Pattern Recognition Letters 145 (2021) 171–177

Contents lists available at ScienceDirect

Pattern Recognition Letters

journal homepage: www.elsevier.com/locate/patrec

Machine learning based blind color image watermarking scheme for

copyright protection ✩
Rishi Sinhal a, Deepak Kumar Jain b, Irshad Ahmad Ansari a,∗
a
Electronics and Communication Engineering, PDPM Indian Institute of Information Technology Design and Manufacturing, Jabalpur - 482005, MP, India
b
Key Laboratory of Intelligent Air–Ground Cooperative Control for Universities in Chongqing, College of Automation, Chongqing University of Posts and
Telecommunications, Chongqing 400065, China

a r t i c l e i n f o a b s t r a c t

Article history: This work presents a blind and robust scheme using YCbCr color space, IWT (integer wavelet transform)
Received 30 April 2020 and DCT (discrete cosine transform) for color image watermarking. During watermark insertion, Y chan-
Revised 18 January 2021
nel is divided into blocks and Mersenne Twister random number generator is used to select the blocks
Accepted 12 February 2021
for embedding. This randomized selection of blocks required a secret key, thus improving the security of
Available online 15 February 2021
the scheme. To reduce the computational complexity, the artificial neural network architecture is devel-
Keywords: oped for watermark embedding. To check the robustness, several signal processing attacks such as JPEG
Robust image watermarking compression, filtering attacks, noise attacks, cropping, resizing and other common attacks are applied
Artificial neural network on the watermarked images. The proposed work is tested on different images to verify the similarity
Machine learning in watermarking results. The scheme provides similar results (having little variation) for different test
YCbCr color space images. Experimental results demonstrate the superior performance in terms of imperceptibility and ro-
bustness. Further, the ANN framework provides faster embedding with approximately similar parametric
results. The performance comparison with existing schemes demonstrates better performance for differ-
ent attacks. The proposed work can be used in robust applications (i.e. copyright protection) for efficient
results and less computational time.
© 2021 Elsevier B.V. All rights reserved.

1. Introduction one of the most shared/communicated digital data types. There-

fore, the digital image watermarking has gained a wide attention
Nowadays, the security of signal processing operations has been of researchers with the development of advanced communication
considered an imperative area of research due to various secu- technologies and signal processing tools.
rity issues. The advancement in the internet services and com- In digital image watermarking, the digital images (i.e. host im-
munication technologies lead to access digital content, news and ages) are modified by embedding a watermark (i.e. digital infor-
other information easily as required. After the invention of Smart- mation) in such a way that the embedded information does not
phone, People are obsessed to online services (e.g. social media, affect the visual appearance of the images [6]. In other words,
online banking, gaming etc.) in such a way that it has been a the embedded information (i.e. watermark) should remain unno-
fundamental need in their lives. Accessibility to the digital con- ticeable. This is an important feature of watermarking and it is
tent become easy and therefore the security of copyrighted dig- known as imperceptibility [7]. The imperceptibility of a water-
ital data is an essential need [1]. Unauthorized access and false marking method is expected to be high and usually represented
ownership claims are prime issues facing by the communication by the watermarking parameters PSNR (Peak Signal to Noise Ra-
industries [2]. Several technologies have been developed in the re- tio) [8] and SSIM (Structural Similarity Index) [9]. In extraction
cent past to secure digital information including digital watermark- process, the watermark information is extracted from the water-
ing [3], stenography [4], cryptography [5] etc. Digital watermark- marked image. Before Extraction, the watermarked image can be
ing is one of the widely acceptable technique by the researchers attacked and hence the extracted watermark may get distorted
to secure digital content (i.e. online or offline). Digital images are (mostly become distorted version of the embedded watermark).
Generally, the extracted watermark is expected to be the same as
the embedded one. This watermarking feature is known as robust-
✩
Edited by Maria De Marsico. ness and expected to be high even after the application of signal
∗
Corresponding author. processing attacks. The robustness of a watermarking method is
E-mail address: irshad@iiitdmj.ac.in (I.A. Ansari).

https://doi.org/10.1016/j.patrec.2021.02.011
0167-8655/© 2021 Elsevier B.V. All rights reserved.
R. Sinhal, D.K. Jain and I.A. Ansari Pattern Recognition Letters 145 (2021) 171–177

represented by the watermarking parameters BER (Bit Error Rate)

