Edge Detection and Compressive Sensing-Oriented

Digital Image Block Watermarking Algorithm

Peiqin Tang Changfan Zhang
College of Electrical and Information Engineering College of Electrical and Information Engineering
Hunan University of Technology Hunan University of Technology
Zhuzhou, China Zhuzhou, China
[email protected] [email protected]
2020 Chinese Automation Congress (CAC) | 978-1-7281-7687-1/20/$31.00 ©2020 IEEE | DOI: 10.1109/CAC51589.2020.9327312

Wenbo Pan Teng Long Jie Zhang

CRRC Zhuzhou Institute Co., Ltd CRRC Zhuzhou Institute Co., Ltd CRRC Zhuzhou Institute Co., Ltd
Zhuzhou, China Zhuzhou, China Zhuzhou, China
[email protected] [email protected] [email protected]

Abstract—This paper investigated the problem of the water- be achieved by CS theory, the drawbacks caused by the unified
mark algorithm for digital color image. An digital image wa- operation directly on the whole image still exist.
termark embedding and blind extraction algorithm is proposed To solve the above problems, the image block watermarking
based on edge detection and Compressive Sensing (CS) theory.
Firstly, the carrier image is divided into non-overlapping image algorithm was proposed which mainly contains two categories.
blocks. Then, the target image blocks suitable for concealing One way is embed watermark in all image blocks directly.
the watermark information are chosen by combining the edge This method is obviously the same to embedding watermarks
detection algorithm. Finally, watermarks are embedded into blindly into the whole image. Another way is select the part of
target image blocks by utilizing the Singular Value Decomposition the image blocks as target image blocks, and embedding the
(SVD) and CS theory. The experimental result shows that
the proposed digital images watermark algorithm has excellent watermark information [7]–[9] in those target image blocks.
invisibility and robustness performance in some scenarios. In reference [9], the variance value of each image block
Index Terms—digital watermarking, digital image, compressive is calculated and the image blocks with large value will
sensing theory, image edge detection be taken as the watermark embedding area. Apparently, the
second category algorithm has better watermark invisibility
I. I NTRODUCTION and robustness performance than the first category algorithm.
However, since semantic information is not considered,the
Digital image watermarking algorithm is a technology for watermark invisibility of the second category algorithm may
solving information encryption and copyright protection issue also can not be ensured. One way to solve this problem
by embedding the information into the digital carrier image is combining the edge detection algorithm with the image
[1]–[3]. How to guarantee the invisibility, robustness and watermark embedding scheme.
security of watermark has received considerable attention in In this study, we proposed a new block watermarking
information safety domain in recent years. However, the invisi- scheme based on edge detection and CS theory-oriented. Edge
bility of watermarks always contradicted against its robustness. detection is conducted using Sobel operator for the whole
The robustness of watermark will certainly be insufficient if its carrier image [10]. The image detected is divided into non-
invisibility is assured seriously. To compromise the robustness overlapping image blocks. The target image block suitable for
and invisibility, it is particularly important to embedding embedding the watermark is selected by the analysis on the
watermark into a carrier image at proper positions. statistical properties of image blocks and their relationship
In most of the watermarking algorithms, the whole image with surrounding ones. The embedding and encrypting of
is always being taken as a watermark embedding object of watermark information is then realized using SVD algorithm
uniform operation [4]–[6]. Most of these algorithms have and compressive sensing theory to ensure the robustness and
a problem, i.e. the local characteristics are often different security of the algorithm [11].
from the overall ones in an image. Therefore, the watermark The remainder of the paper is organized as follow. Section
embedding parameters obtained from overall characteristics II introduces the block CS theory. Section III presents an
are may not suitable for partial carrier image area. In [4], edge detectionoriented image block algorithm, and watermark
Huang et al. introduced the CS theory into watermarking embedding and extraction process. In Section IV, simulation
algorithm. Although the pretty security of the algorithm can test is conducted for the algorithm. Finally, some conclusions
for variance-oriented image block watermarking algorithm are
Corresponding author: Changfan Zhang presented in Section V.

