Optik - International Journal for Light and Electron Optics 269 (2022) 169899

Contents lists available at ScienceDirect

Optik
journal homepage: www.elsevier.com/locate/ijleo

A double-color image watermarking algorithm based on

quaternion Schur decomposition
Yehan Sun, Qingtang Su *, Siyu Chen, Xueting Zhang
School of Information and Electrical Engineering, Ludong University, Yantai 264025, China

A R T I C L E I N F O A B S T R A C T

Keywords: In view of the fact that many watermarking algorithms used for copyright protection have
Watermarking ignored the correlation and synchronization between the color channels, a new double-color
Quaternion Schur decomposition image watermarking algorithm based on quaternion Schur (QSchur) decomposition is pre­
Structure-preserving algorithm
sented in this paper. Firstly, the color watermark image is encrypted by two-dimensional com­
Compound chaotic map
pound chaotic map; then, the QSchur decomposition is used to associate and transform multiple
color channels; finally, the watermark is embedded and extracted blindly by using the high
correlation between the coefficient pairs in the block. Due to the high time complexity of the
quaternion, a real structure-preserving algorithm is used to shorten the running time of the
proposed algorithm by about 20 times compared with the quaternion toolbox of Matlab. Simu­
lation experiments and analysis demonstrate that this algorithm has stronger robustness and
higher security when the actual embedding rate and invisibility are the same as those of other
algorithms given.

1. Introduction

The background of the digital watermarking, some related works of the image watermarking algorithm and a new watermarking
algorithm are introduced successively in this section.

1.1. Background

With the development of today’s society, people can access digital information easily via the Internet. But meanwhile, piracy is
getting easier and easier, and the copyright of digital products has become an urgent issue for the industry [1]. In consequence,
watermarking technology has a wide range of application prospects.
Digital watermarking is a security technology that hides data through embedding and extraction operations [2], and its basic idea is
to hide the owner mark in the form of watermark in the original digital product and then transmit it. At the receiving end, the owner
mark is extracted by computer and used for different purposes, including content authentication, digital signature, secret commu­
nication and so on [3,4].
Color image digital watermarking has always been a research hotspot [43], because if the binary image [44] or gray image [5] as
the copyright mark, they not only carry less effective information, but also do not meet the public’s visual requirements for beauty,

* Corresponding author.
E-mail address: [email protected] (Q. Su).

https://doi.org/10.1016/j.ijleo.2022.169899
Received 17 May 2022; Received in revised form 12 August 2022; Accepted 28 August 2022
Available online 2 September 2022
0030-4026/© 2022 Elsevier GmbH. All rights reserved.
Y. Sun et al. Optik 269 (2022) 169899

which makes them subject to many restrictions in practical applications. For example, in algorithm [44], Hosny et al. constructed a
robust color image watermarking algorithm by using highly accurate fractional-order multi-channel orthogonal exponent moments,
and embedded the binary watermark through quantization processing, which had good visual imperceptibility and strong robustness.
Digital watermarking can be divided into spatial-domain watermarking [6] and frequency-domain watermarking [7–17]. The
spatial-domain watermarking is to directly modify pixel values of the carrier image, which has a greater impact on the visual effect of
the carrier image, but its computational complexity is small; the frequency-domain watermarking is to transform the carrier image into
frequency domain through some algorithms, and then calculate the transform coefficient, so its robustness is good, but its compu­
tational complexity is high. Since the robustness is a very important and critical factor for copyright protection, we decide to use the
frequency-domain watermarking. In addition, there are zero-watermarking techniques that do not modify the original image data. For
example, in algorithm [45], Hosny et al. proposed a new geometrically invariant multiple zero-watermarking algorithm, and the
feature images of original medical images were constructed by a new set of multi-channels shifted Gegenbauer moments of fractional
Orders and then a zero-watermark was obtained by XOR operation with the watermark, which was robust against various standard
attacks.

1.2. Related works

In frequency-domain watermarking, transform algorithms commonly used include discrete Fourier transform (DFT) [7], discrete
cosine transform (DCT) [8], discrete wavelet transform (DWT) [9], other transforms [10,11], matrix decomposition (such as Schur
decomposition [12,13], QR decomposition (QRD) [14], Hessenberg decomposition [15], singular value decomposition (SVD) [16] and
eigenvalue decomposition (EVD) [17]), and so on. In scheme [7], Cedillo-Hernandez et al. put forward a DFT-based watermarking
scheme. For solving one problem of manually adjusting key parameters, this scheme improves DFT and adopts a variety of optimi­
zation techniques, so that the key parameters can be automatically adjusted, which not only makes the invisibility higher, but also
makes robustness stronger. In algorithm [8], Roy et al. came up with a DCT-based watermarking algorithm. It mainly uses repeated
codes to enhance the robustness, and the intermediate significant AC (Alternating-Current) coefficients are modified to maintain
invisibility. In algorithm [9], Wang et al. presented a Haar wavelet-based watermarking algorithm. In order not to affect the quality of
the image and to guarantee the robustness, this algorithm embeds the watermark into mid and low frequency band, which not only
ensures the high visibility, but also guarantees the ability to resist some general attacks. In algorithm [10], Luo et al. put forward a
watermarking algorithm based on feature image extraction. Because of the strong correlation between the color channels of high
dynamic range images, this algorithm first segments the color channels into multiple feature maps through Tucker decomposition, and
then carries out Schur decomposition on blocks obtained by dividing the first feature map to embed watermark. In algorithm [12], Liu
et al. came up with a watermarking algorithm using Schur decomposition. Authors use affine transform for encryption and the key
space increases greatly. In addition, the time complexity of Schur decomposition is low, so the operation efficiency of this algorithm is
relatively high. In algorithm [14], Su et al. presented a QRD-based watermarking algorithm. Its main highlight is that the watermark
information is encoded into ternary information, which greatly improves the maximum watermark embedding capacity. In algorithm
[15], Su et al. presented a watermarking algorithm using upper Hessenberg matrix. As can be seen from the experimental results, the
total time of this algorithm is less than one second, so its real-time performance is very good. In algorithm [16], Hu et al. presented a
SVD-based watermarking algorithm. This algorithm consisting of six modules and each module plays a different role to overcome the
shortcomings of existing SVD-based watermarking algorithms. Through the authors’ in-depth and thorough analysis of SVD, the
invisibility and robustness are improved. In algorithm [17], Liu et al. presented an EVD-based watermarking algorithm. The water­
mark is embedded by quantizing the sum of the absolute values of eigenvalues and image correction is added into this algorithm.
Experimental results show that this algorithm performs very well in most respects.
However, most of the above references only studied the single color channel of the color image, discarding the relative relationship
between color channels. Moreover, the color image needs to be converted into the grayscale image in many algorithms, so the cor­
relation between color channels cannot be captured [18,19]. According to the references [18–25], we know that the image processing
based on the quaternion can deal with the above problem, because it takes a color pixel that vertically contains three color channels as
a vector, which can process the color image as a whole. In this way, color channels are successfully integrated, and the disadvantages of
lack of correlation between color channels and un-synchronization of watermark embedding are overcome. Therefore, in recent years,
image representation and application based on the quaternion have become more and more important, and some algorithms based on
the quaternion have been applied to several image processing fields, such as quaternion QR decomposition (QQRD) [20], quaternion
discrete fractional Krawtchouk transform [26], quaternion discrete cosine transform [27] and other quaternion-based algorithms [28,
39–42]. In algorithm [26], Liu et al. extended the traditional discrete fractional Krawtchouk transform with quaternion theory to
quaternion discrete fractional Krawtchouk transform. The effectiveness and efficiency of the proposed method are proved by the
research on two applications of color image encryption and color image watermarking. In algorithm [39], Hosny et al. defined new
quaternion orthogonal shifted Gegenbauer moments (FrQSGMs) of fractional orders. The proposed FrQSGMs is used for color image
analysis and recognition, and it outperforms the existing orthogonal moments with fractional orders in several properties. In algorithm
[40], Hosny et al. proposed a new set of fractional orthogonal polar harmonic transforms, which were used for gray image and color
image analysis. The proposed fractional polar harmonic transforms and fractional quaternary polar harmonic transform are validated
by experiments and have excellent performance in accuracy and numerical stability, digital image reconstruction, robustness to noise,
computational efficiency and so on. In algorithm [41], Shao et al. proposed an encryption algorithm combining equal modulus
decomposition and quaternion gyrator transformation. The reliability and security of this algorithm are verified by experiments.
However, there is another problem at this time, that is, the computation complexity of the quaternion is very high. If it is applied to the