[10] and NC (normalized correlation) [11]. The embedding capac-
ity is also an important watermarking feature, which represents
amount of digital information that can be inserted into the digital
image; while maintaining other watermarking features to a signif-
icant level. There are various classifications of digital watermark-
ing on the basis of extraction process, robustness, host type, do- Fig. 1. (a) Lifting operation to decompose the signal. (b) Inverse lifting operation to
main etc. According to the nature of extraction process, digital im- reconstruct the signal.
age watermarking can be classified as blind, semi-blind and non-
blind watermarking. Blind watermarking methods do not need the
➢ Predict: The prediction of odd sample values (Xm,2l+1 ) is done
knowledge of actual host and watermark for the extraction. How-
based on even sample values and prediction factor (F). The dif-
ever, the semi-blind watermarking methods need the watermark
ference Derr between original and predicted value of samples is
information and the non-blind watermarking needs the knowledge
calculated as Derrm−1 = Xm,2l+1 − F [Xm,2l ].
of original signal (i.e. host) for the extraction process. Based on the
➢ Update: The new even samples are produced by updating the
robustness, the digital image watermarking can be classified as ro-
predicted odd samples using an updating operator (G) and
bust, semi-fragile and fragile watermarking methods [12]. The ro-
adding them with the original even samples. Mathematically,
bust watermarking is primarily used for copyright protection, own-
it is given as Xm−1,2lnew = Xm,2l + G [Derrm−1 ].
ership detection etc.; whereas the fragile methods are preferred
for tamper detection [13] and image authentication [14]. The semi- The same steps are simply reversed to perform inverse opera-
fragile watermarking contains features of both (i.e. robust and frag- tion except that the split process is replaced with merging process.
ile) types up to a certain level. The research opportunities in the Fig. 1 demonstrates the lifting and inverse lifting process of a sig-
field of watermarking are continuously increasing with the ad- nal Xm .
vancement in digital media, signal processing tools and commu- Fu and Shen [29] offered a linear discriminant analysis based
nication technologies [15]. watermarking scheme for color images. The watermark (for copy-
In recent past, many image watermarking methods have been right protection etc.) and the reference watermark have been in-
proposed by the researchers to achieve the watermarking objec- serted into the host image during embedding. Even though the
tives; such as copyright protection, ownership security, forgery de- scheme showed good imperceptibility, it lacked in achieving high
tection, self-recovery of the tampered part and protection from robustness against many common signal processing attacks such
unauthorized access [16–18]. As per the watermarking literature, as JPEG compression etc. Wen et al. [30] proposed a robust wa-
various transform domain schemes have been offered that attained termarking method using dual tree complex wavelet transform
significant imperceptibility and robustness by using the properties (DTCWT) and probabilistic neural network (PNN). The selected co-
of respective transforms and human visual systems. It is found efficients in the wavelet domain have been modified for water-
that the transform domain methods give improved robustness as mark embedding. The PNN network has been used for water-
compared to the spatial domain schemes; however, they compro- mark extraction. The scheme has moderate performance in terms
mise the watermarking capacity (i.e. payload) to some extent as of imperceptibility and robustness against different attacks. Peng
compared to the spatial domain schemes. Most of the robust wa- et al. [31] presented an image watermarking technique using the
termarking schemes preferred different transforms for implement- concept of multi-wavelet and support vector machine (SVM). The
ing the embedding and extraction approach because robustness is mean value modulation approach has been employed during wa-
an important feature regarding the security of copyright material. termarking to embed the watermark. The watermark extraction
DWT (Discrete wavelet transform) [19], DCT (Discrete cosine trans- was based on the SVM’s learning ability about the mean value
form) [20], SLT (Slantlet transform) [21], FT (Fourier transform) relationship. As per the experimental results, the scheme has ac-
[22], CT (Contourlet transform) [23], LWT (Lifting wavelet trans- ceptable imperceptibility. However, it has poor robustness against
form) [24] and IWT (Integer wavelet transform) [25] are some of different attacks (such as compression, filtering and scaling etc.).
the transforms, which have been used in the existing watermark- Das et al. [32] proposed a DCT domain-based robust watermark-
ing methods. ing scheme for digital images. In this scheme, the modification of
The conventional transforms consider that the data given is an DCT coefficients of the 8 × 8 size blocks has been done to embed
input contains floating point values but in the image processing the binary watermark into the host image. The scheme provided
operations, the input consists of integer values. The coefficients ob- considerable imperceptibility but the robustness against common
tained after the transform contain floating point values. These co- image processing attacks (e.g. JPEG compression, median filtering
efficients are then processed and finally the inverse operation is and salt and pepper noise etc.) needed improvement.
applied on the wavelet coefficients. During the inversion process, Kalra et al. [33] introduced an image watermarking algorithm
the floating point values are round off to integers. Thus, the in- to process color images based on DCT and DWT. The watermark
formation loss occurs and it degrades the imperceptibility feature. was secured by two encryption approaches. The middle band coef-
The integer wavelet transform is a type of Lifting wavelet trans- ficients of DCT were used for insertion. The watermarking results
form (LWT) [26] which maps integer values to integers with no are acceptable but the application of encryption methods and two
round off error. Therefore, it reduces the computational complexity transform domain approaches make the scheme computationally
and performs more efficiently. These features of IWT transform are complex. The execution time analysis should also be performed.
useful to get high imperceptibility and robustness [27]. There are Pan-Pan et al. [34] presented a robust scheme for color image wa-
3 steps that are used to implement any LW T (e.g. IW T) transform termarking to overcome the effect desynchronization attacks on
named as split, predict and update [28]. These steps are briefly ex- the color image. The color invariance model and probability den-
plained as follows: sity based feature point detector were made by the authors to
make local feature region. To insert the watermark into the host
➢ Split: The signal is first separated into two sections namely image, an invariant bit plane histogram was used. The parametric
even and odd. Let the input signal is Xm and the divided parts results illustrate that the presented approach gives acceptable re-
are even (Xm ) = Xm,2l and odd (Xm ) = Xm,2l+1 . Mathematically, sults but the embedding capacity is very low. This method is useful
it can be written as Xm,2l , Xm,2l+1 ∈ Xm . only for the low capacity applications. Moosazadeh et al. [35] pro-

