A New Blind Invisible and Semi-Fragile Colour

Image Watermarking Scheme in Spatial Domain

Joshua Roy Palathinkal1 , Yuvam Bhateja1 , Sudip Ghosh2 , Hafizur Rahaman3

Abstract—As copyright protection is vital and indispensable, 2) Content authentication: Special features are embedded in-
watermarking schemes are employed for securing digital data. side the data. When suspicion for malicious modifications
We propose a novel (to the best of our knowledge) blind and arises, these features can be extracted and compared to
invisible, semi-fragile image watermarking scheme for embed-
ding coloured watermark inside a coloured cover image. The authenticate the content [14].
spatial domain embedding is implemented in such a way that 3) Tamper recovery: Unauthorized tampering or illegal mod-
modification made in the pixel values of the cover image is not ification of data is detectable, and recovery is possible.
noticeable by human perception. However, the watermark pixel Recovery techniques have been explored in [14, 11].
values can be recovered using a unique decoding mechanism,
which has been included in the paper. Extraction of the original Digital watermarking has a wide range of applications, in-
cover image using the embedded image is also discussed. The cluding those involving highly sensitive information that must
obtained MATLAB implementation results are promising when be disseminated, such as medical records, electronic billing
compared to similar state-of-the-art research work. transactions, and military information [5].
Index Terms—Color Image watermarking, Invisible and Blind Digital watermarking is the process of embedding a digital
Watermarking, Semi-fragile Watermarking, MATLAB Imple-
mentation. watermark into a larger set of data known as cover. Different
types of watermarking techniques have distinct properties of
I. I NTRODUCTION characteristics. It can be implemented using the visible water-
marking scheme where the incorporated watermark is visible
The age we live in is the ’Age of Big Data’. From a along with the cover data or using the invisible watermarking
total of 2.62 exabytes (=1018 bytes) of data existing in 1986, scheme where the incorporated watermark data is impercep-
today’s age creates 2.5 exabytes of data everyday [13]. These tible to the viewer. Fragile watermarks are designed to be
large amounts of data consist in the form of text, images, vulnerable to all modifications, and the extraction algorithm
audios [8] and videos. Cloning, modification and distribution fails if the slightest data manipulation is made to embedded
of data stored in non-erasable devices used to be difficult data. Semi-fragile watermarks provide selective robustness to
due to the inadequacy of the technology to handle digital a specific set of manipulations which are considered as legiti-
data. But, recent advancements in devices, processor and mate and allowable [16]. A non-blind watermarking techniques
communication technologies has brought paradigm shifts in require a host image (usually the cover) for extraction of the
the way we handle, use and transfer data. The invention of the watermark, while blind watermarking techniques don’t require
re-writable magneto-optical disk in 1991 by IBM has been one it [2, 3].
of the significant milestones achieved in the history of data. As discernible from the properties, the proposed algorithm
These technological advances in the wrong hands give the can incorporate all the most desirable characteristics, at the
ability for individuals to have illegal access, data content same time can provide results which are optimized in terms
modification, and republication of illegitimate copies of data. of computational complexity and quality of the image.
Due to these, encryption and obfuscation [5] of digital data The rest of the paper is organized as follows. In Section
is inevitable. Digital Watermarking is one of the techniques II, the proposed algorithm is discussed in detail. Section III
employed to protect the interest of content providers. It is provides the experimental results obtained, with it’s analysis.
used primarily for three reasons [1]: Finally, the conclusion is given in Section IV, followed by
1) Copyright protection: The content owner embeds a unique Bibliography.
identification mark on the data, which is later on retrieved
to prove ownership rights. II. P ROPOSED WATERMARK A LGORITHM
We conceal the watermark image by embedding it into the
1 Joshua & Yuvam are undergraduates in Electronics & Telecommunica- pixel values of the cover image by mathematical manipulations
tion Engineering Department at Indian Institute of Engineering Science &
Technology (IIEST), Shibpur.
in such a way that both the watermark and cover is retrievable
2 Dr. Sudip Ghosh is Assistant Professor with the School of VLSI Technol- later with the use of a secret key known to the content
ogy (SOVT), IIEST Shibpur. owner. Only the possession of the key can assist in obtaining
3 Prof. Hafizur Rahaman is the HOD of Information Technology Department
at IIEST Shibpur.
the watermark which was embedded inside the cover image.
Different schemes differ in the way the key is represented.
In the proposed scheme, the key is analogous to the seed
978-1-7281-6564-6/20/$31.00 ©2020 IEEE matrix used for generation of LFSR (Linear-feedback shift

