A Hybrid SVD-Based Image Watermarking Scheme Utilizing Both U and V Orthogonal Vectors for Robustness and Imperceptibility

SVD-based watermarking algorithm is one of the most preferable algorithms for copyright protection due to its singular values (SVs) that have outstanding stability and represents intrinsic algebraic image properties. Hence, there is a good trade-off between robustness and imperceptibility. However, most SVD-based algorithms have been tested against conventional attacks, such as image manipulations, that do not fully exploit adversary’s knowledge. These algorithms are vulnerable to false-positive problem, where an adversary’s watermark can be detected in the watermarked image although it was never inserted. The underlying problem is due to the strong influence of <inline-formula> <tex-math notation="LaTeX">$U$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$V$ </tex-math></inline-formula> orthogonal vectors of SVD on an image. In order to solve false-positive problem, <inline-formula> <tex-math notation="LaTeX">$UV$ </tex-math></inline-formula> can be used to embed the watermark together with SVs. However, this solution is not ideal as <inline-formula> <tex-math notation="LaTeX">$U$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$V$ </tex-math></inline-formula> hold important structural information of an image and is hypersensitivity to even a little change in <inline-formula> <tex-math notation="LaTeX">$UV$ </tex-math></inline-formula> vectors. Therefore, this research has the objectives to analyse the robustness of existing SVD-based watermarking algorithms that are using orthogonal vectors and then propose a new robust algorithm that is able to solve false-positive problem and sensitivity issued caused by scaling factor. Hence, a new transform domain image watermarking scheme that utilized both <inline-formula> <tex-math notation="LaTeX">$U$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$V$ </tex-math></inline-formula> orthogonal vectors is proposed with the usage of Human Visual System (HVS), Discrete Wavelet Transform (DWT) and SVD. Experimental results showed that the proposed scheme is more robust against majority types of image processing and geometrical attacks compared to existing schemes while achieving good quality watermarked image level. The significance of new algorithm comes at the right time during the Covid-19 epidemic as organizations involving in business and financial services can be assured of the integrity of its downloadable/streamable/shareable digital files, which are copyrighted through the unique SVD and robustness features of the algorithm ensuring piracy prevention of their content.


I. INTRODUCTION
Digital data especially digital media such as images and videos can be seen everywhere now as digital technologies being part of our lives. As digital media is easily accessible and obtainable by everyone on the online platform, ownership The associate editor coordinating the review of this manuscript and approving it for publication was Mehul S. Raval . or copyright of the media itself is quite difficult to be proven or protected. Hence, researchers came across techniques such as image watermarking scheme allowing media owners to prove their ownership and protect their copyright. Image watermarking scheme fulfilled it's purpose by embedding a watermark that belongs to the owner into the media in which the watermark is invisible on the media and can be extracted after attacks by an adversary. There are several types of image watermarking techniques proposed by researchers to ensure copyright protection such as digital image watermarking [1], [2] and optical image watermarking [3], [4], [5]. This research focuses on digital image watermarking and one of the techniques that is used by many researchers in the area of digital watermarking scheme is the singular value decomposition (SVD) [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22].
Image watermarking scheme can be defined based on two working domains, 1) Spatial domain and 2) Transform domain. Spatial domain watermarking schemes embed the watermark into the pixels of host image and is known by it's advantages of being fast, simple and can handle large capacity. Watermark can be embedded several times in the host image to decrease the probability of being destroyed after attacks. However, spatial-based watermarking schemes have low-resistance against non-geometrical attacks and the embedding steps are well-known by adversary which lower the security of the scheme. In the other hand, transform domain watermarking scheme generate frequency coefficients from the host image using transformation method. Watermark is embedded by modifying the frequency coefficients and applying small modification on the frequency coefficients has less probability to be detected by Human Visual System (HVS). Several transformation methods are being used in existing watermarking schemes such as Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT), Redundant Discrete Wavelet Transform (RDWT) and Integer Wavelet Transform (IWT).
Watermark extraction process can also be used to define an image watermarking scheme. There are three types of watermarking scheme, 1) Blind, 2) Non-blind and, 3) Semi Blind. Blind watermarking scheme is favourable as it only requires the watermarked image and a key to extract the watermark out from the image. Non-blind scheme requires an additional component for example, original host image or watermark image apart from the watermarked image and key for watermark extraction where semi-blind scheme uses a key and the original watermark image for extraction.
There are several watermarking methods used to embed the watermark into a host image such as singular value matrix watermarking (SVWM), direct watermarking (DW), singular value and comparison (SVC), comparison and threshold (CT), principle components (PC) and singular vectors (U or V). Methods that are able to avoid false-positive problem is prior to researchers and only three methods are able to do that which is CT, PC and U or V. This is because the methods include the use of U matrix which contains geometrical information of an image preventing the adversary on extracting the watermark.
Scale factor plays an important role in image watermarking because it determines the embedding trade-offs. A single scaling factor (SSF) can lead to a stable level of embedding, but cannot fulfil the desired balance between robustness and imperceptibility. A small SSF value can lead to high imperceptibility and low robustness, vice versa. Researchers have used multiple scaling factors (MSF) to achieve the desired goals of robustness and imperceptibility. Optimal MSF values can be determined by using optimization algorithms such as Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Multi-Objective Ant Colony Optimization (MOACO) and Differential Evolution (DE). MSF-based optimization techniques achieve all requirements of a good image watermarking scheme. However, the use of optimization techniques are computationally expensive, and there is a limited range of suitable MSF values for each watermarking scheme.

