Robust Blind Watermarking Method for High Capacity RGB Image in Wavelet Domain

Multimedia content has been increasingly distributed on the Internet and is now easily available to view and download [1]. However, this widespread availability of digital information raises the issue of protecting this digital content. Digital watermarking has emerged as a powerful technique that has emerged to address the issue of digital content protection by embedding identification or proprietary marks directly into the digital host, such as images, audio, and video. This protection technique is used for authentication, copyright, secret communication, biomedical data authentication, and many others [2-4]. Watermarking embeds the secret data, i.e., watermark, into the digital content of multimedia directly using embedding algorithms instead of making the multimedia content unintelligible/readable. In addition, the watermark can be visible or invisible, depending on the application [5]. However, in digital image watermarking, two of the most important principles, i.e. imperceptibility and robustness, are given priority in such an application [6]. The watermark that is hidden within the content should ideally remain undetectable by humans, ensuring it does not compromise the quality or disrupt the experience of the original material. At the same time, it needs to be strong enough to withstand typical signal processing actions without being compromised.

RGB image is one of the digital content that has been widely used in watermarking. The watermarking of RGB images plays a vital role in the protection of digital visual content by embedding imperceptible but resilient information within the red, green, and blue channels of an image. Unlike grayscale watermarks, RGB watermarks embed information in multiple color channels, enhancing robustness against attacks such as compression, crop ping, and noise addition [7]. Furthermore, color images can significantly increase the capacity and comprehension of information [8]. Therefore, RGB image watermarking plays a significant role in fields such as medical imaging, e-Commerce, and digital forensics, where data integrity and source identification are paramount. However, working on RGB space poses sev eral significant challenges that impact its effectiveness and reliability. One primary challenge is achieving the balance between imperceptibility and ro bustness, which is inherently difficult. Another issue is the vulnerability to desynchronization attacks, where modifications, such as cropping or rota tion, can interfere with the alignment between the embedded watermark and the detection algorithm, leading to failed watermark retrieval [9]. Further more, the growing adoption of deep learning techniques presents challenges such as the requirement for larg training datasets and the potential risk of overfitting, which can affect the generalizability of watermarking models [10]. These challenges require ongoing research to develop more sophisticated and resilient watermarking methods that can effectively protect digital images digital images while maintaining their quality. Digitally, several domain-based watermarking methods have been presented in both spatial and frequency domains. In the spatial domain, the watermarks are embedded in the pixels directly by modifying the values of intensities. One of the most popular methods for embedding watermarks in the spatial domain is least significant bit (LSB), which slightly affects the perceptibility of the image [11]. Quantization-based methods have also been used in this domain [12]. Despite the simplicity and low computation rate of applying the watermark in the spatial domain, the watermarking performed in the spatial domain is relatively weak against image-processing-based attacks, such as JPEG compression, filtering, and noise addition [5, 13, 14].

In the frequency (spectral) domain, various transforms have been experienced in watermarking applications to embed the watermark. In these transforms, the carrier image is firstly transformed into their frequency-based version and information hiding is carried out in the frequency components instead of the spatial content. The frequency domain based watermarking algorithms have the advantages of robustness against image processing based attack [13] and producing a watermarked image with less distortion [14]. The frequency domain digital watermarking can be achieved using the most common transforms such as discrete wavelet transform (DWT), Discrete Cosine Transformation (DCT), Discrete Fourier Transformation (DFT). These transforms offer greater robustness against various attacks [15]. DFT is the early transform that is explored to hide the watermark in image. In this context, we found the examples in studies [16-19]. DWT is also extensively used to meet the authentication copyright protection criteria in applications for protecting multimedia content. In DWT, the host image is decomposed into levels producing details and approximation sub-bands. DWT is commonly applied to embed a single watermark in one of the produced bands [20], however, two bands are also used to embed two watermarks [21]. Using these sub-bands for embedding provides better imperceptibility and quality of watermarked images. Consequently, the number of embedded watermarks can be increased without affecting the quality of images. In the proposed method, RGB carrier image is decomposed into two and three levels for each color band using DWT. Then, a single watermark is inserted in a low-level sub-band of each color band based on measuring the homogeneity of the region of embedding. This method provides better imperceptibility, quality of watermarked image, high-capacity watermark, and high authentication standard. The main contributions of the proposed method are:

1. Proposing a new blind RGB image watermarking method using DWT, embedding three watermark logos instead of just one. Each RGB channel contains a separate watermark, increasing authentication reliability and watermark capacity.
2. Enhancing the imperceptibility and robustness of the watermarking against various attacks by embedding the watermarks in the frequency domain instead of directly modifying pixel values. This scheme provides an additional layer of protection, leading to better invisibility and higher resistance to attacks like noise addition, filtering, and sharpening.
3. Utilizes all three RGB color channels for embedding multiple logos, significantly increasing payload capacity without degrading the quality of the watermarked host image.
4. Blind extraction of the watermark without requiring the original image, making it more practical for real-world applications.

The remainder of this paper is structured as follows: Section 2 discusses the related studies on DWT based watermarking techniques. Section 3 explains the proposed watermarking method. The results are analyzed and discussed in Section 4. Finally, conclusions are drawn in Section 5.

This section introduces a robust blind watermarking technique utilizing the DFT domain instead of the spatial domain. Figure 1 shows the pipeline of the proposed method. This technique offers a robust watermarking approach that reduces the distortion and increases the imperceptibility. The proposed method examines the correlation among the frequencies of the adjacent pixels to estimate their smoothness or contrast levels, thereby determining the payload capacity for each embedding location. Accordingly, the blocks located in the edge area embed more data than those in the smooth area. In addition, the data embedded within the watermarked image is retrieved without the need to reference the original image.

9-1.png

Figure 1. The block diagram of the proposed watermarking method

3.1 Watermark logo embedding

Given the host RGB image, $I$, with size N × N and the binary watermark logo image, w, with size n × n. $I$ is firstly decomposed into its R, G, and B color components. Then, l-levels DWT is applied to these components, separately, resulting in the low-frequency component, i.e. approximation, for each color channel. The approximation coefficients are selected to embed w. Suppose $A_i$ represents the matrix of approximation coefficients of the color components i, where i ∈ {R,G, and B}, after applying the $l$-levels DWT. $A_i$ is firstly partitioned into square blocks of size 3 × 3 with 30% of overlapping to ensure that the embedding positions are neither altered nor involved in calculating the embedding payload. For each overlapping block, the logo data is then hidden in $A_i$ bit by bit. During embedding, the bit e, where e = 1,2,...,n × n, of w is added to the scanned overlapped block. This implies that when the embedded bit is equal to 1, an additional bit is added to the magnitude of the centered location in $A_i$. In contrast, if the embedded bit is equal to 0, the average of the four surrounding coefficients replaces the original coefficient magnitude of the centered location at the current scanned overlapped block, as demonstrated in Eq. (1).

$\hat{\mathrm{A}}_i^c=\left\{\begin{array}{c}A_i^c+w(e) \ if \ w(e)=1 \\ \bar{A}_i \quad { if }\ w(e)=0\end{array}\right.$          (1)

where, $\hat{\mathrm{A}}_i^c$ and $A_i^c$ represent the watermarked approximation and the original approximation coefficients of the centered position $c$, respectively, at the current scanned overlapped block in $A_i.  \overline{A_i}$ is the average approximation coefficient of the current scanned overlapped block, which is computed using the following formula:

$\bar{A}_i=\frac{A_i^l+A_i^r+A_i^b+A_i^u}{4}$           (2)

where, $A_i^l, A_i^r, A_i^b$, and $A_i^u$ represent the left, right, bottom, and upper neighboring of the current scanned location, respectively.