978-1-7281-7687-1/20/$31.00 2020 IEEE 6684

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on May 17,2021 at 11:57:23 UTC from IEEE Xplore. Restrictions apply.
 II. P RELIMINARY WORK where, the ΦM is measurement matrix. To the whole image
X, the measurement matrix Φ is a diagonal matrix, expressed
A. Compressive Sensing (CS) Theory
as  
CS theory is also known as compressive sampling or sparse Φ0 
 
sampling. It is a signal reconstruction technology using sparse  Φ0 
 
or compressible signals [12], [13]. Set a one-dimensional Φ= .. . (7)
 . 
discrete signal f ∈ RN ×1 and express it with a orthogonal  
 Φ0 
basis matrix Ψ ∈ RN ×N as
Block CS algorithm could just storage the matrix Φ0 instead of
f = Ψh, (1) storaging Φ. Obviously, it can greatly save the storage space,
where h ∈ RN ×1 is the orthogonal transformation sparse coef- which is more conducive to the rapid implementation of the
ficient vector. The signal f is K-sparse signal when the number algorithm.
of non-zero elements n in h satisfies n ≤ K. This is the III. B LOCK CS WATERMARKING A LGORITHM
expression of signal sparsity. Ψ is called sparse decomposition
matrix. It is indicated that the signal is compressible if the A. Image Block Strategy
number of non-zero elements of h is much smaller than N . The watermark algorithm can obtain better security by the
The measurements vector Y ∈ RM ×1 may be obtained as addition of CS theory. However, the CS algorithm has also
shown in the equation (2) below if a linear measurement is following shortcomings: high computational complexity, large
conducted for the signal f . Formula (3) can be obtained by scale of memory required for storing compressed sampling
combination of (1) and (2). The process is called the perceptual operator, etc. Therefore, we can not combine the CS with
measurement of the signal. watermark algorithm directly. This problem can be solved
effectively by adopting the image edge detection and blocking
Y = Φf = Θh, (2) operation. Moreover, the invisibility and robustness of water-
marking algorithm can be effectively improved by localization
where
of embedding location and selection of proper areas for water-
Θ = ΦΨ (3) mark information embedding after image blocking processing.
is referred to CS matrix. The way to chose complex image areas for hiding water-
Now it is considered to reconstruct h using the measurement mark information are given below. Firstly, a binary image is
Y. To reconstruct signals, Cand and Tao have proven that, obtained from the edge detection of the whole carrier image
when the signal h meets the condition of K-sparse, the using Sobel operator [10]. The binary image is then subjected
measurement matrix Φ meets the condition of RIP (Restricted to non-overlapping block processing. The sum of the gray
Isometry Property) [14] and the number of measurements M value of each binary image block Pi is calculated. When Pi
satisfies meets one of following conditions, it is determined that it is
the target image block:
M ≥ K × log10 N
. (4)
1. The sum of gray values of image blocks Pi exceeds the
The signal h can be precisely reconstructed by solving threshold T set, i.e. Pi > T ;
minimization norm optimization equation [15] 2. Pi is smaller than T , but more than the gray values of
the four surrounding image blocks, then the image block is
ĥ = arg minh0 s.t. Φh = Y, (5) also determined as the target image block, i.e.
where,  · 0 is the 0 order norm of the vector and it refers to Pi ≥ max(P1 , P2 , P3 , P4 ). (8)
the number of non-zero elements in h. ĥ is the reconstructed
signal. B. Embedding of Watermark
After target image blocks are selected adopting the method
B. Block Compressed Sensing Algorithm specified in Section 3.1, watermark embedding operation can
When the large size of whole image was sampled, high be conducted for the target blocks. The color images are
computational complexity and long reconstruction time are the processed in three channels (R, G and B). The image of blue
troubles faced by the existing CS method. These problems can channel (B) is then selected as watermark carrier image. The
be solved effectively by using the block based compression process of embedding algorithm as shown in Fig. 1.
sampling method. The specific steps are as below:
Dividing the image X ∈ Rn×n into small image blocks of Step 1: Decomposition of the original color carrier image
2
size m × m. Denote the xi ∈ Rm ×1 to the signal for i-th im- X based on channels R, G and B is conducted. The image
2
age block, i = 1, 2, . . . , (n/m) . CS processing is conducted XB of channel B is processed according to the block strategy
for xi , and get the measurement vector xi , partitioned as specified in Section 3.1. The target image block XBi , (i =
1, 2, . . . , 16) is selected as the position of embedding and its
y i = Φ0 x i , (6) position i is saved as No.1 key for watermark extraction.