II. RELATED WORK
In this section, algorithm of existing schemes [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22] that embeds the watermark into either U or V orthogonal matrix are analyzed and explained. The schemes analyzed in this paper does not affected by false-positive problem (FPP) as FPP is solved by utilizing the U component instead of S component of the host image. To allow comparison to be made easier, Table 1 summarized the algorithm used by each scheme.
Chang et al. [12] proposed an image watermarking scheme that explores the U component and S component. The algorithm proposed by Chang et al. first partitioned the host image into 8 × 8 non-overlapping blocks. Then, the non-zero coefficients of the S component of each block is calculated as the authors stated that the number of non-zero coefficients in the S component can be used to determine the complexity of each block whereas the blocks with the most complexity is more favourable for watermark embedding. Pseudo-random number generator (PRNG) which is able to increase the security is then being used to select a set of high complexity blocks and the watermark is embedded into the host image by modifying the second and third coefficients of the first column of U matrix as Chang et al. discovered that the relationship between the first column of U matrix will be preserved after general image processing. The experiments performed by the authors show that the proposed algorithm has good performance in both robustness and security.
Two notes are presented by Chung et al. [13] as a guideline to modify the coefficients in U component and in V component for the watermarking schemes that embeds the watermark in the U component or V component to increase the imperceptibility and the embedding capacity. The notes presented are, 1) Modification of coefficients in the column vector of U component will results in less visible distortion; and 2) Modification of coefficients in the row vector of V component will results in less visible distortion. The authors then suggested that modifying the column vector of U component and row vector of V component simultaneously can increase the imperceptibility of the watermarked content and embedding capacity in which their suggestion was proven by the experiments performed. VOLUME 11, 2023  Later the year, Fan et al. [14] also proposed two notes to enhance the invisibility and robustness of a SVD-based watermarking scheme. The authors stated that all the column vectors of U component and V components can contribute towards robust watermarking and it is known that the magnitude relationship of the first column of the U component and V component can be preserved and are the most stable among other coefficients. Hence, the first note proposed was: 51020 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.
''Only the coefficients in the first column of U component and V component may be modified for robust SVD-based watermarking''. Then, the second note presented was to use U or V component to embed the watermark bit and compensate the visible distortion by modifying V or U component. The authors came to a conclusion where the proposed notes improved the peak signal-to-noise ratio (PSNR) greatly and as compensation can be made to lower the visible distortion, higher threshold can be used in which will enhance the robustness of the watermarking scheme.
Lai [15] proposed an improved SVD-based image watermarking scheme that used HVS instead of PRNG. The author explained that PRNG selects blocks based on their ranks and blocks with higher ranks will be chose to embed the watermark which will lower the distortion level of the watermarked content whereas the implementation of HVS characteristics helps to select embedding regions based on the visual entropy and edge entropy of the host image to balance between robustness and imperceptibility of watermarked image. Lai's algorithm starts with partitioning the host image into 8 × 8 non-overlapping blocks. Then, DCT and SVD are applied to the blocks that were selected using HVS characteristics. The watermark is embedded in the third and fourth element of the first column of the U matrix. The author concluded that the proposed scheme shows significant result towards the robustness and imperceptibility.
A blind color image watermarking scheme based on SVD was proposed by Su et al. [16] which succeeds in producing a watermarked image that has high imperceptibility and is robust against various signal processing operation and geometrical attacks. Firstly, pre-processing of watermark image is performed by dividing the 3D watermark image into 3 components (Red, Green and Blue) using dimension-reduction treatment and each 2D component is then permuted by Arnold transformation and is converted into binary sequence. Next, the color host image is split into Red, Green and Blue components and each component is then partitioned into 4 × 4 non-overlapping blocks. SVD is applied on each blocks and the watermark is embedded by modifying the second and third row of the first column of the U matrix.
Makbol et al. [17] proposed to use DWT in their SVD image watermarking scheme using HVS as DWT has less computational time compare to DCT and is more accurate in the modelling aspects of HVS. Host image is first partitioned into 8 × 8 non-overlapping blocks and the HVS characteristics are applied to select the low information blocks which are the best regions for watermark embedding. The coordinates of the selected blocks which will be used in the extraction process were encrypted by AES-192 to ensure security of the scheme. After the blocks are selected, DWT is applied on each block and 4 sub-bands which are LL, LH, HL and HH were obtained. SVD is applied on the LL sub-band and the watermark is embedded by examining the second and third element of the first row of the U matrix. The authors mentioned that the second and third element is modified because better performance was achieved instead of the modifying the third and fourth element as proposed by Lai [15]. The scheme proposed by Makbol's et al. outperforms several other existing schemes and results with a watermarked image that has high imperceptibility and strong robustness.