2
Y. Sun et al. Optik 269 (2022) 169899

watermarking algorithm, the real-time requirement in the era of fast information transmission is difficult to be met. For the calculation
of quaternion matrix decomposition, in addition to using the quaternion toolbox of Matlab [33], another way is real/complex
counterpart algorithms [35], which can not only convert quaternion operations into real/complex operations, but also take advantage
of the quaternion algebraic structure for the development of high-efficiency structure-preserving algorithms. To sum up, we come up
with a double-color image watermarking based on quaternion Schur (QSchur) decomposition, which is to embed the color watermark
image into the color carrier image according to the proposed rules. What’s more, according to the ability of digital watermarking to
resist external attacks, the proposed watermarking belongs to robust watermarking; according to different host image processing
methods, the proposed watermarking belongs to frequency-domain watermarking; according to different embedding objects, the
proposed watermarking belongs to image watermarking; according to different watermark extraction methods, the proposed water­
marking belongs to blind watermarking. Therefore, in general, the proposed watermarking belongs to the robust frequency-domain
image blind watermarking algorithm.

1.3. Main contributions

For the algorithm we proposed, firstly, the watermark image is encrypted by using the two-dimensional Logistic-adjusted-Sine map
(2D-LASM), and then embedded blocks are randomly selected and the dither noise is added to them. After that, the QSchur decom­
position is performed, and then the high correlation coefficients of the unitary quaternion matrix are modified according to the
watermark bit. Finally, after performing inverse QSchur decomposition, the watermarked pixel block is adjusted by fine-tuning
mechanism. According to the above ideas, the major contributions of proposed algorithm are as follows:

(1) QSchur decomposition is realized by the structure-preserving algorithm instead of the quaternion toolbox, which is nearly 20
times faster than using the quaternion toolbox;
(2) 2D-LASM, as an encryption algorithm with more chaotic structure, is applied in this algorithm, which more evenly encrypts the
watermark and enhances its security;
(3) For solving the problem that the watermark bit cannot be extracted correctly when all pixels are zeros in the embedded block,
the dither noise and pixel fine-tuning mechanism are added in this algorithm, which improves the robustness of the algorithm.

The remaining sections of this paper mainly include: Section 2 combs the main mathematical theories required by the proposed
algorithm; Section 3 describes the entire process of the proposed algorithm; Section 4 presents the experimental content, experimental
results and its analysis of the proposed algorithm; Section 5 gives a general summary of the proposed algorithm.

2. Preliminaries

In this section, the main mathematical theories involved in the proposed algorithm are introduced. The first one is two-dimensional
compound chaotic map, that is, 2D-LASM; the second one is the basic theory of the quaternion; the third one is the theorems and
implementation algorithm of QSchur decomposition.

2.1. Two-dimensional compound chaotic map

To enhance the security level of the proposed algorithm, 2D-LASM presented in [29] is used to make watermark images more
random. 2D-LASM has complicated characteristics such as ergodicity, unpredictability and sensitivity to initial states, and it also has a
wide range of values. 2D-LASM is obtained by combining Sine and Logistic maps [30], and its definition is as follows:
{
xi+1 = sin(πμ(yi + 3)xi (1 − xi ))
(1)
yi+1 = sin(πμ(xi+1 + 3)yi (1 − yi ))

where μ ∈ [0,1]; Sine map is defined as xi+1 = α sin(π xi ), α ∈ [0,1], xi ∈ [0, π] and this map is chaotic when α ∈ [0.87,1]; Logistic map is
defined as xi+1 = 4βxi (1 − xi ), β ∈ [0, 1], xi ∈ [0, 1] and this map is chaotic when β ∈ [0.8925, 1]. Firstly, through the factor of μ, the
logistic expression xi (1 − xi ) is scaled and taken as the input to the Sine map. Lastly, the phase plane is expanded from one-dimension to
two-dimension. It can be seen from Eq. (1) that two input pairs and their output pairs influence each other. Therefore, by comparing
the above maps, it can be found that 2D-LASM has a more complicated chaotic structure and its results are more difficult to predict. In
the proposed algorithm, initial states (x0 , y0 ) and μ are viewed as secret keys.

2.2. Quaternion representation of color images

Introduced by the mathematician Hamilton [34], the quaternion q can be expressed as Eq. (2):
q = w + xi + yj + zk (2)

where w, x, y and z belong to real numbers, i, j and k belong to imaginary numbers, and when w= 0, q is the form of pure quaternion.
Three imaginary numbers obey the rules in Eq. (3):

3
Y. Sun et al. Optik 269 (2022) 169899

i2 + j2 + k2 = − 1
(3)
ij = ji = k, jk = − kj = i; ki = − ik = j

where i, j and k correspond to the three channels of the color image, which are red, green, and blue respectively. Each image pixel of a
color image A can be represented using Eq. (4) as follows:
( ) ( ) ( ) ( )
Aq x, y = AR x, y i + AG x, y j + AB x, y k (4)

where (x, y) is the pixel coordinate, AR, AG and AB are the red, green and blue channels, respectively. Obviously, a color image can be
represented as a pure quaternion matrix [25].