6685

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on May 17,2021 at 11:57:23 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1. Watermark Embedding Flow Chart


YA = UA ΣA VA
T
(14)
where α is embedding strength. α and ΞA are saved as No.2
key for watermark extraction; UW and VW T
are saved as No.
3 key for watermark extraction.
Step 6: CS optimal reconstruction is conducted using mea-

surement matrix Φ and the matrix YA in Step 5 to obtain the
wavelet subband matrix A with watermark embedded. Inverse
Fig. 2. Encryption and Restoration of Scrambling Transformation
wavelet transform is then carried out using A and the target
image block XBi with watermark embedded is recovered and
Step 2: The first-level discrete wavelet transform is conduct- obtained.
ed for XBi and its approximation subband coefficient matrix Step 7: The image of channel B XB is combined with
A is taken out. images of original channel images R and G (XR and XG )
Step 3: Scrambling transformation is carried out for binary to obtain the watermarked color image X .
watermark image sized 64 × 64. The scrambling times k is In this paper, the orthogonal wavelet theory is taken as
saved as No.4 key for extraction of watermarks. The results sparse transform base Ψ and the random bernoulli matrix
of scrambling transformation and recovery as shown in Figure is taken as measurement matrix Φ. In addition, orthogonal
2. The image after scrambling W is divided into four image matching pursuit algorithm (OMP) [16] is adopted for opti-
blocks Wj , (j = 1, 2, 3, 4) sized 32 × 32. mization and reconstruction.
Step 4: CS processing is conducted for the approximate C. Watermark Extraction
subband matrix A and the watermark image Wj scrambled The watermark extraction process is closely related to the
using sparse transform base Ψ and measurement matrix Φ. embedding method and it is the inverse process of watermark
Then the measurement matrixes YA and YW are obtained, embedding. The four keys in Section B are used for extrac-
expressed as tion. Then the security of the watermark information can be
YA = ΦA = ΦΨSA = ΘSA (9) guaranteed. The process of watermark extraction as shown in
Fig. 3.
and The specific steps are as below:
YW = ΦWj = ΦΨSW = ΘSW (10) Step 1: Decomposing the watermarked color images X and
where SA = ΨT A, SW = ΨT Wj , ΨT refers to the transpose extract the the image of channel B XB . XB is decomposed
of the matrix Ψ. into non-overlapping small pieces sized 64 × 64. Then the
Step 5: SVD transformation is conducted for YA and YW target image block XBi with watermark embedded are found
to generate orthogonal matrixes U and V and a diagonal using No. 1 key in Section 3.2.
matrix Ξ. ΞA is obtained for realizing watermark embedding Step 2: The first order wavelet decomposition transform
by the superposition of matrixes ΞA and ΞW . SVD inverse is conducted for XBi to obtain its wavelet subband matrix
transformation is then carried out for UA , ΞA and VA T
to X . CS is carried out for matrix A using the sparse base Ψ

obtain the measurement matrix YA containing watermark and measurement matrix Φ in Section 3.2, and then the CS

information. The specific embedding algorithm is as below: measurement matrix YA is obtained.