172
R. Sinhal, D.K. Jain and I.A. Ansari Pattern Recognition Letters 145 (2021) 171–177

posed a YCoCg-R color space and DCT based robust scheme for wa- • Select the Y channel and perform block-wise division to divide
termarking the color images. The binary watermark is encrypted it into uniform 4 × 4 blocks.
using Arnold transform before insertion into the host image. At • Choose the blocks with the help of randomization for embed-
embedding stage, the image is transformed using DCT followed by ding the watermark bits.
block-wise division. The complexity of the blocks gets checked to • Now, sequentially select block to embed a watermark bit watbit
select blocks for embedding. Adaptive embedding strength based into it.
on the energy of the block is used for watermark insertion. Ex- • Perform IWT transform to get the sub bands (LL, LH, HL and
perimental results show the significant parametric values, how- HH) and apply DCT transform on the LL sub band.
ever adaptive embedding strength may increase the time complex- • Select a coefficient (embedcoe f f ) from the DCT matrix and mod-
ity and it should be analyzed. Su and Chen [36] offered a spa- ify it to get newcoe f f as per the Eq. (1).
tial domain based robust watermarking technique for color im-   
ages to protect copyright. The scheme is blind in nature and used (rembed × Q _ f ac ) +  Q _ 2f ac  , i f Econd = 0
newcoe f f = (1)
blue channel to embed the binary watermark. To insert the wa- (rembed × Q _ f ac ) − Q _2f ac , i f Econd = 1
termark, the pixels are modified based on the properties of DC
(Direct current) coefficients. At extraction, the watermark extrac- Where rembed = round ( (embedcoe f f )/Q _ f ac ), Q _ f acis the
tion is done using DC coefficients and the concept of quantiza- quantization factor and Econd has been calculated as Econd =
tion. Simulation results show that the method has good impercep- XOR(mod (rembed , 2 ), watbit ).
tibility but the time complexity and robustness needs further im- Here, mod indicates a modular operation and watbit denotes the
provement. Abdulrahman and Ozturk [37] presented a robust wa- watermark bit to be embedded.
termarking scheme based on DWT and DCT. The Arnold transform • Replace embedcoe f f with the updated coefficient value newcoe f f .
was applied on the watermark to provide security. The host im- • Perform the inverse DCT on the modified DCT matrix to get
age and the encrypted watermark were transformed using DWT
modified LL band. Further perform inverse IWT transform to get
and DCT respectively. Further, the DWT sub bands are modified via
the watermarked block.
DCT coefficient of the equally divided watermark portions. Signifi- • Perform these steps for each selected block to embed all the
cant robustness achieved during experimentation, however the ap-
watermark bits in a sequential manner and get the modified
plication of transforms and Arnold scrambling increase the com-
blocks to get the watermarked Y channel.
putational complexity of the method. Sinhal et al. [21] presented • Finally, convert the updated YCbCr image into the RGB image to
a deep neural network (DNN) based robust watermarking method
get the watermarked image.
for identifying the source of the online digital information. The
scheme used SLT transforms for watermark embedding. An op- During transmission through the communication media, the
timization approach had been used to reduce the computational watermarked image may get distorted (e.g. attacked) via signal
time for embedding. The use of DNN provided fast extraction. Ex- processing attacks such as JPEG compression, filtering etc. There-
perimental performance proved that the scheme has efficient re- fore, various signal processing attacks are considered to evaluate
sults in terms of robustness, imperceptibility and computational the performance of the method.
time. However, the embedding capacity had been compromised to
enhance the reliability. Valandar et al. [38] proposed a robust wa- 2.2. Extraction steps
termarking scheme using different transforms (i.e. DW T, IW T, CT)
and block-wise division. In this blind scheme, the security was ob- • Select the watermarked/attacked image and convert it into
tained by Henon chaotic map and the chaotic nature of the map YCbCr from RGB color space.
is presented by using Lyapunov exponent, bifurcation and trajec- • Choose Y channel and divide it into uniform non-overlapping
tory diagram. During embedding, the image is divided into uniform blocks of 4 × 4 size.
blocks and the transform is applied on blocks. Finally, LL band is • Select blocks used for embedding with the help of secret key
used to embed the watermark bit. The significant imperceptibil- (i.e. randomization) and apply the following steps on each block
ity results were achieved but the robustness against various signal to extract the watermark bit.
processing attacks should be improved. • Perform IWT on the block followed by the DCT transform of the
The objective of this work is to propose a new image water- LL band.
marking scheme that would be useful in copyright protection and • Select the coefficient value (extcoe f f ) of DCT matrix that has
ownership verification in an efficient way. This scheme is robust been used at the embedding stage and extract the watermark
and blind in nature. It uses randomization [39] for selecting the bit (extbit ) as per the Eq. (2).
blocks to embed the watermark bits. The concept of randomization ⎧  
⎪
⎪ 0, i f (extcond = 0 )and extcoe f f > 0
improves the security of the watermarking method and the use ⎨  
1, i f (extcond = 0 ) and extcoe f f ≤ 0
of YCbCr color space offer better visual experience to the viewers.
tbit =   (2)
Further, to improve the computational complexity of the scheme, ⎪0, i f (extcond = 1) and extcoe f f ≤ 0
⎪
⎩
an artificial neural network based embedding framework is pre- 1, i f (extcond = 1 ) and extcoe f f > 0
pared which gives approximately similar parametric values with
less computational time. Where,extcond = mod (rext , 2 ) and rext is obtained as rext =
f ix(extcoe f f /Q _ f ac ).
2. Proposed watermarking method Similarly, extract the watermark bits from selected blocks and
reshape to get the extracted binary watermark.
The embedding and extraction steps of the proposed water-
marking scheme are presented as follows: 3. Experimental results and performance evaluation

2.1. Embedding steps The experimentation is performed with different images to

check the ability of the method for practical use. To perform ex-
• Convert the selected color image from RGB to YCbCr color periments, 80 test color images of size 512 × 512 and four 32 × 32
space. size binary watermarks are used [40]. For succinct representation