2.3. Quaternion Schur decomposition

Theorem 2.3.1. (Real representation of quaternion matrix) When A∈Hm×m , two quaternion matrices will be generated by per­
forming QSchur decomposition, namely a quaternion Schur matrix V = V 0 +V 1 i +V 2 j +V 3 k and a unitary quaternion matrix U = U0 +
U1 i + U2 j + U3 k, so that A = U × V × U′ , where H represents quaternion skew-field and U′ = UT0 − UT1 i − UT2 j − UT3 k is the conjugate
transpose of U. For the quaternion matrix A = A0 + A1 i + A2 j + A3 k, where Ai∈Rm×n (i = 0, 1, 2, 3), its definition of real representation
can be descripted as follows:

⎡ ⎤
A0 − A1 − A2 − A3
⎢ A1 A0 − A3 A2 ⎥
A ≡⎢
R
⎣ A2
⎥ ∈ ℝ4m×4n (5)
A3 A0 − A1 ⎦
A3 − A2 A1 A0

where R represents real number field and AR is called JRS-symmetric [24]. The quaternion operations of a quaternion matrix can be
equivalent to the real-number operations of its real representation by using the real structure-preserving algorithm [21], which is
shown in following Algorithm 2.3.
Theorem 2.3.2. (Quaternion Householder based transform) A quaternion reflection is a matrix of form H1 =I − 2uu′ , where u∈Hn
y− x
and uu′ = 1. Suppose that x, y∈Hn , x ‡ y, ||x||= ||y|| and y′ x∈R, if u = ||y− x||, then H1 y = (I − 2uu )y = x, where ||.|| is unitary
′

invariant norm [22]. If y∈Hn isn’t a multiple of e1, choose x = αe1, where e1 is the first column of identity matrix In and a reflection
{ y1
− =0
||y||, ify1 ∕
mapping y to αe1 is provided by α = |y1 | . However, α is a quaternion in the matrix H1, so a new form H2 is obtained by
− ||y||, else
modifying the matrix H1 because it is handier for α to be a real number [23]. Suppose that y∈Hn , y ‡ 0 and y isn’t a multiple of e1,
{ y1
− ||y||, ify1 ∕=0 ( )
y− αe1
denote u = ||y− αe1 ||, where α = |y1 | . Denote αM = diag |αα|, In− 1 and H2 = αM H1 , then H2 maps y to |α|e1, where α is the
− ||y||, else
conjugate of α. As a matter of convenience, H1 can be rewritten as H1 = I − βuu′ , so.

H2 A = αM H1 A = αM (A − (βu)(u′ A)) (6)

is used for householder transforms of real matrix to reduce the amount of calculation.
Algorithm 2.3. (QSchur decomposition) For the given quaternion matrix A = A0 + A1 i + A2 j + A3 k∈Hm×m , the specific QSchur
decomposition is shown in the following pseudo-code. In steps 2–7, loop each quaternion vector of ARc to eliminate and update the
elements according to Theorem 2.3.2, where ARc is the first column block of real representation AR of the quaternion matrix A, and
then perform an Givens transform [24] to obtain the orthogonal matrix by the function orthogonalization(). In step 8, the function
realRepresent() generates a 4 m× 4 m real representation of matrix according to Eq. (5). Finally, the real representation’s first
column block URc of the quaternion Schur matrix U and the real representation’s first column block V Rc of the unitary quaternion

4
Y. Sun et al. Optik 269 (2022) 169899

Fig. 1. The flowchart of embedding process.

matrix V are obtained.

Remark 2.3. After Householder matrix is reduced, the JRS-symmetric algebraic structure of AR is still retained known from Algo­
rithm 2.3. So only the first column block ARc of real representation AR needs to be stored and calculated in order to greatly reduce the
calculation amount of the whole algorithm. What’s more, two algorithms of implementing QSchur decomposition are compared, one
using quaternion toolbox of Matlab and the other using the presented algorithm. It can be obtained from the comparison results of
Table 5, compared with using quaternion toolbox, the total average running time of the proposed algorithm is about 20 times shorter.
The reason is that the real matrix instead of quaternion matrix is used by Algorithm 2.3 to avoid the complex calculation of quaternion,
and then the four component matrices are synchronously simplified into the structural matrix consisting of four upper triangular
matrices. Therefore, the real structure-preserving algorithm of the proposed algorithm greatly reduces the running time of the algo­
rithm combined with quaternion matrix, which is more conducive to the practical application of copyright protection.

3. The proposed algorithm

The watermarking algorithm mainly consists of two parts, namely, watermark embedding and watermark extraction. In this
section, we mainly focus on the two parts vividly in the form of pseudo code and flowchart.

3.1. The process of watermark embedding

The entire embedding process is explained by the following pseudo code. For the pseudo code of watermark embedding, the first
step is the encryption of watermark image; the second step is the selection of embedded blocks via Matlab’s built-in randperm function;
the fifth step is the addition of dither noise, if the pixel value becomes negative after this step, make it negative again; the sixth to
eighth steps are the embedding of watermark bits, which include the core algorithm QSchur decomposition and inverse QSchur
decomposition; the ninth step is pixel fine-tuning mechanism; the tenth step is to update the original block to the watermarked block;
and the third to eleventh steps are cyclic embedding of all watermark bits. Finally, the watermarked image C*is obtained. As shown in

5
Y. Sun et al. Optik 269 (2022) 169899

Fig. 2. The flowchart of extraction process.

Fig. 1, in order to make the process more vivid and easier to understand, the process is also organized in the form of a flowchart.

6
Y. Sun et al. Optik 269 (2022) 169899

Fig. 3. Five carrier images used for testing.

3.2. The process of watermark extraction

The entire extraction process is explained by the following pseudo code. For the pseudo code of watermark extraction, the first step
is the selection of embedded blocks; the fourth and fifth steps are the extraction of watermark bits, which include the core algorithm
QSchur decomposition; the second to sixth steps are the cyclic extraction of all watermark bits; the seventh step is to decrypt the
extracted watermark bits and convert them into the final extracted watermark image W* . Similarly, to make the process more un­
derstandable, the process is shown as a flowchart, as shown in Fig. 2.

4. Experimental results

The performances of the proposed algorithm are evaluated through simulation experiments, so this section mainly introduces the
dataset and evaluation indicators required by the simulation experiment, as well as the experimental data and detailed analysis of each
performance.

4.1. Test data and basis

This part introduces the preparation of the simulation experiment, including the dataset used in the experiment, the determination
and setting of various parameters in the process of the experiment, and the use of appropriate evaluation indicators to evaluate the
experimental results.