Another image watermarking scheme was proposed by Makbol et al. [18] and unlike the previous scheme, MOACO were used in order to obtain optimal scaling factor for balancing the imperceptibility and robustness. The scheme first perform IWT on the host image to obtain LL, LH, HL and HH sub-bands and SVD is applied to the LL sub-band. Next, the watermark also undergo SVD and a new S matrix of the LL sub-band is calculated by adding the multiplication of the MZF value and U components of the watermark with the original S matrix of the LL sub-band. New LL sub-band was obtained by performing inverse SVD and watermarked image is obtained after inverse IWT is applied. The MZF value is chosen by MOACO which is the optimal scaling factor that allow the watermarked image to have high imperceptibility and strong robustness against attacks. Hence, with the help of MOACO, the scheme exibits high resistance against attacks and high imperceptibility.
On the other hand, the authors in [19] proposed to use DCT instead of DWT because DCT achieved less computational cost as opposed to the claims made by Makbol et al. [17]. Host image is first divided into 4 × 4 non-overlapping blocks and the modified entropy of each block is calculated and sorted in ascending order. The blocks with lower modified entropy value are selected and are transformed by DCT. Application of SVD on transformed blocks are performed and watermark is then embedded by modifying the second and third element of the first column of the U matrix. To increase the security of the scheme, the watermark image is encrypted by performing logical XOR operation using a private key. The authors claims that the proposed scheme produced high watermarked image quality and offers more robustness as compared to Lai's [15] image watermarking scheme.
Ernawan and Kabir [20] proposed a watermarking scheme that uses RDWT transformation method. RDWT was selected because the elimination of up-sampling and down-sampling of coefficients that appears in DWT during filter-bank iterations. 8 × 8 pixels are obtained after the division of host image and the blocks will the lowest modified entropy value were selected after calculation. Then, the binary watermark is scramble using Arnold chaotic map which can improve the security and confidentiality of the watermark. The selected blocks were transformed by RDWT and four sub-bands were obtained (LL, LH, HL and HH). SVD is applied on the LL sub-band and modification is made on the third and fourth coefficients in the first column of U component. Test results show that the proposed scheme produced less distortion and better perceptual quality which also has higher robustness compared to other image watermarking scheme.
The authors in [20] also proposed another watermarking scheme that uses RDWT transformation and Arnold chaotic map [21] which exhibits better performance against some attacks compare to the previous proposed watermarking VOLUME 11, 2023 scheme. Instead of dividing the host image to 8 × 8 pixels, this scheme chose to partitioned the host image into 4 × 4 non-overlapping blocks. The scrambled watermark is embedded into the host image by performing modification in the second and third row of the first column of U matrix which is different from the previous scheme.
Sajjad et al. [22] proposed an intelligent and blind image watermarking scheme based on DWT and SVD using HVS and PSO. The host image is first undergoing DWT to obtain the LL, LH, HL and HH sub-bands and the LL sub-band is then partitioned into 4 × 4 non-overlapping blocks. HVS is then used to select the blocks for watermark embedding. Later, SVD is applied on the selected blocks and the watermark is embedded into the second and third element of the first column of both U and V matrices. The watermarking scheme is said to be intelligent because the embedding threshold was chosen by PSO which allows the watermarked image to have balance between imperceptibility and robustness. The watermarking scheme also has an increase in the embedding capacity as embedding in both U and V matrices allow more data to be embedded.
To conclude the analysis of existing schemes, the security criteria of an image watermarking scheme can be enhanced by encrypting the watermark image and the locations of embedded watermark. However, this security enhancement technique does not directly solve the false-positive issue which is a big faulty of an image watermarking scheme. From the existing schemes, we can see that the schemes [12], [15], [16], [17], [18], [19], [20], [21] utilized the U component of the host image which holds the geometrical information to avoid false-positive problem. Having said that, the V component which holds the same information can also be used to embed the watermark like scheme [22]. The usage of U and V components to embed the watermark is able to reduce the space/blocks used for embedding, hence increase the embedding capacity. Besides that, most existing schemes do not use an optimization algorithm to determine the optimal scaling factor that is able to balance between robustness and imperceptibility. Although the usage of optimization algorithm is computationally expensive, it is able to help the image watermarking scheme to search the suitable scaling factor for different host image and different watermark.
Based on the result of analysis of existing schemes, this research has the aim to propose an image watermarking scheme that utilizes both the U and V orthogonal vectors to generate a watermarked image that is able to avoid false-positive problem with enhanced robustness and imperceptibility. The techniques implemented in the proposed scheme such as Arnold transformation (AT), Human Visual System (HVS), Discrete Wavelet Transform (DWT) and Singular Value Decomposition (SVD) are further explained in Section III. The embedding and extraction processes of the proposed scheme are then illustrated in Section IV. This is followed by the experimental results of benchmarking the proposed scheme with other existing schemes in Section V.
Last but not least, Section VI concludes the research on the proposed scheme.