173
R. Sinhal, D.K. Jain and I.A. Ansari Pattern Recognition Letters 145 (2021) 171–177

ages have been used to compare the performance in terms of im-

perceptibility and robustness. Table 3 compares the imperceptibil-
ity parameters (PSNR, SSIM) with the existing schemes. Although
the PSNR values are lower as compared to [36] and [38], the val-
ues are almost similar to the schemes [32] and [33]. Although the
PSNR of the proposed scheme is less than the other schemes, yet it
has a significant and acceptable value to maintain the visual qual-
ity. Moreover, the proposed scheme has less variation in the PSNR
values (obtained for different images) as compared to the schemes
[33], and [38]. This shows that the scheme provides more consis-
tent results than the existing schemes. On the other side, the pro-
posed scheme proves to be better as compared to other schemes
in terms of structural similarity (SSIM). SSIM relates the structural
Fig. 2. Host and watermark images. similarity of images, that is an effective way to compare images
[41].
Table 1 The performance comparison in terms of robustness has been
Imperceptibility results (PSNR, SSIM) for different presented in Table 4. Four host images (Lena, Mandrill, Plane, Pep-
host images. per) have been used to compare the normalized correlation (NC)
S. No. Host Image PSNR (in dB) SSIM of the extracted watermarks. The high NC values are obtained as
compared to the other schemes [32,33], and [36] against most of
1 Lena 40.1304 0.9976
2 Mandrill 39.7387 0.9961
the attacks. However, the performance of the proposed scheme is
3 Plane 40.0244 0.9700 comparatively low in case of JPEG compression attacks with qual-
4 Pepper 40.2522 0.9976 ity factor equal to 10. Robustness comparison (as given in Table 4)
5 Sailboat 40.0188 0.9931 shows that the proposed scheme gives superior/equivalent results
6 House 39.8059 0.9892
as compared to the existing methods. The parametric comparison
7 Barbara 39.8599 0.9920
shows that the proposed scheme can withstand more efficiently
against attacks with improved visual quality than the existing tech-
niques.
in this manuscript, a binary watermark image and seven host im- Furthermore, the statistical comparison has been done to check
ages are selected and shown in Fig. 2. In this method, the values of the overall robustness performance of the proposed scheme as
some embedding parameters such as quantization factor, position compared to the recently offered scheme [38]. All the test images
of embedding coefficient and sub-band (i.e. LL) are decided after (i.e. 80 color images) and four watermark images have been used
the wide-ranging experimentation on different images to get effec- to obtain the NC values (using the proposed scheme) against each
tive results. The quantization factor is fixed to 32 and the position attack presented in Table 5. It can be observed from Table 5 that
of embedding coefficient in the DCT matrix is (2, 1). To check the the proposed scheme has better results than the scheme [38] for
performance of the method, some watermarking parameters like most attacks. For a few attacks such as GF (2 × 2), GN (0.01), S&P,
PSNR, SSIM, BER and NC are calculated. resizing, and MF (3 × 3), the proposed scheme provides slightly
Eighty host images and four watermark images have been used lower but acceptable NC values than [38].
to calculate the statistical results presented in this manuscript. Af- In addition, the time complexity of the scheme has been an-
ter thorough experimentation, it is found that the average val- alyzed to check its viability for practical applications that need
ues of PSNR and SSIM are 40.025 and 0.9907 respectively. Table 1 quick processing. The average embedding time is 2.23 s and the av-
presents the imperceptibility results for the selected host and wa- erage extraction time is 1.36 s. Although the computational times
termark images. The standard deviation and variance in the PSNR of both the approaches are less, sometimes it is not considerable
values are found to be 0.1828 and 0.0334 respectively. The varia- for practical applications that require fast response. Therefore, a
tions in parametric values are very less for different images and neural network framework is presented to reduce the computa-
watermarks, which signify that the proposed method can be used tional complexity up to an acceptable limit for practical applica-
to achieve consistent performance for different type of images. tions with approximately the same imperceptibility and robust-
The robustness is evaluated by measuring BER and NC values ness. The neural network [42] is a type of machine learning mech-
for the extracted watermark. Table 2 illustrates the robustness re- anism that can be trained using the input and target dataset. The
sults for different watermarked images against various signal pro- large training dataset should be used to train the neural network.
cessing attacks. Different signal/ image processing attacks includ- It helps the neural network to efficiently learn the relationship be-
ing speckle noise (SN), Gaussian noise (GN), salt and pepper noise tween the input dataset and the target dataset [42,43]. After train-
(S&P), JPEG and JPEG 20 0 0 compression, median filtering (MF), ing, it can be used to achieve effective results for the new data (i.e.
Gaussian filtering (GF), Wiener filtering (WF), and other attacks digital signals/images, etc.).
(such as cropping, resizing, blurring, and gamma correction, etc.)
have been applied on the watermarked images. The obtained re- 3.1. Embedding via neural network framework
sults during experimentation confirmed that the proposed scheme
performs well against most of the attacks like JPEG compression Fig. 3. represents the proposed artificial neural network hav-
attacks, noise attacks; filtering, resizing, cropping and other com- ing single hidden layer with 10 neurons. This framework accepts
mon attacks. The proposed method achieves the significant im- 20 values as the input that contains 16 pixel values and 4 copies
perceptibility and robustness. The obtained results signify that the of watermark bit to be inserted into the block. At output end, 16
method can be used for robust applications in an efficient manner. modified pixel values are obtained. The proposed neural network
The comparison of parametric values with other existing framework for embedding has been trained by a dataset made by
schemes has been done to judge the performance of the proposed 100 different types of images. The proposed ANN network reduces
scheme as compared to other robust watermarking schemes having the embedding time and gives approximately same parametric val-
same watermarking payload (i.e.1024 bits). Some of the test im- ues. In the proposed ANN architecture, 4 × 4 size blocks and the 4

174
R. Sinhal, D.K. Jain and I.A. Ansari Pattern Recognition Letters 145 (2021) 171–177

Table 2
BER, NC for different signal processing attacks.