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 03,2021 at 16:51:05 UTC from IEEE Xplore. Restrictions apply.
 Figure 2. Embedding & extraction process

Figure 1. Preprocessing of cover and watermark image

are partitioned into non-overlapping sub-blocks of size (8×8).
Step 2: Partitioning of sub-blocks is done into partitions
register) sequence in the embedding process. Given the knowl- of (4 × 4) size. We denote these partitions as F = {Fi where
edge of the seed matrix, we can decipher the watermark and i ∈ 1, 2, 3, 4}.
cover with an appropriate decoding mathematical algorithm. • W atermark Image :
In traditional watermarking, the decoding mechanism is the Step 1: We partition each colour-plane of the watermark
reverse-process of the embedding process. This raises security image into non-overlapping blocks of size (2 × 2). The 4
and vulnerability issues, thus prompting us to propose a decod- elements obtained are then arranged in a matrix form as shown
ing mechanism which is entirely different and more complex in Figure.1, with the MSB being in the 1st column, and the
than the embedding process. This gives us the advantage of LSB in the last.
fast encryption, while the embedded content is invulnerable to Step 2: Neglecting the first 4 least-significant bits, we obtain
unauthorised extraction. a (4 × 4) matrix, with 1-bit as each element. We partition this
We divide the algorithm into three parts: 4x4 matrix into non-overlapping (2 × 2) matrix, where each
P re − processing P hase −→ W atermark Embedding −→ elements are denoted as S = {Si where i ∈ 1, 2, 3, 4}.
Extraction P rocess
B. Watermark Embedding: The watermark embedding phase
A. Pre-processing Phase: The pre-processing phase is shown is shown in Figure.2
in Figure.1. • Linear F eedback Shif t Register :
• Cover Image : LFSRs are used to design hardware for PRBS (pseudo-
Step 1: We partition each colour-plane of the cover image random bit sequence) generation. LFSR circuits generate high-
into non-overlapping blocks, each of size (16 × 16). Blocks speed binary sequences using minimum hardware. Given an

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 03,2021 at 16:51:05 UTC from IEEE Xplore. Restrictions apply.
L-bit seed, the LFSR is capable of generating a PRBS of 2L −1
bits long.
Step 1: We select four 4-bit secret-key (manually). We use
these keys as the seed for the LFSR circuit and generate four
15-bits sequence.
Step 2: We pad the 16th-bit of each sequence with the
1st-bit, to make the sequences 16-bit long, thus enabling us
to reshape the sequences to create four (4 × 4) binary-valued
matrices.
Alternative: The proposed algorithm is flexible for the use
of other Random Sequence Generator, such as Gollman Cas-
cade Filtered-Feedback with Carry Shift Register (F-FCSR)
[15], for providing more complex sequences, with the trade-
off of increment in hardware requirements.
• Embedding M athematical Operation :
Step 1: The data embedding rule is expressed as follows:

if Si = 1 ⇒ Fi0 = Fi + KLi
(1)
if Si = 0 ⇒ Fi0 = Fi − KLi

where K is the modulation index. This is repeated until the

whole image is processed upon. Figure 3. Reconstruction of extracted watermark
Step 2: We re-construct the Fi0 (as we have fragmented
Fi in the pre-processing phase) to acquire the watermarked
image. where i0 ∈ {1, 2}, j 0 ∈ {1, 2} and d is a (4 × 4) matrix
Alternative: Scalability can be achieved in this algorithm by consisting of 1-bit elements. We compute steps 1-3 thrice again
varying the size of 1) cover image 2) watermark image and the to obtain four d matrices, denoted by di where i ∈ {1, 2, 3, 4}.
3) LFSR matrix. Watermarking on different sized cover and Step 5: We concatenate the four di ’s to obtain a (4 × 4)
watermark images were implemented successfully, and minor matric, and pad it with a random (4×4) LFSR matrix to obtain
changes in the results have been observed. a (8 × 8) matric containing values of 4 watermark pixels. The
process is depicted in Figure.3.
C. Extraction Process: Due to the flexibility of the proposed Step 6: We reconstruct the obtained watermark pixels
algorithm, it is possible to extract both the watermark and systematically to obtain the watermark.
cover. • Cover Extraction : The embedded image is passed
through the same embedding algorithm (as shown in Figure.2)
• W atermark Extraction : The watermark extraction with reverse embedding process. In reverse embedding pro-
process is shown in Figure.2 and Figure.3 cess, the following changes are incorporated:
Step 1: Partitioning the watermarked image using same Change 1: In place of the (512 × 512 × 3) cover image,
procedure as done in pre-processing phase, we obtain (4 × 4) we use the (512 × 512 × 3) embedded image. Also in place of
sized Fi0 where i ∈ {1, 2, 3, 4} for a single sub-block. We the (64 × 64 × 3) watermark image, we use the (64 × 64 × 3)
then compute the average value of each Fi0 , by taking mean extracted watermark image.
all elements within each Fi0 . We denote the average value as Change 2: The Si conditions should be interchanged, i.e.
Ai where i ∈ {1, 2, 3, 4}. Si = 1 should result in subtraction process, while Si = 0
Step 2: We perform the following operation: should result in addition process.
Fi00 = (Fi0 − Ai ) × Li (2) III. E XPERIMENTS & R ESULTS
The implementation of this work is done in MATLAB.
It is important to note here that element-wise subtraction is
All the experimental results have been obtained and validated
being performed between Fi0 (matrix) and Ai (element). We
using MATLAB programming. Below are the simulation result
denote the result obtained as Fi00 where i ∈ {1, 2, 3, 4}.
of the algorithm:
Step 3: We then compute average value of each Fi00 , by
taking the mean of all elements within each Fi0 . We denote 1) elasped time for encryption - 0.991091 seconds
the average value as A0i where i ∈ {1, 2, 3, 4}. 2) elasped time for extraction - 0.411357 seconds
3) elasped time for combined operation - 1.409701 seconds
Step 4: The data extraction rule is expressed as follows:
The low elapsed time indicates the efficiency of this algo-
if A0i < 0 ⇒ d[i0 , j 0 ] = 1 rithm, by which it is discernible that the computational load
(3)
if A0i ≥ 0 ⇒ d[i0 , j 0 ] = 0 of the algorithm is minimal.

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 03,2021 at 16:51:05 UTC from IEEE Xplore. Restrictions apply.
A. Image Quality Metrics
Image quality can degrade due to distortions in the water-
marking embedding and extraction process. Objective mea-
sures of quality are necessary, but for blind watermark appli-
cations, it is vital to evaluate a quality metric which correlates
well with the subjective perception of quality by a human
observer. Considering the above points, we calculate four
quality metrics, namely:
1) Peak Signal-to-Noise Ratio (PSNR)
2) Normalized Correlation (NC)
3) Quality Factor (Q-factor)
4) Structural Similarity (SSIM)
We use the original image (cover and watermark) as the Figure 5. NC of embedded image & extracted watermark vs Modulation
reference to compute the quality of the extracted watermark Index
and extracted cover. For our application, three different inde-
pendent evaluation is required for each metrics:
1) Between the watermarked image and the original cover: 3) Quality Factor: The Q-factor is given as [17]:
human visual perceptibility should be very low
4σIK × µI µK
2) Between the extracted watermark and the original water- Q= (6)
mark: similarity index should be high (σI2 + σK
2 )(µ2 + µ2 )
I K
3) Between the extracted cover and the original cover:
where for images I & K, µI & µK denotes the mean,
similarity index should be high 2
σIK denotes the correlation coefficient and σI2 & σK is the
Discussion on the quality metrics is given below:
variance.
1) Peak Signal-to-Noise Ratio: The PSNR between the
reference image I(i, j, k) and the test image K(i, j, k) is given
by [12]:
Ψ2 × 3mn
P SN R = 10 log10 Pm Pn P3 2
i=1 j=1 k=1 [I(i, j, k) − K(i, j, k)]
(4)
Here, Ψ is the maximum possible pixel value of the image
(=255, in our case) and m, n respectively represents the height
& width of the image.
The PSNR result is shown in (Figure.4).