Step 3: SVD decomposition is conducted for YA to extract
YA = U A Ξ A VA
T
(11) 
the diagonal matrix ΞA . Then get the diagonal matrix ΞW by
used the No. 2 key information in Section 3.2 . The operation
YW = U W Ξ W VW
T
(12)
is
Ξ A = ΞA + αΞW (13) YA 
= UA ΞA VA T
(15)

6686

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on May 17,2021 at 11:57:23 UTC from IEEE Xplore. Restrictions apply.
 Fig. 3. Watermark Extraction Flow Chart

and
ΞW = (ΞA − ΞA ) /α (16)
Step 4: SVD inverse transformation is conducted with ΣW
and the No.3 key in Section 3.2 to obtain measurement matrix
YW ;
Step 5: CS optimization and reconstruction of ΞW is carried
out adopting OMP algorithm and measurement matrix Φ, and
the watermark images Wi with scrambling and encryption can
be obtained. (a) Lena Image (b) Baboon Image
Step 6: Scrambling restoration is conducted for Wi using
No.4 key in Section 3.2 to obtain binary watermark image
block Wi . After that, spliced the Wi in order and the binary
watermark image can then be obtained.
Since the watermark information is embedded repeatedly
for 4 times due to 16 targets image blocks are selected and
the original watermark image is divided into 4 small pieces,
thus the Wi shall be averaged after all Wi has been obtained.
The operation is also helpful for enhancing the robustness of
watermark. (c) Airplane Image (d) Landscape Image
Fig. 5. PSNR versus the embedding strength α
IV. S IMULATION T EST AND A NALYSIS
In this section, we present numerical examples to show the
performance of the watermarking algorithm we proposed. Four Fig. 4 displays the test results of embedding the watermark
color images (Lena, Baboon, Airplane, and Landscape) with into the carrier images when the α = 0.15, which shows that
different texture features are selected as test images, in which there is no obvious change in the whole image and it is hard to
the size of them are 8-bit ones of 512 pixel×512 pixel×3. detect the traces where the image is modified and has superior
The original watermark images are the two-dimensional code watermark invisibility performance.
binary image sized 64×64. Moreover, for comparison purpos- To evaluate the algorithm more objectively, the PSNR value
es, we compare our watermarking algorithm results with the changes of the whole image by adopting the algorithm of the
block image watermark algorithm in [9] and the whole image paper, watermark algorithm in reference [4] and [9] are tested
watermark algorithm in [4]. in the paper when the embedding strength α is increasing
from 0.05 to 0.3. The text results are shown in Fig. 5. As
A. Watermark Invisibility Test
it can be seen that, the PSNR value obtained by adopting the
Watermark invisibility refers to the hiding ability of wa- algorithm of the paper is higher than those gotten by the other
termark in carrier image and its performance is evaluated two algorithms. The value can keep at around 30 even though
objectively using PSNR value. The PSNR is defined as the embedding strength α is up to 0.3. In addition, with the
 
2552 × M × N increase of α, the advantage of the algorithm of the paper is
PSNR = 10 × lg N (17) more obvious and better invisibility.
 2
[X(i, j) − X (i, j)]
j=1 B. Watermark Robustness Test
Where, M and N refers to the length and width of the image To test the robustness of watermark, the attack such as
respectively; X and X are the original image and the one Gaussian noise, multiplicative noise, filtering and JPEG com-
with watermark embedded respectively. pression is applied for the images. In addition, the NC value is

6687

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on May 17,2021 at 11:57:23 UTC from IEEE Xplore. Restrictions apply.
 Name Lena Baboon Airplane Landscape