4.1.1. Dataset
The carrier images and watermark images used for testing in the proposed algorithm are 24-bit color images, where carrier images
are from USC-SIPI [31] and CVG-UGR [32] databases. For illustrating the performance of proposed algorithm in all aspects, we decided

7
Y. Sun et al. Optik 269 (2022) 169899

Fig. 4. Two watermark images used for testing.

Fig. 5. The broken line chart of the three indicators changing with quantization steps.

to select five carrier images as shown in Fig. 3 and two watermark images as shown in Fig. 4, in which the sizes of the selected carrier
image and watermark image are 512 × 512 and 32 × 32 respectively.

4.1.2. Metrics
For evaluating the invisibility and robustness of the proposed algorithm, two indicators are used in this paper, namely Peak Signal
to Noise Ratio (PSNR) [36] and Normalized Cross-correlation (NC) [37]. As for the invisibility of watermark, the larger the PSNR value,
the better the invisibility. As for the robustness of watermarking, the greater the NC value, the better the robustness, when the NC value
is 1, the algorithm can perfectly extract all watermark bits without any error. The definitions of PSNR and NC are shown in Eqs. (13)
and (14) respectively.
⎛ ⎛ ⎞⎞
/
⎜∑ 3 ⎜ 2
255 × M × N ⎟⎟
⎜ ⎜ ⎟⎟
PSNR = ⎜ 10 × log10 ⎜ M N ⎟⎟ 3 (13)
⎝ p=1 ⎝∑ ∑ [ ]2 ⎠⎠
Cp (i, j) − Cp ∗ (i, j)
i=1 j=1

where M× N is the size of the color image, Cp (i, j) and C∗p (i, j) are the pixel values of the carrier image and the watermarked image at the
position (i, j) in the p channel respectively, and the value of p represents the channel of the color image.
∑
3 ∑
M ∑
N
[W(x, y, p) × W ∗ (x, y, p)]
(14)
p=1 x=1 y=1
NC = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ 3 ∑ M ∑ N ∑ 3 ∑ M ∑ N
[W(x, y, p)]2 [W ∗ (x, y, p)]2
p=1 x=1 y=1 p=1 x=1 y=1

8
Y. Sun et al. Optik 269 (2022) 169899

Fig. 6. Test results for PSNR with different watermark images. (a) The PSNR variation values of different watermarked images with “Logo 1′′ . (b)
The PSNR variation values of different watermarked images with “Logo 2′′ .

where W(x, y, p) and W∗ (x, y, p) are the pixel values of the color watermark image and the extracted watermark image at the position
(x, y) of the p channel, respectively.

4.1.3. Parameter setting

In the proposed algorithm, the most important parameter is quantization step T and it is the key to balance the invisibility and
robustness. Fig. 5 is obtained by using the images shown in Figs. 3 and 4 to test and evaluate the proposed algorithm with the indicators
referred in Section 4.1.2. PSNR is the average value of all watermarked images, U_NC represents the average NC value of the
watermark image extracted from the watermarked image without attacks, and A_NC represents the average NC value of the watermark
image extracted from the watermarked image subject to common attacks. The basic principle for determining T is that U_NC= 1, PSNR
＞35 dB, and A_NC＞0.9. As shown in Fig. 5, in order to have better robustness on the basis of good invisibility, we choose the T of the
intersection point of A_NC and PSNR, that is, T = 0.026.
What’s more, we set the block size to 4 × 4. Based on this parameter, it is found that the coefficients on the second and third rows of
the first column of the imaginary part matrix Ui (i = 0, 1, 2, 3) in the unitary quaternion matrix U have the strongest correlation after
QSchur decomposition, which determines the values of two parameters p and q. According to these parameters, we carried out the next
experiments.

4.2. Invisibility test and analysis

Good invisibility is an essential factor for a good watermarking algorithm, which is mainly reflected by the visual quality of the

9
Y. Sun et al. Optik 269 (2022) 169899

Table 1
The attack and its serial number.
Attack number Attack types and parameters

A1 JPEG (90)
A2 JPEG 2000 (5:1)
A3 Gaussian noise (0, 0.001)
A4 Salt & Peppers noise (5%)
A5 Poisson noise
A6 Cropping (left 30%)
A7 Scaling (2:1)
A8 Gaussian low pass filtering (3 ×3)
A9 Motion blur (9, 0)
A10 Histogram equalization
A11 Brighten (from [0,1] to [0.4, 1])
A12 Darken (from [0,1] to [0, 0.6])
A13 Rotation (5◦ )
A14 Translation (35, − 25)

watermarked image. Here, we use watermark and carrier images in Section 4.1.1 to evaluate and test the proposed algorithm. As
shown in Fig. 6, the PSNR values corresponding to all quantization steps are displayed, where the range of T tested is [0.01, 0.05].
Fig. 6(a)-(b) shows results that embedded watermark images are “Logo 1′′ and “Logo 2′′ respectively. When the PSNR value of the
watermarked image is greater than 35 dB, the image quality is good [38], while all PSNR values are greater than 35 dB when T is in the
interval [0.01, 0.032]. What’s more, in order to balance the performance of two mutual constraints, namely invisibility and robustness,
T is selected as 0.026. Therefore, the watermark invisibility of proposed algorithm meets the basic requirement.