BER, NC
Signal processing
Lena Mandrill Plane Pepper Sailboat House Barbara
S. No. Attacks
BER NC BER NC BER NC BER NC BER NC BER NC BER NC

1 Attack free 0 1 0 1 0 1 0 1 0 1 0 1 0.000 1.000

Noise attacks
2 Speckle noise (0.01) 0.001 0.998 0 1 0.003 0.994 0 1 0.001 0.998 0.001 0.998 0 1
3 Speckle noise (0.02) 0.008 0.984 0.008 0.984 0.038 0.922 0.021 0.958 0.024 0.950 0.027 0.944 0.007 0.986
4 Speckle noise (0.05) 0.067 0.863 0.074 0.849 0.173 0.647 0.091 0.816 0.088 0.820 0.134 0.727 0.056 0.887
5 Gaussian noise (0.01) 0.019 0.962 0.022 0.956 0.021 0.958 0.026 0.946 0.026 0.946 0.023 0.952 0.023 0.952
6 Gaussian noise (0.02) 0.113 0.770 0.121 0.753 0.109 0.777 0.112 0.771 0.099 0.799 0.111 0.773 0.125 0.745
7 Gaussian noise (0.05) 0.278 0.438 0.295 0.408 0.273 0.446 0.257 0.480 0.254 0.487 0.291 0.414 0.287 0.422
8 Salt & pepper (0.01) 0.003 0.994 0.001 0.998 0.005 0.990 0.005 0.990 0.006 0.988 0.003 0.994 0.003 0.994
9 Salt & pepper (0.02) 0.026 0.946 0.012 0.976 0.023 0.952 0.028 0.942 0.024 0.950 0.019 0.962 0.015 0.970
10 Salt & pepper (0.05) 0.104 0.789 0.088 0.822 0.109 0.778 0.104 0.788 0.110 0.778 0.106 0.785 0.096 0.805
JPEG attacks
11 JPEG (10) 0.320 0.337 0.174 0.640 0.263 0.473 0.291 0.402 0.241 0.506 0.227 0.536 0.246 0.497
12 JPEG (20) 0.036 0.927 0.015 0.970 0.041 0.917 0.038 0.922 0.025 0.948 0.032 0.935 0.023 0.954
13 JPEG (30) 0 1 0.002 0.996 0 1 0.001 0.998 0 1 0.001 0.998 0 1
14 JPEG (40) 0 1 0 1 0 1 0 1 0 1 0 1 0 1
15 JPEG (50) 0 1 0 1 0 1 0 1 0 1 0 1 0 1
16 JPEG (60) 0 1 0 1 0 1 0 1 0 1 0 1 0 1
17 JPEG 2000 (CR=10) 0 1 0 1 0 1 0 1 0.001 0.998 0 1 0 1
18 JPEG 2000 (CR=20) 0 1 0.053 0.892 0 1 0 1 0.004 0.992 0.007 0.986 0.003 0.994
19 JPEG 2000 (CR=30) 0.007 0.986 0.128 0.737 0.006 0.988 0.010 0.980 0.123 0.746 0.088 0.819 0.021 0.958
20 JPEG 2000 (CR=40) 0.027 0.944 0.217 0.555 0.021 0.958 0.045 0.908 0.140 0.712 0.173 0.643 0.178 0.639
Filtering attacks
21 Median filter (2 × 2) 0.025 0.948 0.162 0.671 0.059 0.880 0.032 0.934 0.078 0.841 0.076 0.845 0.047 0.904
22 Median filter (3 × 3) 0.011 0.978 0.138 0.719 0.026 0.946 0.014 0.972 0.050 0.899 0.040 0.918 0.034 0.930
23 Gaussian filter (2 × 2) 0.020 0.960 0.151 0.692 0.062 0.874 0.033 0.932 0.077 0.843 0.082 0.833 0.042 0.914
24 Gaussian filter (3 × 3) 0 1 0.002 0.996 0 1 0 1 0 1 0 1 0 1
25 Wiener filter (2 × 2) 0 1 0.018 0.964 0 1 0 1 0 1 0 1 0 1
26 Wiener filter (3 × 3) 0 1 0.040 0.918 0.004 0.992 0.001 0.998 0.000 1.000 0.007 0.986 0.005 0.990
Other attacks
27 Crop 5% pixel@center 0.004 0.992 0.004 0.992 0.005 0.990 0.004 0.992 0.005 0.990 0.005 0.990 0.004 0.992
28 Crop 10% pixel@center 0.008 0.984 0.008 0.986 0.008 0.984 0.008 0.984 0.008 0.984 0.010 0.980 0.008 0.984
29 Contrast adjust (20%) 0.002 0.996 0.003 0.968 0.048 0.904 0.003 0.994 0.022 0.956 0.484 0.021 0.491 0.005
30 Blurring (θ −7, len=10) 0.066 0.865 0.068 0.572 0.110 0.773 0.068 0.859 0.137 0.724 0.022 0.956 0.005 0.990
31 Resize(1–0.5–1) 0.005 0.990 0.008 0.821 0.021 0.958 0.008 0.984 0.022 0.956 0.132 0.733 0.079 0.837
32 Resize(1–2–1) 0 1 0 1 0 1 0 1 0 1 0.023 0.952 0.009 0.982
33 Gamma 0.002 0.996 0.006 0.992 0.015 0.970 0.006 0.988 0.007 0.986 0 1 0 1
correlation(0.6)
34 Gamma 0 1 0 1 0 1 0 1 0 1 0.006 0.988 0.016 0.968
correlation(0.8)

Table 3
Performance comparison of imperceptibility results (PSNR, SSIM).