Water-
marked
image

Fig. 4. Watermarked color images

TABLE I missing of any of the keys. Therefore, the CS-oriented block

NC VALUE OF THE WATERMARK WHEN THERE IS NO ATTACKS watermark algorithm proposed in the paper has relatively high
security.
Images Lena Baboon Airplane landscape
Our method 1 0.9941 1 1 V. C ONCLUSION
Reference [9] 1 0.9915 0.9894 1
Reference [4] 0.9773 0.9641 0.9627 0.9639 An edge-detection and CS-oriented blind watermarking
algorithm for digital image is introduced in this paper. The
image blocked edge detection algorithm is proposed to select
used to evaluate objectively the quality of extracted watermark the area for hiding watermark information. By embedding the
images. The NC is defined as watermarks into those selected areas and combining CS algo-
rithm, the digital image watermarking algorithm we proposed

M 
N
exhibits more reliable invisibility and robustness performance.
W(i, j)W (i, j)
i=1 j=1 Moreover, the guarantee can be provided for the security of
NC = , (18)

M 
N
2
watermark by scrambling transformation processing and CS
W(i, j) transform, which also proved the the effectiveness of the
i=1 j=1
proposed watermarking algorithm in some scenarios.

where W and W refer to the original watermark image and Acknowledgments
that extracted respectively; M and N are the length and width This work was supported by the National Key R&D Program of
of the watermark image respectively. China (no. 2018YFD0400705).
When the image is not attacked, the normalized correlation R EFERENCES
coefficient of the watermark image extracted is shown in Table
[1] B. Sridhar, “Cross-layered embedding of watermark on image for high
1. When noise, filtering or JPEG compression attack is added, authentication,” Pattern Recognition and Image Analysis, vol. 29, no. 1,
the extracted watermark are as shown in Table 2 and Table pp. 194–199, 2019.
3. It can be seen from the results that, when an attack is [2] H. Singh, “Watermarking image encryption using deterministic phase
mask and singular value decomposition in fractional mellin transform
added, compared with the other two algorithms, the quality domain,” Iet Image Processing, vol. 12, no. 11, pp. 1994–2001, 2018.
of extracted watermark images by adopting the algorithm of [3] S. Saadi, A. Merrad, and A. Benziane, “Novel secured scheme for blind
the paper is relatively high subjectively and most of watermark audio/speech norm-space watermarking by arnold algorithm,” Signal
Processing, vol. 154, pp. 74–86, 2019.
images can still be identified. In the meantime, the advantage [4] Q. Y. Zhang, Y. Sun, and Y. Yan, “A reversible watermarking algorithm
of the algorithm of the paper is also proven objectively through based on block adaptive compressed sensing,” Dianzi Yu Xinxi Xue-
the comparison of NC values. Therefore, the robustness of the bao/Journal of Electronics and Information Technology, vol. 35, no. 4,
pp. 797–804, 2013.
algorithm in this paper is better than the watermark algorithm [5] W. Hu, R. Zhou, J. Luo, and B. Liu, “Lsbs-based quantum color
of Reference [4] and [9]. images watermarking algorithm in edge region,” Quantum Information
Processing, vol. 18, no. 1, p. 16, 2019.
C. Safety Performance Analysis [6] L. Hsu and H. Hu, “Blind image watermarking via exploitation of inter-
block prediction and visibility threshold in dct domain,” Journal of
The watermark information of the algorithm of the paper is Visual Communication and Image Representation, vol. 32, pp. 130–143,
generated by CS. It can be deemed as a completely reliably 2015.
signal encryption process. Furthermore, the encryption can [7] M. Agoyi, E. Celebi, and G. Anbarjafari, “A watermarking algorithm
based on chirp z-transform, discrete wavelet transform, and singular
be completed without being responsible for the additional value decomposition,” Signal, Image and Video Processing, vol. 9, no. 3,
calculation caused by separate encryption agreement. In the pp. 735–745, 2015.
meantime, it can provide signal compression and encryption. [8] D. Dejey and R. S. Rajesh, “Robust discrete wavelet-fan beam
transforms-based colour image watermarking,” Iet Image Processing,
When the sparse base Ψ and measurement matrix Φ are un- vol. 5, no. 4, pp. 315–322, 2011.
known (non-authorized extraction), the watermark information [9] Q. Zhang, Y. Sun, and Y. Yan, “A reversible watermarking algorithm
cannot be extracted successfully. In addition, the four keys based on block adaptive compressed sensing: A reversible watermarking
algorithm based on block adaptive compressed sensing,” Journal of
must act simultaneously during the extraction procedure. The Electronics Information and Technology, vol. 35, no. 4, pp. 797–804,
watermark information hidden cannot be extracted in case of 2014.