III. PRELIMINARIES
This section will show brief mathematical explanations of the techniques used in the proposed scheme such as Arnold Transform (AT), Human Visual System (HVS), Discrete Wavelet Transform (DWT) and Singular Value Decomposition (SVD).

A. ARNOLD TRANSFORM (AT)
Arnold transformation is a simple and efficient chaos system that is popularly used in image encryption area. It modified the original image by repeatedly changing the image pixels position to generate a chaotic image with the purpose of hiding original image information. This technique is able to enhance the image security and increase the image's robustness towards illegal attacks [23], [24]. The two-dimensional pixel-scramble tool can be defined as follow: where x y and x ′ y ′ represents the vector position before and after AT shifting, respectively and mod in the other hand denotes the modulus operation with the divisor N in which N is the size of the image to be scrambled [24]. As AT is used to encrypt the watermark image, hence, an inverse equation must be used to decrypt the scrambled image to recover the watermark image [20], [21]. The inverse Arnold transformation equation can be calculated as, A secret key, k, will be used as the number of scrambling iterations to further scramble the image for k times as the transformation in (1) only scrambles the image once, hence, the watermark image will be further secured.

B. HUMAN VISUAL SYSTEM (HVS)
Human Visual System model helps image watermarking schemes to further obtain optimal balance between robustness and imperceptibility. Watermarked image with no visible distortions can be generated if the watermarking algorithm adopts the characteristics of HVS. Visual entropy and edge entropy are the characteristics of HVS model and both the attributes helped in selecting significant embedding regions which also known as the region of interest as the regions selected holds significant information about the host image. Visual entropy quantifies the texture content of host image and can be used to calculate the spatial correlation of neighbouring pixels [15], [17], [22]. The visual entropy of an image can be defined using the following mathematical formula 51022 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.
according to Shannon's definition: p i represents the occurrence probability of an event i with 0 ≤ p ≤ 1 and n i=1 p i = 1. Visual entropy is considered as a global measure of the image as the value calculated strongly relies on the probability distribution of pixel intensities and did not take co-occurrence of the pixel values in the formula [17].
Edge entropy which contains the edge information of an image is introduced with the aim to obtain a better entropy value compared to only using visual entropy. A more suitable embedding regions can be chosen with the help of visual entropy and edge entropy [15], [22]. The calculation of edge entropy is shown as follows: where u i = 1 − p i refers to the ignorance or uncertainty of the pixel value.