Applying Eq. (1) leads to increase or decrease the homogeneity in the target block, which contributes in the extraction stage without needing the origin host image. On one hand, the level of homogeneity inside each block is increased when the embedded bit is 0 by replacing $A_i^c$ with $\bar{A}_i$. On the other hand, embedding a 1-bit secrete  decreases the homogeneity inside the target block. The aforementioned framework is repeated on each color channel, allowing the algorithm to hide three logo watermarks instead of a single watermark. Consequently, the payload increases by including three watermark logos, and the authentication is improved. After completion of the embedding task, the inverse discrete wavelet transform, IDWT, is applied to reconstruct the watermarked image $\hat{I}$.

Furthermore, making minimal changes using 2 leads to less noticeable and higher quality watermarking. In addition, utilizing the three color bands in hiding significantly increases the embedding capacity, allowing three watermarks to be embedded in a single image and enhancing the level of security. In the next section, blind extraction of the watermark logo images will be explained in detail.

3.2 Watermark logo extraction

The proposed extraction stage is blind where the embedded logo image is achieved without relying on the original carrier image or any information that can help as a reference. In this context, the same kernel that was used in embedding is used again to recover the watermark logos from the watermarked image.

Given the RGB watermarked image, $\hat{I}$, the watermarked color components R, G and B are obtained. Each color component is decomposed, separately, with l-levels DWT, resulting in the approximation and detail coefficients of this component. Suppose $\hat{\mathrm{A}}_i$ represents the matrix of approximation coefficients that includes a watermark logo of the color components i, where i ∈ {R, G, and B}, after performing l-levels of DWT to $\hat{I}$. Then, ${A}_i$ is partitioned into overlapped blocks with 30% overlapping and it is scanned in a raster mode. At the current scanned overlapped block, a window mask, e.g. the kernel shown in Figure 2, is applied to extract the logo data as follows

$\mathcal{D}=\left|\frac{\hat{A}_i^u+\hat{A}_i^r+\hat{A}_i^b+\hat{A}_i^l}{4}-\hat{A}_i^c\right|$          (3)

where, $\mathcal{D}$ represents the absolute value of the difference between the approximation coefficient at the center of current block, $\hat{A}_i^c$, and its neighboring. Then, logo image bits, $\widehat{w}$, are collected as follows

9-2.png

Figure 2. The cross operator of getting the approximation coefficient and its neighbors

The kernel structure shown in Figure 2 enables the algorithm to extract the logo bits without requiring knowledge of the embedding locations. This means that the embedding approach is blind. Then, logo image bits, $\widehat{w}$, are collected as follows

$\widehat{\mathrm{w}}(e)=\left\{\begin{array}{c}0 \text { if } \mathcal{D} \leq \tau \\ 1 \ { otherwise },\end{array}\right.$           (4)

Eq. (4) implies that a watermark logo bit of 0 is extracted when the coefficients of the current block are homogeneous according to a predefined threshold τ. In contrast, a watermarked logo bit of 1 is obtained when the coefficients of the current block vary with the surrounding coefficients, less homogeneity. The level of homogeneity inside each block was already changed, increased or decreased, during the embedding phase. Therefore, this indicator of homogeneity level is used to extract the embedded data, making the extraction is blind. The value of ꚍ is chosen between 0 and 1 to examine the degree of change in the homogeneity of the region being analyzed. Consequently, when the value of $\mathcal{D}$  exceeds ꚍ, this difference indicates that there is a change has been made in this region, which can be utilized to extract a bit with a value of 1; otherwise, a bit with a value of 0 is extracted.The idea behind using a weight value of -1 in the center location of the kernel is to measure the disparity between the center point and the neighboring points by giving more importance to the center point where the logo data is assumed to be included. The coefficient values in the kernel in Figure 2 sum to zero, so they would give a response of zero in areas that contain a magnitude similar to its neighbors, which means a 0 bit of logo image. Otherwise, a 1 bit logo image is obtained.

The calculation of $\mathcal{D}$ using the kernel in Figure 2 is a linear operation and is implemented using convolution. This means that the extraction stage using the proposed kernel operators is a lightweight process since the computational burden is linear.