Schemes
Images
[32] [33] [36] [38] Our

Lena 41.78, 0.9704 42.01, 0.9787 49.98, 0.9872 55.19, 0.9964 40.13, 0.9976
Mandrill 40.24, 0.9893 36.11, 0.9754 49.89, 0.9957 51.29, 0.9944 39.74, 0.9961
Plane 40.79, 0.9872 39.76, 0.9851 49.86, 0.9854 50.00, 0.9919 40.02, 0.9700
Pepper 41.01, 0.9733 42.68, 0.9816 50.08, 0.9859 52.48, 0.9964 40.25, 0.9976

• RGB image has been transformed to the YCbCr space and Y

channel is selected for embedding.
• Y channel has been broken into uniform 4 × 4 size blocks and
the blocks are selected for embedding in a randomized manner
using secret key.
• To embed the watermark bits in the selected blocks via neural
network, prepare the dataset using the selected blocks and wa-
Fig. 3. Proposed ANN architecture.
termark bits (using the embedding process (as in Section 2.1)).
• The dataset has been applied to the network to get the water-
marked blocks.
• Convert the modified YCbCr color space into an RGB image to
copies of the watermark bit (that would be inserted into the block) get the watermarked image.
are used to get the 20 values as an input to modify pixel values of
a block. Similarly, all selected blocks would be updated. The water- During extraction stage, the same procedure is used as pre-
marking process using ANN framework is presented in the follow- sented in the Section 2.2. Experimental results obtained from ANN
ing steps: framework are shown in Table 6. To compare the results, average

175
R. Sinhal, D.K. Jain and I.A. Ansari Pattern Recognition Letters 145 (2021) 171–177

Table 4
Performance comparison of robustness results (NC).

Images

Lena Mandrill Plane Pepper

Attacks
[32] [33] [36] Our [32] [33] [36] Our [32] [33] [36] Our [32] [33] [36] Our

S&P (0.002) 0.81 1.00 1.00 1.00 0.82 0.99 0.99 1.00 0.81 1.00 1.00 1.00 0.80 1.00 1.00 1.00
S&P (0.004) 0.78 0.96 0.97 1.00 0.77 0.98 0.98 1.00 0.79 0.98 0.98 1.00 0.78 0.98 0.99 0.99
S&P (0.006) 0.76 0.92 0.94 0.99 0.77 0.96 0.97 1.00 0.77 0.97 0.98 1.00 0.76 0.96 0.97 1.00
S&P (0.008) 0.72 0.90 0.90 1.00 0.73 0.95 0.96 0.99 0.72 0.93 0.93 1.00 0.71 0.95 0.94 0.99
S&P (0.010) 0.67 0.89 0.90 0.99 0.70 0.89 0.90 0.99 0.71 0.89 0.89 0.99 0.79 0.92 0.94 0.99
S&P (0.012) 0.65 0.84 0.87 0.97 0.67 0.89 0.91 0.98 0.68 0.84 0.86 0.96 0.71 0.86 0.91 0.95
MF (2 × 2) 0.83 0.93 0.95 0.95 0.70 0.75 0.74 0.67 0.85 0.95 0.95 0.88 0.81 0.93 0.94 0.93
MF (3 × 3) 0.91 0.95 1.00 0.98 0.88 0.75 0.88 0.72 0.94 0.99 1.00 0.95 0.92 0.94 1.00 0.97
JPEG (10) 0.62 0.64 0.77 0.34 0.50 0.80 0.58 0.64 0.28 0.63 0.21 0.47 0.57 0.60 0.86 0.40
JPEG (20) 0.72 0.72 0.73 0.93 0.49 0.80 0.60 0.97 0.42 0.70 0.39 0.92 0.61 0.78 0.80 0.92
JPEG (30) 0.67 0.85 0.67 1.00 0.51 0.86 0.66 1.00 0.57 0.84 0.68 1.00 0.65 0.87 0.65 1.00
JPEG (40) 0.62 0.98 0.67 1.00 0.68 0.96 0.72 1.00 0.63 0.96 0.72 1.00 0.66 0.96 0.71 1.00
JPEG (50) 0.68 1.00 0.72 1.00 0.67 0.99 0.70 1.00 0.67 0.97 0.70 1.00 0.70 0.98 0.67 1.00
JPEG (60) 0.68 1.00 0.74 1.00 0.73 1.00 0.75 1.00 0.71 1.00 0.74 1.00 0.73 0.99 0.79 1.00
JPEG (70) 0.80 1.00 0.77 1.00 0.79 1.00 0.77 1.00 0.79 1.00 0.81 1.00 0.77 1.00 0.83 1.00
JPEG (80) 0.87 1.00 0.91 1.00 0.84 1.00 0.94 1.00 0.87 1.00 0.93 1.00 0.86 1.00 0.95 1.00
JPEG (90) 1.00 1.00 0.99 1.00 0.98 1.00 0.98 1.00 0.98 1.00 0.98 1.00 0.94 1.00 1.00 1.00
JPEG (100) 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.99 1.00 1.00 1.00 0.97 1.00 1.00 1.00
Resize(50%) 0.96 0.98 1.00 0.99 0.92 0.94 0.99 0.82 0.97 0.96 0.99 0.96 0.99 0.97 1.00 0.98
Resize(200%) 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00