4.3. Robustness test and analysis

Strong robustness is essential factor for a good watermarking algorithm, which is mainly embodied by the quality of the watermark
image extracted from the attacked watermarked image.
For reflecting the better robustness of proposed algorithm more objectively, the comparison algorithms are reproduced and their
PSNR values of different images are adjusted to be consistent with the PSNR values of the proposed algorithm shown in Fig. 6. In other
words, the robustness is compared while maintaining consistency in invisibility. Therefore, by changing the parameters of each al­
gorithm according to Fig. 6, the PSNR values of each identical watermarked image in different algorithms are adjusted to approxi­
mately the same value, that is, the PSNR values of watermarked image “Barbara”, “House”, “Sailboat”, “Lena”, “Peppers” are 36.27 dB,
36.8 dB, 36.85 dB, 36.68 dB, 37.9 dB, respectively.
Next, we compare the proposed algorithm with some advanced watermarking algorithms whose watermarking mechanisms are
similar to that of the algorithm based on QSchur decomposition. As shown in Table 1, a variety of common attacks are used to test the
proposed algorithm, such as JPEG compression, noise, filtering, geometric attacks, etc. Tables 2 and 3 show the comparison results of
NC values of different watermarking algorithms, and Fig. 7 shows the robustness comparison result by NC values and visual effects of
extracted watermark images.
Based on the data obtained from experiments, we made a series of analyses as follows. For compression attacks (A1 and A2), it’s a
lossy compression and most of the energy is retained, so all watermarking algorithms have high NC values and good robustness. For
Gaussian noise (A3), since this attack seriously changes pixel values, the robustness of all algorithms is not very good, where the
robustness of algorithms [12,17] are poor and the robustness of proposed algorithm is similar to that of the other two algorithms [13,
15]. For Salt & Peppers noise (A4), proposed algorithm and algorithms [12,17] show good robustness, but algorithms [13,15] are not
robust. For Poisson noise (A5), since this attack reduces image contrast and hides image details, the robustness of all algorithms is poor,
but proposed algorithm is better than most of comparison algorithms. For cropping attack (A6), the robustness of proposed algorithm is
between algorithms [12,13,15] and algorithm [17]. For scaling attack (A7), all algorithms show good robustness, with NC values
around 0.99. This is mainly because the enlarged image is realized by inserting new pixels into the original image according to the
surrounding pixels and certain rules, so the newly inserted pixels basically have the same relationship as the original pixels. When the
image is scaled back, there is still a certain relationship between the remaining pixels, so this attack has little impact on the image. For
Gaussian low pass filtering (A8), except for the algorithm [12], all other algorithms have good resistance. For motion blur attack and
histogram equalization attack (A9 and A10), other algorithms show poor robustness, especially for resisting A10 attack, while the NC
values of the proposed algorithm are almost 1, so the proposed algorithm is more robust to the attack. For the brightness adjustment
attacks (A11 and A12), the NC values of other algorithms are lower than 0.8, showing extremely poor robustness, while the NC values
of the proposed algorithm are all 1, which represents proposed algorithm is particularly robust to the attack. For rotation attack (A13),
the NC values of all algorithms are mostly around 0.98, which shows strong robustness. In addition to the serious loss of pixels at the
edge of the image, the pixels in the image block still have a certain correlation and the value range of the maximum energy element will
not fluctuate too much after the rotation and recovery of the attacked image, so they still have certain robustness under this attack. In
particular, the algorithm [17] with image correction performs detailed processing on the image, so it has the strongest robustness, with
NC value almost 1. For translation attack (A14), the NC value of algorithms [12,17] is around 0.98 and the NC value of the proposed
algorithm is around 0.95, showing strong robustness. However, the NC value of algorithms [13,15] is around 0.88, showing poor

10
 Y. Sun et al.
Table 2
NC values of different watermarking algorithms with “Logo 1′′ after different attacks.
Attack number Barbara House

Algorithm[12] Algorithm[13] Algorithm[15] Algorithm[17] Proposed algorithm Algorithm[12] Algorithm[13] Algorithm[15] Algorithm[17] Proposed algorithm

A1 0.99999 0.99765 0.99575 0.99982 0.96045 1.00000 0.99982 0.97426 1.00000 0.99124
A2 1.00000 0.99710 0.99673 0.99982 0.97147 1.00000 1.00000 0.97454 1.00000 1.00000
A3 0.74306 0.93978 0.91009 0.83788 0.83995 0.71017 0.93703 0.88943 0.78654 0.87472
A4 0.96850 0.68642 0.78370 0.96424 0.94895 0.98026 0.71548 0.77333 0.97039 0.95777
A5 0.72545 0.84478 0.82313 0.73840 0.78483 0.60448 0.81744 0.76184 0.64930 0.79977
11

A6 0.98364 0.70432 0.70999 0.99465 0.86497 0.77893 0.70667 0.69833 0.99535 0.86524
A7 1.00000 0.99857 0.99652 0.99982 0.97351 1.00000 0.99989 0.97602 0.99996 0.99488
A8 0.87900 0.97401 0.98269 0.99939 0.97790 0.93738 0.97495 0.98424 0.99947 0.99452
A9 0.64399 0.67220 0.62087 0.63208 0.83802 0.79805 0.81332 0.72775 0.76172 0.87259
A10 0.68876 0.58077 0.57037 0.56762 0.99145 0.77385 0.48089 0.52144 0.50803 0.99527
A11 0.40131 0.46888 0.60444 0.52587 1.00000 0.50745 0.58309 0.47100 0.44768 1.00000
A12 0.41213 0.51348 0.44087 0.39761 1.00000 0.53212 0.53564 0.46797 0.52438 1.00000
A13 0.99599 0.99254 0.99065 0.99924 0.98153 0.99848 0.99877 0.97674 0.99899 0.98576
A14 0.99430 0.88347 0.89373 0.99951 0.94820 0.99321 0.88297 0.88423 0.99940 0.94820
Average 0.81687 0.80385 0.80854 0.83257 0.93437 0.82960 0.81757 0.79151 0.83152 0.94857

Optik 269 (2022) 169899

 Y. Sun et al.
Table 3
NC values of different watermarking algorithms with “Logo 2′′ after different attacks.
Attack number Sailboat Lena

Algorithm[12] Algorithm[13] Algorithm[15] Algorithm[17] Proposed algorithm Algorithm[12] Algorithm[13] Algorithm[15] Algorithm[17] Proposed algorithm

A1 0.99988 0.99700 0.98791 0.99926 0.98239 0.99999 0.99699 0.99975 1.00000 0.98370
A2 0.99998 0.99828 0.98800 0.99999 0.99947 0.99970 0.99903 0.99926 1.00000 0.99452
A3 0.80056 0.95476 0.91901 0.78748 0.90684 0.81521 0.96388 0.92277 0.82552 0.92199
A4 0.98972 0.82176 0.85157 0.96695 0.97072 0.98426 0.79951 0.86183 0.96798 0.96500
A5 0.79088 0.87389 0.85596 0.68054 0.87470 0.76526 0.88917 0.85617 0.69678 0.87739
12

A6 0.78027 0.69480 0.69407 0.99513 0.90692 0.78026 0.69417 0.90321 0.99540 0.90632
A7 0.99977 0.99869 0.98843 0.99830 0.99179 0.99974 0.99995 0.99969 1.00000 0.99325
A8 0.92484 0.98042 0.99307 0.99832 0.98875 0.94418 0.99021 0.98919 0.99819 0.99564
A9 0.78917 0.77814 0.73231 0.65860 0.89962 0.79702 0.80517 0.73426 0.66224 0.90291
A10 0.80288 0.83550 0.80737 0.65716 0.99821 0.81011 0.76890 0.85913 0.47167 0.99579
A11 0.56183 0.65621 0.70322 0.48560 1.00000 0.60376 0.72304 0.63741 0.54323 1.00000
A12 0.65158 0.70291 0.59496 0.47498 1.00000 0.57328 0.69529 0.58649 0.40640 1.00000
A13 0.99912 0.99781 0.98671 0.99969 0.98750 0.99800 0.99632 0.99716 0.99962 0.98838
A14 0.97514 0.88252 0.89357 0.97681 0.96881 0.97536 0.88529 0.89576 0.97842 0.96881
Average 0.86183 0.86948 0.85687 0.83420 0.96255 0.86044 0.87192 0.87443 0.82468 0.96384

