filtering and other image attacks and extract watermark is good. Figure 3a-c is shown that the extract watermark from JPEG compression with different quality factors and Fig. 3d-g extract watermark from Gaussian noise (0.1, 0.2 and 0.5%) respectively and this indicated that the proposed DCT method is robust against JPEG and Gaussian noise. And the watermarked image is attacked by low pass filtering, histogram equalization, sharpening and resizing and salt and pepper noise and the extracted watermark is shown in Fig. 4. As we can seen from the results the proposed DCT method is also robust against the above mentioned attacks.

Fig. 4:	Extracted watermark (a-c) LPF (Gamma: 0.9, 0.8 and 0.5) and (d-g) Histogram equalization, sharpening, resizing and 1% salt and pepper noise, respectively

Fig. 5:	(a-c) Distorted watermark image with salt and pepper noise (10, 3 and 1%) and (d) histogram equalization, respectively

The proposed DWT method is tested against JPEG compression, salt and pepper noise, low pass filtering and histogram equalization and the extracted watermark is shown in Fig. 6a-d and 7a-d show the extracted watermark from Gaussian noise, sharpening and resizing. As we can observed from Fig. 6 and 7 the extracted watermark can be identified easily and this shows that the proposed DWT method is robust against various image distortions.

Fig. 6:	Extracted watermark (a and b) JPEG compression with quality factor 85 and 90% and (c-e) salt and pepper noise (10, 3 and 1%) and (f and g) LPF and histogram equalization, respectively

Fig. 7:	Extracted watermark (a and b) Gaussian noise 0.1 and 0.3% and (c and d) sharpening and resizing, respectively

Table 2 displays the NC value between the original watermark and the extracted watermark from the attacked watermarked images. The experimental results demonstrate that the NC value is higher. The robustness of the proposed schemes is evident from the experimental evaluation. And Table 3 shows the robustness performance and both proposed methods show better performance than the existing method.

Table 2:	Robustness experimental results

Table 3:	Robustness performance