Table 5 posed scheme has robust and blind nature. It used the random-
Statistical comparison in terms of average NC values.
ization process based on Mersenne Twister random generator for
Attacks [38] our Attacks [38] our selecting the blocks to embed the watermark bits. Further, an ANN
No attack 1.00 1.00 Blur (2.5,50) 0.90 0.96 framework is developed to reduce the computational complexity
GF (2 × 2) 1.00 0.90 Blur (1.5,50) 0.98 1.00 of embedding process for speedy response. Experimental results
GF (3 × 3) 0.99 1.00 Blur (0.5,50) 1.00 1.00 proved the significance of the method as compared to the other re-
GF (5 × 5) 0.96 1.00 Resize (0.5) 0.98 0.97 cently developed robust watermarking schemes. Future work will
GC (1.5) 0.91 0.99 MF (3 × 3) 0.99 0.95
include the development of more efficient neural network archi-
GC (1.0) 1.00 1.00 WF (2 × 2) 1.00 1.00
GC (0.5) 0.83 0.83 WF (3 × 3) 0.99 0.99 tecture to achieve optimal performance in terms of watermarking
GN (0.001) 0.99 1.00 JPEG 90 0.96 1.00 parameters and improving the embedding strategy. The extraction
GN (0.01) 0.98 0.93 JPEG 80 0.95 1.00 process will also be investigated using neural network architec-
S&P (0.01) 0.99 0.98 JPEG 70 0.90 1.00
tures to improve the performance.
S&P (0.02) 0.97 0.94 SN (0.001) 0.96 1.00
S&P (0.05) 0.82 0.79 SN (0.005) 0.94 1.00
Declaration of Competing Interest

Table 6
The authors declare that they have no known competing finan-
Performance comparison with and without ANN framework.
cial interests or personal relationships that could have appeared to
Host Without ANN influence the work reported in this paper.
Image
Avg. Avg.
Embed Embed Acknowledgments
time (in time (in
PSNR Avg. BER sec) PSNR Avg. BER sec)
This work was supported by Faculty Initiation Grant of PDPM
Lena 40.13 0.0476 1.93 39.90 0.0515 0.41 Indian Institute of Information Technology Design and Manufactur-
Mandrill 39.74 0.0668 1.91 37.47 0.0991 0.42
ing Jabalpur, India.
Plane 40.02 0.0554 1.94 39.78 0.0685 0.41
Pepper 40.25 0.0493 1.93 39.82 0.0527 0.38
sailboat 40.02 0.0595 1.92 39.39 0.0731 0.40 References
House 39.81 0.0609 1.91 39.28 0.0718 0.39
Barbara 39.86 0.0520 1.92 38.73 0.0600 0.40 [1] L. Me, G.R. Arce, A class of authentication digital watermarks for secure multi-
media communication, IEEE Trans. Image Process. 10 (11) (2001) 1754–1764.
[2] Y. Wang, J.F. Doherty, R.E. Van Dyck, A wavelet-based watermarking algorithm
for ownership verification of digital images, IEEE Trans. Image Process. 11 (2)
(2002) 77–88.
BER is calculated for 34 different signal processing attacks. From
[3] R. Liu, T. Tan, An SVD-based watermarking scheme for protecting rightful own-
Table 6, it is observed that the embedding time has been reduced ership, IEEE Trans. Multimed. 4 (1) (2002) 121–128.
significantly by using ANN framework and the degradation in other [4] I. Cox, M. Miller, J. Bloom, J. Fridrich, T. Kalker, Digital Watermarking and
parametric values is very less. Therefore, the ANN framework can Steganography, Morgan kaufmann, 2007.
[5] C.C. Chang, H.C. Wu, A copyright protection scheme of images based on visual
be used effectively for practical applications that required fast re- cryptography, Imag. Sci. J. 49 (3) (2001) 141–150.
sponse . [6] P.T. Yu, H.H. Tsai, J.S. Lin, Digital watermarking based on neural networks for
color images, Signal Process. 81 (3) (2001) 663–671.
[7] M. Khalili, D. Asatryan, Colour spaces effects on improved discrete wavelet
4. Conclusion transform-based digital image watermarking using Arnold transform map, IET
Signal Process. 7 (3) (2013) 177–187.
[8] Y. Huang, B. Niu, H. Guan, S. Zhang, Enhancing image watermarking with adap-
In this study, an improved color image watermarking based on tive embedding parameter and psnr guarantee, IEEE Trans. Multimed. 21 (10)
machine learning and transform domain was elaborated. The pro- (2019) 2447–2460.