C. DISCRETE WAVELET TRANSFORM (DWT)
DWT is a mathematical tool based on time which is able to point out the local and global image elements providing efficiency for image processing. DWT decomposes the host image to produce high frequency and low frequency coefficients by passing the image through a sequence of high-pass and low-pass filter [1].
To explain the DWT decomposition process, x[n], which is the signal with the frequency range of 0 to p rad/s passes through a high-pass filter (g[n]) and a low-pass filter (h[n]) leads to the removal of half the samplings as in accordance to Nyquist's rule. The signal is sub-sampled by 2 as it only has a ceiling frequency of π/2 radians instead of π radians as a result after the filtering process. This process is expressed as follows: and where y high and y low are the outputs of the high-pass and lowpass filters after sub-sampling by 2 [17]. DWT decomposes the host image into four sub-bands, LL, LH, HL, and HH as shown in Figure 1, where each subband's size is a quarter of the host image's size holds different information about the original image.
• An illustration of the host image, LL sub-band which passed through low-pass filter twice during decomposition has the highest coefficients and contains majority of the host image's information [22].
• LH sub-band passed through low-pass filter followed by high-pass filter holds the horizontal image's information whereas HL sub-band holds vertical information passed through high-pass filter before low-pass filter.
• HH sub-band, in the other hand holds the diagonal information of the host image passed through high-pass filter twice during decomposition. LL sub-band is chosen for watermark embedding to achieve the desired balance between robustness and imperceptibility because it allows the watermarked image to resist various attacks such as geometric distortions [1].

D. SINGULAR VALUE DECOMPOSITION (SVD)
SVD is a numerical tool that is famous for its good stability property. In image processing, SVD can be applied to an image, I , which can be denoted as a real number matrix resulting in three matrices, U , S and V . Among the three matrices, U and V are orthogonal matrices that holds geometric information about the host image which are highly sensitive to modifications. The S matrix which in charge of the luminance of the host image is a diagonal matrix that consists of descending positive singular values [17], [18]. SVD process can be defined as follows:

IV. PROPOSED ALGORITHM
This section explains the algorithm used to embed the watermark into the host image and also the watermark extraction process done from the watermarked image.

A. EMBEDDING PROCESS
The security of the watermark can be enhanced by applying Arnold transformation to scramble the binary watermark image and by encrypting the coordinates of the HVS selected blocks which will be used for extracting the watermark correctly.
The watermark embedding procedure are as follows: 1) Host image of size 512 × 512 is converted to grayscale image and 32 × 32 watermark image is binarized into a binary matrix. 2) Arnold transformation is applied on the binarized watermark image with a secret key, k.  5) DWT transformation is applied to the selected blocks to generate LL, LH, HL, and HH sub-bands. 6) LL sub-band is selected for application of SVD. 7) Modification of coefficients in U and V matrices are made to embed the binary watermark. Modifications are made according to the following conditions: • Odd number column of binary watermark is embedded into the U matrix.
• Even number column of binary watermark is embedded into the V matrix.

B. EXTRACTION PROCESS
The watermark extracting procedure are as follows: 1) Watermarked image is partitioned into 8 × 8 nonoverlapping blocks. 2) Load the recorded coordinates of the 512 blocks that were selected by the HVS characteristics and decrypt the file. 3) DWT transformation is applied to the 512 blocks. 4) LL sub-band undergoes SVD. 5) Examination of coefficients in U and V matrices to extract the binary watermark according to the following conditions: • Extracted binary bit from the U matrix is the odd column of the binary watermark.
• Extracted binary bit from the V matrix is the even column of the binary watermark.
• If |matrix 2.1 | is larger than |matrix 3,1 |, the extracted binary bit is 1; else if |matrix 3,1 | is larger than |matrix 2,1 |, the extracted binary bit is 0. 6) Inverse Arnold transformation is applied to the extracted binary matrix with the secret key, k used 51024 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. in embedding process to obtain the extracted binary watermark. The experimental results of applying the proposed watermarking algorithm will be explained in details in the next section.