Optik 269 (2022) 169899

Y. Sun et al. Optik 269 (2022) 169899

Fig. 7. Visual effects and NC values comparison of watermark images extracted from different algorithms.

robustness against this attack. The main reason is the distribution of watermark bits, if the distribution of watermark bits is not uniform
or concentrated, more watermarked blocks will be lost after this attack. Algorithms [13,15] adopts classical Arnold transform, al­
gorithms [12,17] adopts affine transform with large key space, and the proposed algorithm adopts 2D-LASM which is sensitive to the
initial state, so different block selection algorithms will affect the robustness under this attack. Obviously, the average NC value of

13
Y. Sun et al. Optik 269 (2022) 169899

Fig. 8. Watermark images and their encrypted images: (a) “Logo 1′′ , (b)-(d) the encrypted images of (a) by Arnold transform, Affine transform and
2D-LASM respectively, (e) “Logo 2′′ , (f)-(h) the encrypted images of (e) by Arnold transform, Affine transform and 2D-LASM respectively.

Table 4
Distribution of the number of pixels before and after encryption.
Logo 1 Logo 2

Red Green Blue Red Green Blue

Number of pixels before encryption (0/1) 1912/6280 2027/6165 1998/6194 3672/4520 3021/5171 3819/4373
Number of pixels after encryption (0/1) in algorithms [12,13,15,17] 1912/6280 2027/6165 1998/6194 3672/4520 3021/5171 3819/4373
Number of pixels after encryption (0/1) in proposed algorithm 4089/4103 4106/4086 4095/4097 4017/4175 4108/4084 4050/4142

proposed algorithm exceeds 0.9, which is much higher than other algorithms, so the robustness of the proposed algorithm is superior to
other algorithms, showing the superiority of this algorithm.
In order to show that the proposed algorithm has better performance more clearly, “Logo 1′′ is embedded into “Peppers” to show the
comparison results. Fig. 7 shows effect diagrams and NC values of watermark images extracted from different algorithms. The whole
watermark image can be extracted without error observed from Fig. 7 when the attack type is none, A11 and A12. While under other
attacks, the visual quality of extracted watermark images decreases slightly, but it is also recognizable. In contrast, the watermark
image extracted by other algorithms under some attacks is of poor quality and cannot be recognized visually. For example, the
watermark image extracted by algorithm [12] under the attack of A3, A5, A6, and A9-A12, the watermark image extracted by al­
gorithm [13] under the attack of A4, A6, and A9-A12, the watermark image extracted by algorithm [15] under the attack of A3-A6, and
A9-A12, and the watermark image extracted by algorithm [17] under the attack of A3, A5 and A9-A12. Obviously, under the attack of
A9-A12, other algorithms cannot extract useful watermark information at all, but only the proposed algorithm can extract identifiable
watermark, which reflects the advantages and functions of the quaternion. In addition, the average NC value of proposed algorithm
exceeds 0.9, while that of other algorithms is only about 0.8. Therefore, the proposed algorithm has greater advantages than other
algorithms.

4.4. Watermark security analysis

In order to effectively prevent infringement, 2D-LASM is used to encrypt watermark images. In watermarking algorithm, the

14
Y. Sun et al. Optik 269 (2022) 169899

Table 5
Time comparison of different methods and ways (second).
Algorithm Watermark embedding Watermark extraction Total time

Algorithm [12] 0.5575 0.7039 1.2614

Algorithm [13] 0.9301 0.2850 1.2151
Algorithm [15] 0.5663 0.3205 0.8868
Algorithm [17] 1.6441 1.5610 3.2051
Use the toolbox 92.0502 28.2809 120.3311
Proposed algorithm 4.5002 2.3233 6.8235

common encryption algorithms include Affine transform [12,17] and Arnold transform [13,15], but because only the order of pixels of
the watermark image encrypted by them is disrupted, while the relative number of ‘0′ and ‘1′ pixels has not changed, so it is easy for
infringers to discover the rules and crack the algorithm to a certain extent. Fig. 8(a) is the original watermark “Logo 1′′ , Fig. 8(b)-(d) are
the encrypted images of “Logo 1′′ by Arnold transform, Affine transform and 2D-LASM respectively, and Fig. 8(e)-(h) are also the
original and corresponding encrypted images of “Logo 2′′ respectively. Table 4 shows the change in the number of ‘0′ and ‘1′ pixels
before and after using 2D-LASM to encrypt watermark images, where ‘Red’, ‘Green’ and ‘Blue’ represent three image channels of the
watermark. As a result, as shown in Fig. 8 and Table 4, 2D-LASM not only exhibits more chaotic behavior, but also encrypts the
watermark image into a more evenly distributed ciphertext image, so it’s difficult for infringers to detect the rules through statistical
algorithms and obtain the original information.

4.5. Watermark embedding capacity analysis

In existing watermarking algorithms, embedded watermark images include binary watermark image, grayscale watermark image,
and color watermark image, which indicate that watermark embedding capacity has an indispensable impact on the performance of
watermarking algorithms. Relatively speaking, the more embedded watermark information, the more pixels of the carrier image need
to be modified, and the worse the invisibility of the watermark will be. Therefore, the amount of embedded watermark information
should be appropriate.
For the proposed algorithm, the maximum embedding capacity is three times of total number of non-overlapping blocks of a carrier
image, for example, for a 512 × 512 carrier image, the maximum embedding capacity is 128 × 128 × 3 when block size is 4 × 4.
Nevertheless, The proposed algorithm is actually a 512 × 512 color carrier image embedded a 32 × 32 color watermark image, so the
actual watermark embedding rate is (32 ×32 ×24)/(512 ×512 ×3)= 0.03125bpp (bit per pixel). We also analyze actual embedding
rates of algorithms [12,13,15,17], because the carrier image and watermark image are 24 bits and the sizes of the two kinds of image
are the same respectively, their actual embedding rates are the same as that of the proposed algorithm. As a result, the proposed
algorithm’s actual watermark embedding capacity is consistent with that of algorithms used for comparison, which also ensures that
these algorithms are fair in other performance comparisons.

4.6. Running-time test and analysis

The hardware environment and software platform of the proposed method are AMD Ryzen 7 4700 U with Radeon Graphics
2.00 GHz, 16.00 GB RAM, x64, Windows 10 and Matlab (R2017a). Table 5 shows time comparison of different methods. The time
efficiency of algorithms [12,13,15,17] is high, because they embed watermark information into a single-channel image block without
considering the internal relationship between image channels. The total running time of the proposed algorithm is about 7 seconds, but
this time efficiency is still a limitation in the 5G era with high speed and low delay. Nonetheless, compared with the realization of this
algorithm using the quaternion toolbox of Matlab, the proposed algorithm has greatly improved in terms of time efficiency, and the
time has been shortened by about 20 times, which also proves the effectiveness of the proposed algorithm.