176
R. Sinhal, D.K. Jain and I.A. Ansari Pattern Recognition Letters 145 (2021) 171–177

[9] Z. Wang, A.C. Bovik, H.R. Sheikh, E.P. Simoncelli, Image quality assessment: [26] N.M. Makbol, B.E. Khoo, A new robust and secure digital image watermarking
from error visibility to structural similarity, IEEE Trans. Image Process. 13 (4) scheme based on the integer wavelet transform and singular value decompo-
(20 04) 60 0–612. sition, Digit. Signal Process. 33 (2014) 134–147.
[10] S. Thakur, A.K. Singh, B. Kumar, S.P. Ghrera, Improved DWT-SVD-based med- [27] I. Daubechies, W. Sweldens, Factoring wavelet transforms into lifting steps, J.
ical image watermarking through hamming code and chaotic encryption, in: Fourier Anal. Appl. 4 (3) (1998) 247–269.
Advances in VLSI, Communication, and Signal Processing, Springer, Singapore, [28] V.S. Verma, R.K. Jha, A. Ojha, Significant region based robust watermarking
2020, pp. 897–905. scheme in lifting wavelet transform domain, Expert Syst. Appl. 42 (21) (2015)
[11] C.S. Hsu, S.F. Tu, Enhancing the robustness of image watermarking against 8184–8197.
cropping attacks with dual watermarks, Multimed. Tool. Appl. (2019) 1–27. [29] Y.G. Fu, R.M. Shen, Color image watermarking scheme based on linear discrim-
[12] I.J. Cox, M.L. Miller, J.A. Bloom, C. Honsinger, Digital Watermarking, Morgan inant analysis, Comput. Stand. Interface. 30 (3) (2008) 115–120.
Kaufmann, San Francisco, 2002 (Vol. 53). [30] X.B. Wen, H. Zhang, X.Q. Xu, J.J. Quan, A new watermarking approach based on
[13] P.L. Lin, C.K. Hsieh, P.W. Huang, A hierarchical digital watermarking method probabilistic neural network in wavelet domain, Soft. Comput. 13 (4) (2009)
for image tamper detection and recovery, Pattern Recognit. 38 (12) (2005) 355–360.
2519–2529. [31] H. Peng, J. Wang, W. Wang, Image watermarking method in multiwavelet
[14] C.S. Lu, H.Y. Liao, Multipurpose watermarking for image authentication and domain based on support vector machines, J. Syst. Softw. 83 (8) (2010)
protection, IEEE Trans. Image Process. 10 (10) (2001) 1579–1592. 1470–1477.
[15] M.M. Yeung, Digital watermarking, Commun. ACM 41 (7) (1998) 31–33. [32] C. Das, S. Panigrahi, V.K. Sharma, K.K. Mahapatra, A novel blind robust image
[16] A. Khan, F. Jabeen, F. Naz, S. Suhail, M. Ahmed, S. Nawaz, Buyer seller water- watermarking in DCT domain using inter-block coefficient correlation, AEU-Int.
marking protocols issues and challenges–A survey, J. Netw. Comput. Appl. 75 J. Electron. Commun. 68 (3) (2014) 244–253.
(2016) 317–334. [33] G.S. Kalra, R. Talwar, H. Sadawarti, Adaptive digital image watermarking for
[17] A.S. Panah, R. Van Schyndel, T. Sellis, E. Bertino, On the properties of non-me- color images in frequency domain, Multimed. Tool. Appl. 74 (17) (2015)
dia digital watermarking: a review of state of the art techniques, IEEE Access 6849–6869.
4 (2016) 2670–2704. [34] N. Pan-Pan, W. Xiang-Yang, L. Yu-Nan, Y. Hong-Ying, A robust color image wa-
[18] M.A. Nematollahi, C. Vorakulpipat, H.G. Rosales, Security enhancement of termarking using local invariant significant bitplane histogram, Multimed. Tool.
digital watermarking, in: In Digital Watermarking, Springer, Singapore, 2017, Appl. 76 (3) (2017) 3403–3433.
pp. 191–203. [35] M. Moosazadeh, G. Ekbatanifard, An improved robust image watermark-
[19] X. Kang, J. Huang, Y.Q. Shi, Y. Lin, A DWT-DFT composite watermarking scheme ing method using DCT and YCoCg-R color space, Optik (Stuttg) 140 (2017)
robust to both affine transform and JPEG compression, IEEE Trans. Circuit. Syst. 975–988.
Video Technol. 13 (8) (2003) 776–786. [36] Q. Su, B. Chen, Robust color image watermarking technique in the spatial do-
[20] M.A. Suhail, M.S. Obaidat, Digital watermarking based DCT and JPEG model, main, Soft. Comput. 22 (1) (2018) 91–106.
IEEE Trans. Instrum. Meas. 52 (5) (2003) 1640–1647. [37] A.K. Abdulrahman, S. Ozturk, A novel hybrid DCT and DWT based robust wa-
[21] R. Sinhal, I.A. Ansari, D.K. Jain, Real-time watermark reconstruction for the termarking algorithm for color images, Multimed. Tool. Appl. 78 (12) (2019)
identification of source information based on deep neural network, J. Real– 17027–17049.
Time Image Process. (2019) 1–19. [38] M.Y. Valandar, M.J. Barani, P. Ayubi, A blind and robust color images water-
[22] R.N. Bracewell, R.N. Bracewell, The Fourier transform and Its Applications, Mc- marking method based on block transform and secured by modified 3-dimen-
Graw-Hill, New York, 1986 (Vol. 31999). sional Hénon map, Soft. Comput. 24 (2) (2020) 771–794.
[23] M.N. Do, M. Vetterli, The contourlet transform: an efficient directional mul- [39] M. Saito, M. Matsumoto, Variants of Mersenne twister suitable for graphic pro-
tiresolution image representation, IEEE Trans. Image Process. 14 (12) (2005) cessors, ACM Trans. Math. Softw. (TOMS) 39 (2) (2013) 1–20.
2091–2106. [40] “Image databases”, Accessed: Sep. 2020. [Online]. Available: http://www.
[24] B. Lei, Y. Soon, F. Zhou, Z. Li, H. Lei, A robust audio watermarking scheme based imageprocessingplace.com/root_files_V3/image_databases.htm.
on lifting wavelet transform and singular value decomposition, Signal Process. [41] Hore, A., & Ziou, D. (2010, August). Image quality metrics: PSNR vs. SSIM. In
92 (9) (2012) 1985–2001. 2010 20th International Conference on Pattern Recognition (pp. 2366–2369).
[25] I.A. Ansari, M. Pant, C.W. Ahn, Robust and false positive free watermarking IEEE.
in IWT domain using SVD and ABC, Eng. Appl. Artif. Intell. 49 (2016) 114– [42] M.H. Hassoun, Fundamentals of Artificial Neural Networks, MIT press, 1995.
125. [43] B. Yegnanarayana, Artificial Neural Networks, PHI Learning Pvt. Ltd, 2009.

177