6688

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on May 17,2021 at 11:57:23 UTC from IEEE Xplore. Restrictions apply.
 TABLE II
WATERMARK EXTRACTED AFTER BEING ATTACKED

Attacks Gaussian noise Salt-and-pepper noise Multiplicative noise Random noise

Strength (0.003) (0.02) (0.02) (0.03)

Our Method results

Reference [9] results

Reference [4] results

Attacks Wiener filter Wiener filter Median filter JPEG compression

Strength (3×3) (4×4) (3×3) (80)

Our Method results

Reference [9] results

Reference [4] results

TABLE III
ROBUSTNESS EXPERIMENT RESULTS

Salt-and- Salt-and-
Multiplica-tive Multiplica-tive
Attacks Gaussian noise Gaussian noise pepper pepper
noise noise
noise noise
Strength (0.002) (0.003) (0.01) (0.02) (0.01) (0.02)
Our method 0.9615 0.9226 0.9643 0.9432 0.9591 0.9417
Reference [9] 0.9353 0.8671 0.9314 0.9226 0.9237 0.9185
Reference [4] 0.8726 0.6617 0.9135 0.8753 0.6469 0.5978
JPEG JPEG
Attacks Random noise Median filter Wiener filter Wiener filter
compression compression
Strength (0.03) (33) (33) (44) (80) (85)
Ourmethod 0.7236 0.7987 0.8741 0.8376 0.9208 0.9657
Reference [9] 0.7148 0.7594 0.8657 0.8012 0.8745 0.9293
Reference [4] 0.8653 0.5381 0.7548 0.5283 0.6259 0.7165

[10] N. Kanopoulos, N. Vasanthavada, and R. L. Baker, “Design of an image 2005.

edge detection filter using the sobel operator,” IEEE Journal of Solid- [15] E. J. Candes, J. Romberg, and T. Tao, “Robust uncertainty principles:
state Circuits, vol. 23, no. 2, pp. 358–367, 1988. exact signal reconstruction from highly incomplete frequency informa-
[11] D. Kalman, “A singularly valuable decomposition: The svd of a matrix,” tion,” IEEE Transactions on Information Theory, vol. 52, no. 2, pp.
College Mathematics Journal, vol. 27, no. 1, pp. 2–23, 1996. 489–509, 2006.
[12] R. G. Baraniuk, “Compressive sensing [lecture notes],” IEEE Signal [16] J. A. Tropp and A. C. Gilbert, “Signal recovery from random mea-
Processing Magazine, vol. 24, no. 4, pp. 118–121, 2007. surements via orthogonal matching pursuit,” IEEE Transactions on
[13] W. Dai and O. Milenkovic, “Subspace pursuit for compressive sens- Information Theory, vol. 53, no. 12, pp. 4655–4666, 2007.
ing signal reconstruction,” IEEE Transactions on Information Theory,
vol. 55, no. 5, pp. 2230–2249, 2009.
[14] E. J. Candes and T. Tao, “Decoding by linear programming,” IEEE
Transactions on Information Theory, vol. 51, no. 12, pp. 4203–4215,

6689

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on May 17,2021 at 11:57:23 UTC from IEEE Xplore. Restrictions apply.