5. Conclusions

A new QSchur decomposition algorithm is designed and presented in this paper, which is an effective algorithm for copyright
protection of digital products. Firstly, the watermark image is encrypted and scrambled by chaotic system to ensure the security of the
algorithm. Then, the split color channels are combined by the quaternion, and they are decomposed as a whole, so that the correlation
between the color channels and the synchronization of watermark embedding are not lost. Finally, the watermark is embedded and
extracted blindly by modifying the relative relationship between multiple coefficient pairs. Besides, a great quantity of comparative
experiments demonstrate that the invisibility of the proposed algorithm can meet the basic requirement, the robustness is better than
other related watermarking algorithms under different kinds of attacks, and the security is higher than other encryption algorithms
commonly used in other watermarking algorithms. Although the watermarking robustness of proposed algorithm is better than that of
related algorithms after some attacks, watermark recognition under some attack types is still a little difficult from the renderings. In the
future, we will try to overcome these shortcomings and continue to study matrix factorization and its application to digital
watermarking.

15
Y. Sun et al. Optik 269 (2022) 169899

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work report in this paper.

Data Availability

The authors do not have permission to share data.

Acknowledgments

The work was supported by the National Natural Science Foundations of China (No. 61771231, 62171209, 61772253, 61873117,
61872170 and 61803253), the Key Project of Shandong Natural Science Foundation (No. ZR2020KF023), Shandong Big Data
Development and Innovation Laboratory of Public Resources, and China Scholarship Council (202008370052).

References

[1] X.T. Zhang, Q.T. Su, Z.H. Yuan, D.C. Liu, An efficient blind color image watermarking algorithm in spatial domain combining discrete Fourier transform, Optik
219 (2020), 165272, https://doi.org/10.1016/j.ijleo.2020.165272.
[2] Q.T. Su, D.C. Liu, Y.H. Sun, A robust adaptive blind color image watermarking for resisting geometric attacks, Inf. Sci. 606 (2022) 194–212, https://doi.org/
10.1016/j.ins.2022.05.046.
[3] Q.T. Su, Y.G. Niu, G. Wang, S.L. Jia, J. Yue, Color image blind watermarking scheme based on QR decomposition, Signal Process. 94 (2014) 219–235, https://
doi.org/10.1016/j.sigpro.2013.06.025.
[4] Q.T. Su, B.J. Chen, Robust color image watermarking technique in the spatial domain, Soft Comput. 22 (2018) 91–106, https://doi.org/10.1007/s00500-017-
2489-7.
[5] T. Huynh-The, C.H. Hua, N.A. Tu, T. Hur, J. Bang, D. Kim, M.B. Amin, B.H. Kang, H. Seung, S. Lee, Selective bit embedding scheme for robust blind color image
watermarking, Inf. Sci. 426 (2018) 1–18, https://doi.org/10.1016/j.ins.2017.10.016.
[6] X.T. Zhang, Q.T. Su, A spatial domain-based color image blind watermarking scheme integrating multilevel discrete Hartley transform, Int. J. Intell. Syst. 36
(2021) 4321–4345, https://doi.org/10.1002/int.22461.
[7] M. Cedillo-Hernandez, A. Cedillo-Hernandez, F.J. Garcia-Ugalde, Improving DFT-based image watermarking using particle swarm optimization algorithm,
Mathematics 9 (2021) 1795, https://doi.org/10.3390/math9151795.
[8] S. Roya, A.K. Pal, A blind DCT based color watermarking algorithm for embedding multiple watermarks, AEU-Int. J. Electron. Commun. 72 (2017) 149–161,
https://doi.org/10.1016/j.aeue.2016.12.003.
[9] J.Y. Wang, Z.G. Du, A method of processing color image watermarking based on the Haar wavelet, J. Vis. Commun. Image Represent. 64 (2019), 102627,
https://doi.org/10.1016/j.jvcir.2019.102627.
[10] T. Luo, G.Y. Jiang, M. Yu, H.Y. Xu, W. Gao, Robust high dynamic range color image watermarking method based on feature map extraction, Signal Process. 155
(2019) 83–95, https://doi.org/10.1016/j.sigpro.2018.09.024.
[11] B. Ma, L.L. Chang, C.P. Wang, J. Li, G. Li, Z.Q. Xia, X.Y. Wang, Double medical images zero‑watermarking algorithm based on the chaotic system and ternary
accurate polar complex exponential transform, J. Math. Imaging Vis. 63 (2021) 1160–1178, https://doi.org/10.1007/s10851-021-01048-w.
[12] D.C. Liu, Z.H. Yuan, Q.T. Su, A blind color image watermarking scheme with variable steps based on Schur decomposition, Multimed. Tools Appl. 79 (2020)
7491–7513, https://doi.org/10.1007/s11042-019-08423-1.
[13] Q.T. Su, X.F. Zhang, G. Wang, An improved watermarking algorithm for color image using Schur decomposition, Soft Comput. 24 (2020) 445–460, https://doi.
org/10.1007/s00500-019-03924-5.
[14] Q.T. Su, Y.H. Liu, D.C. Liu, Z.H. Yuan, H.Y. Ning, A new watermarking scheme for colour image using QR decomposition and ternary coding, Multimed. Tools
Appl. 78 (2019) 8113–8132, https://doi.org/10.1007/s11042-018-6632-y.
[15] Q.T. Su, B.J. Chen, A novel blind color image watermarking using upper Hessenberg matrix, AEU-Int. J. Electron. Commun. 78 (2017) 64–71, https://doi.org/
10.1016/j.aeue.2017.05.025.
[16] H.T. Hu, L.Y. Hsu, H.H. Chou, An improved SVD-based blind color image watermarking algorithm with mixed modulation incorporated, Inf. Sci. 519 (2020)
161–182, https://doi.org/10.1016/j.ins.2020.01.019.
[17] D.C. Liu, Q.T. Su, Z.H. Yuan, X.T. Zhang, A blind color digital image watermarking method based on image correction and eigenvalue decomposition, Signal
Process.: Image Commun. 95 (2021), 116292, https://doi.org/10.1016/j.image.2021.116292.
[18] C.P. Wang, X.Y. Wang, Y.W. Li, Z.Q. Xia, C. Zhang, Quaternion polar harmonic Fourier moments for color images, Inf. Sci. 450 (2018) 141–156, https://doi.org/
10.1016/j.ins.2018.03.040.
[19] Y.N. Liu, S.S. Zhang, J. Yang, Color image watermark decoder by modeling quaternion polar harmonic transform with BKF distribution, Signal Process.: Image
Commun. 88 (2020), 115946, https://doi.org/10.1016/j.image.2020.115946.
[20] Y. Chen, Z.G. Jia, Y. Peng, Y.X. Peng, D. Zhang, A new structure-preserving quaternion QR decomposition method for color image blind watermarking, Signal
Process. 185 (2021), 108088, https://doi.org/10.1016/j.sigpro.2021.108088.
[21] Z.G. Jia, M.S. Wei, M.X. Zhao, Y. Chen, A new real structure-preserving quaternion QR algorithm, J. Comput. Appl. Math. 343 (2018) 26–48, https://doi.org/
10.1016/j.cam.2018.04.019.
[22] A. Bunse-Gerstner, R. Byers, V. Mehrmann, A quaternion QR algorithm, Numer. Math. 55 (1989) 83–95.
[23] Y. Li, M.S. Wei, F.X. Zhang, J.L. Zhao, Real structure-preserving algorithms of Householder based transformations for quaternion matrices, J. Comput. Appl.
Math. 305 (2016) 82–91, https://doi.org/10.1016/j.cam.2016.03.031.
[24] Z.G. Jia, M.S. Wei, S.T. Ling, A new structure-preserving method for quaternion Hermitian eigenvalue problems, J. Comput. Appl. Math. 239 (2013) 12–24,
https://doi.org/10.1016/j.cam.2012.09.018.
[25] L.Y. Hsu, H.T. Hu, Blind watermarking for color images using EMMQ based on QDFT, Expert Syst. Appl. 149 (2020), 113225, https://doi.org/10.1016/j.
eswa.2020.113225.
[26] X.L. Liu, Y.F. Wu, H. Zhang, J.S. Wu, L.M. Zhang, Quaternion discrete fractional Krawtchouk transform and its application in color image encryption and
watermarking, Signal Process. 189 (2021), 108275, https://doi.org/10.1016/j.sigpro.2021.108275.
[27] J.Z. Li, Q. Lin, C.Y. Yu, X.C. Ren, P. Li, A QDCT- and SVD-based color image watermarking scheme using an optimized encrypted binary computer-generated
hologram, Soft Comput. 22 (2018) 47–65, https://doi.org/10.1007/s00500-016-2320-x.
[28] Z.Q. Xia, X.Y. Wang, C.P. Wang, B. Ma, H. Zhang, Q. Li, Novel quaternion polar complex exponential transform and its application in color image zero-
watermarking, Digit. Signal Process. 116 (2021), 103130, https://doi.org/10.1016/j.dsp.2021.103130.
[29] Z.Y. Hua, Y.C. Zhou, Image encryption using 2D Logistic-adjusted-Sine map, Inf. Sci. 339 (2016) 237–253, https://doi.org/10.1016/j.ins.2016.01.017.