V. RESULTS AND DISCUSSIONS
This section will explain the experiment and comparison results of proposed scheme with existing schemes [15], [17], [19], [20], [21], [22]. To provide fairness for comparison, grayscale images of Lena, m83 and Peppers with size 512 × 512 as shown in Figures 2(a), 2(b) and 2(c) are being used as the host images, whereas grayscale watermarks of Ju and Curtin Logo with size 32 × 32 as shown in Figures 2(d) and 2(e) are used for embedding. In addition, the grayscale SB watermark with size 32 × 32 as shown in Figure 2(f) is only being used to embed in the host image, Lena for the comparison with Sajjad's scheme [22]. Other than using the same image, same scaling factor of T = 0.4, T = 0.2, T = 0.012 and T = 0.002 will be used in each existing watermarking scheme as well as the proposed scheme. However, for Sajjad's watermarking scheme [22] which uses optimization algorithm to define the optimal scaling factor, the threshold used in Sajjad's scheme which is T = (21.73, 17.73) will be used in the proposed scheme for fair comparison but as the matrix modification algorithm is different, the scaling factor will be divided by 1000.
To evaluate the robustness of each scheme, various types of image processing and geometrical attacks are practiced on the generated watermarked images such as Gamma Cor-  Table 2. Peak signal-to-noise (PSNR) and normalized correlation (NC) are the two main methods used to determine the imperceptibility and robustness aspects of a watermarking scheme.

A. IMPERCEPTIBILITY ANALYSIS
Peak signal-to-noise (PSNR) is used to define the imperceptibility of the watermarking scheme and can be defined as, where max(I (i, j)) is the largest pixel value in the host image, while the mean square error (MSE) between the host image, I and the watermarked image, I W can be calculated as, where M and N are the rows and columns number of an image. The relationship between the PSNR value and the imperceptibility aspect of a watermarked image is: The higher the PSNR value, the lesser the difference between the host image, I and watermarked image, I W . According to Naffouti et al. [11], a watermarked image with good quality has the PSNR value between 35dB to 48dB, whereas if a watermarked image has the PSNR value of more than 48dB, it means that the image has excellent quality and no difference can be spotted between the watermarked image and original host image. However, if the watermarked image has the quality ranging between 29dB to 35dB, it is considered as acceptable image quality. Hence, to ensure the watermarked image's quality for the imperceptibility test, the PSNR value of the watermarked image generated by the existing schemes and the proposed scheme should be more than or equal to 35dB.
watermark are tabulated in Tables 3, 4 and 5 respectively, whereas the results of using Curtin Logo watermark are tabulated in Tables 6, 7 and 8. As for the scheme in [22], PSO is used to obtain the optimal MZF value. Hence, result of experiments recorded by the authors will be used to conduct the comparison in which the scaling factor used for U and V components are 21.23 and 17.73, respectively. As for the comparison, the proposed scheme used the same scaling factor used in [22] but due to the difference in the matrix modification algorithm, the scaling factor is divided by 1000.  [17] and Ernawan's [19], [20], [21] schemes with different threshold, T using m83 as host image and Curtin Logo as watermark image.  [19], [20], [21] schemes with different threshold, T using Peppers as host image and Curtin Logo as watermark image.  Tables 3, 4, 5, 6, 7 and 8, we can see that the scheme [21] has the highest PSNR value no matter the value of threshold used. As in Table 9, our proposed scheme with the usage of scaling factor (0.02123, 0.01773) obtained the PSNR of 46.9467 dB, 47.0675 dB and 46.9359 dB with the watermark Ju, Curtin Logo and SB, respectively. However, the scheme [22] having the imperceptibility target of PSNR = 43 dB obtained the average PSNR of 43.3281 dB according to the results recorded by Sajjad et al. [22]. However, all the proposed scheme and existing schemes [15], [17], [19], [20], [21], [22] have the PSNR results of good quality image and above, according to Naffouti et al. [11]. In other words, the PSNR results passed the imperceptibility criteria of an image watermarking scheme.