16
Y. Sun et al. Optik 269 (2022) 169899

[30] X.B. Kang, F. Zhao, Y.J. Chen, G.F. Lin, C.N. Jing, Combining polar harmonic transforms and 2D compound chaotic map for distinguishable and robust color
image zero-watermarking algorithm, J. Vis. Commun. Image Represent. 70 (2020), 102804, https://doi.org/10.1016/j.jvcir.2020.102804.
[31] [Dataset] University of Southern California, Signal and Inage Processing Institute, USC-SIPI Image Database, 1997. http://sipi.usc.edu/database/. (Accessed 20
May 2021).
[32] [Dataset] University of Granada, Computer Vision Group, CVG-UGR Image Database, 2002. http://decsai.ugr.es/cvg/dbimagenes/c512.php. (Accessed 20 May
2021).
[33] Sangwine S.J., Bihan N.L. Quaternion toolbox for matlab. http://qtfm.sourceforge.net/. (Accessed 15 June 2021).
[34] W.R. Hamilton, Elements of Quaternions, Longmans Green, London, 1866.
[35] Q.W. Wang, X.X. Wang, Arnoldi method for large quaternion right eigenvalue problem, J. Sci. Comput. 82 (2020) 58, https://doi.org/10.1007/s10915-020-
01158-4.
[36] Q.T. Su, Y.H. Sun, X.T. Zhang, H.Y. Wang, G. Wang, T. Yao, A watermarking scheme for dual color images based on URV decomposition and image correction,
Int. J. Intell. Syst. (2022), https://doi.org/10.1002/int.22893.
[37] Q.T. Su, X.T. Zhang, H.Y. Wang, A blind colour image watermarking algorithm combined spatial domain and SVD, Int. J. Intell. Syst. (2021), https://doi.org/
10.1016/10.1002/int.22738.
[38] Q.T. Su, X.T. Zhang, A blind color image watermarking algorithm combined spatial domain and SVD, Int. J. Intell. Syst. (2021), https://doi.org/10.1007/
s00500-019-03924-5.
[39] K.M. Hosny, M.M. Darwish, Novel quaternion discrete shifted Gegenbauer moments of fractional-orders for color image analysis, Appl. Math. Comput. 421
(2022), 126926, https://doi.org/10.1016/j.amc.2022.126926.
[40] K.M. Hosny, M.M. Darwish, T. Aboelenen, Novel fractional-order polar harmonic transforms for gray-scale and color image analysis, J. Frankl. Inst. 357 (2020)
2533–2560, https://doi.org/10.1016/j.jfranklin.2020.01.025.
[41] Z.H. Shao, X.L. Liu, Q.J. Yao, N. Qi, Y.Y. Shang, J.J. Zhang, Multiple-image encryption based on chaotic phase mask and equal modulus decomposition in
quaternion gyrator domain, Signal Process.: Image Commun. 80 (2020), 115662, https://doi.org/10.1016/j.image.2019.115662.
[42] Z.H. Shao, Y.Y. Shang, R. Zeng, H.Z. Shu, G. Coatrieux, J.S. Wu, Robust watermarking scheme for color image based on quaternion-type moment invariants and
visual cryptography, Signal Process.: Image Commun. 48 (2016) 12–21, https://doi.org/10.1016/j.image.2016.09.001.
[43] X.L. Liu, Y.F. Wu, P.T. Gao, J.L. Ouyang, Z.H. Shao, Color image watermarking based on singular value decomposition and generalized regression neural
network, Multimed. Tools Appl. (2022), https://doi.org/10.1007/s11042-022-12990-1.
[44] Khalid M. Hosny, Mohamed M. Darwish, Mostafa M. Fouda, Robust color images watermarking using new fractional-order exponent moments, IEEE Access 9
(2021) 47425–47435, https://doi.org/10.1109/ACCESS.2021.3068211.
[45] Khalid M. Hosny, Mohamed M. Darwish, New geometrically invariant multiple zero-watermarking algorithm for color medical images, Biomed. Signal Process.
Control 70 (2021), 103007, https://doi.org/10.1007/s11042-022-12990-1.

17