B. ROBUSTNESS ANALYSIS
Normalized correlation (NC) is a method used to measure the difference between an extracted watermark, W new and the original watermark, W to determine the robustness of an image watermarking scheme. When the original and extracted images are nearly similar to each other, the NC value will be close to 1 (NC ≈ 1). In case where NC = 1, the original watermark and extracted watermark are the same as 51026 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.  [15], [17], [19], [20], [21] after image processing attack and geometrical attack using threshold, T = 0.04 using Lena as host image and Ju as watermark image.  [15], [17], [19], [20], [21] after image processing attack and geometrical attack using threshold, T = 0.04 using m83 as host image and Ju as watermark image. each other. NC can be calculated as in (11), as shown at the bottom of the page, where µ 1 and µ 2 refer to the mean values of W and W new .
and Peppers, respectively. The threshold of 0.04 is chosen for comparison because the usage of high scaling factor will result in stronger robustness. The results of NC value after performing different attacks on the proposed scheme with different watermark images (Ju, Curtin Logo and SB) embedded and Sajjad's scheme [22] (as recorded in paper [22]) are shown in Table 16. The best recorded NC value among the results is indicated in bold.
Based on the results shown in Table 16, the proposed scheme is more robust than the scheme proposed by Sajjad [22]. On the other hand, the proposed scheme outperformed other existing schemes [15], [17], [19], [20], [21] against different attacks according to the results recorded in Table 10, 11, 12, 13, 14, and 15 that used different host images and different watermark images.

C. EMBEDDING CAPACITY ANALYSIS
Embedding capacity is another criteria to evaluate an image watermarking scheme. Embedding capacity of an image watermarking scheme is defined as the amount of data or payload that can be embedded into the host image and is highly related to the size of watermark and the blocks' size that  [15], [17], [19], [20], [21] after image processing attack and geometrical attack using threshold, T = 0.04 using m83 as host image and Curtin Logo as watermark image.  [15], [17], [19], [20], [21] after image processing attack and geometrical attack using threshold, T = 0.04 using Peppers as host image and Curtin Logo as watermark image.
the host image is partitioned into. The embedding capacity of the image watermarking scheme can be analyzed through the embedding process and can be calculated by using the following equation: MEC = (Size host /Size block ) 2 (12) where MEC is the Maximum Embedding Capacity of an image watermarking scheme, Size host is the host image's size and Size block is the block size that the host image is being partitioned into. Table 17 showed the comparison of embedding capacity between the proposed scheme and existing schemes [15], [17], [19], [20], [21], [22]. The proposed scheme and Sajjad's scheme [22] are able to double the embedding capacity of existing image watermarking schemes [15], [17], [20] that partitioned the host image into 8 × 8 blocks as the proposed scheme and Sajjad's scheme [22] embeds the watermark in both the orthogonal vectors, U matrix and V matrix, whereas the other existing schemes only embed in the U matrix. For the schemes [19], [21], the host image was partitioned into smaller blocks (4 × 4) resulting in larger MEC. However ,  TABLE 16. NC results on proposed scheme with the scaling factor (0.02123, 0.01773) and different watermark and scheme [22] with the scaling factor (21.23, 17.73) recorded by Sajjad et al. [22] after image processing attack and geometrical attack.  [15], [17], [19], [20], [21], [22] with host image's size 512 × 512.
if the schemes [19], [21] were modified to partition the host image into 8 × 8 blocks, the MEC of both the schemes are just 4096 which is half the MEC of the proposed scheme and Sajjad's scheme [22].

VI. CONCLUSION
As we are using HVS to find the favourable blocks to embed the watermark, three transformation methods are being compared in this paper, namely DCT, DWT and RDWT. Experimental results showed that DWT performs better with the usage of HVS. The watermarking schemes that are compared in this paper avoided false-positive problem because the watermark is embedded in the U component of the host image which contains the geometrical information instead of S component which only controls the luminance of the image.
As a conclusion, the hybrid DWT-SVD scheme using HVS such as our proposed scheme works the best among other existing schemes as the proposed scheme showed strong robustness against many different attacks which ensures the ownership and copyright protection of the media. Although the proposed scheme did not perform well in the imperceptibility test compared to [15], [17], [19], [20], and [21], the proposed scheme still achieved the PSNR value of 40.8396 dB, which is in the good quality level according to Naffouti et al. [11]. On the other hand, the proposed scheme is able to double the embedding capacity compared to other watermarking schemes [15], [17], [19], [20], [21]. Furthermore, in comparison with Sajjad's scheme [22] that embeds watermark into both U and V orthogonal vectors as well, the results showed that the proposed scheme outperformed Sajjad's scheme [22] in both imperceptibility and robustness tests. As for the embedding capacity, both schemes achieved the same maximum embedding capacity.

VII. FUTURE WORK
Future research will be done to further enhance the proposed image watermarking scheme by selecting a suitable optimization algorithm in order to generate a watermarked image with enhanced and balanced imperceptibility and robustness.