Figure 6. Q-factor of embedded image & extracted watermark vs Modulation

Index

The NC result is shown in (Figure.6).

4) Structural Similarity: The SSIM is given as [18]:

(2µI µK + C1 )(2σIK + C2 )
SSIM = (7)
(µ2I + µ2I + C1 )(σI2 + σK
2 +C )
2

[C1 C2 ] = (K[1] × Ψ)2 (K[1] × Ψ)2

 
(8)
Figure 4. PSNR of embedded image & extracted watermark vs Modulation
Index
where we choose K = [0.01 0.03]
2) Normalized Correlation: The normalized correllation is The SSIM result is shown in (Figure.7).
given as [6]: Conclusion for Image Quality Metrics: The proposed work
m n 3 has been implemented on eight different covers (of size (512×
1 XXX
NC = [I(i, j, k) × K(i, j, k)] (5) 512×3)) and eight different watermarks (of size (64×64×3)).
3mn i=1 j=1
k=1 The optimal value of K to obtain the most desirable result is
The NC result is shown in (Figure.5). for K = 10. Table.I details the data.

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 03,2021 at 16:51:05 UTC from IEEE Xplore. Restrictions apply.
 Table III
C OMPARISON BETWEEN NOISE INFLUENCED WATERMARK AND ORIGINAL
WATERMARK .

Extracted Watermark (w.r.t. Original watermark)

Noise
PSNR NC Q-factor SSI
Original 34.72319 0.99 1.00 0.93
Gaussian 30.59 0.97 0.99 0.86
Poissons 31.03 0.96 0.99 0.85
Speckle 32.09 0.99 1.00 0.92
Salt & Pepper 32.04 0.98 1.00 0.90

Table IV
C OMPARISON BETWEEN NOISE INFLUENCED WATERMARK AND ORIGINAL
WATERMARK .
Figure 7. SSIM of embedded image & extracted watermark vs Modulation
Extracted Cover (w.r.t Original Cover)
Index Noise
PSNR NC Q-factor SSI
Original 57.10 0.99 0.98 0.99
Table I Gaussian 49.13 0.88 0.97 0.17
I MAGE Q UALITY M ETRICS FOR K = 10. Poisson 51.93 0.98 0.97 0.64
Speckle 54.12 0.99 0.98 0.93
Image PSNR NC Q-factor SSI Salt & Pepper 54.23 0.99 0.98 0.96
Embedded image
34.21 0.98 0.96 0.72
(w.r.t. original Cover)
Extracted watermark
34.72 0.99 1.00 0.93
(w.r.t original Watermark) 2) Poisson noise: Also called ’shot noise’, this noise can be
Extracted cover
(w.r.t. original Cover)
57.10 0.99 0.98 0.99 modelled by a Poisson process. This noise mainly originates
from the discrete nature of the electric charge.
3) Speckle noise: Speckle is a granular interference that
inherently degrades the quality of medical ultrasound images
B. Noise Attack Analysis [7]. Though known as noise, it is not in itself a noise, but
As noise is one of the widely used techniques to distort blind rather an unwanted modification to the desired signal. Speckle
watermarks intentionally, we deliberately performed noise noise may severely diminish the performances of information
attack on the embedded image to determine the robustness extraction techniques [4].
of the watermarking algorithm. After injecting noise, we have 4) Salt & Pepper noise: Also known as ’impulse noise’, this
computed the image quality metrics between the noise injected is probably the most used form of noise. The presence of salt
embedded image and the cover image in Table.II. After & pepper noise can be identified by sparsely occurring black
extracting the watermark, we have compared the image quality & white pixels [9], thus resulting in the name. The median
metrics between the extracted (noise-influenced) watermark filter is an effective solution for this noise [10].
and the original watermark in Table.III. Likewise, we have
C. Cropping Attack Analysis
compared the image quality metrics between the extracted
(noise-influenced) cover and the original cover in Table.IV. Cropping attack is an effective way to eradicate blind
The modulation index chosen is as K = 10. watermarking. These algorithms are susceptible to cropping
attack of small ratio, resulting in the complete removal of
Table II the hidden information. Our proposed algorithm is partially
C OMPARISON BETWEEN NOISY EMBEDDED IMAGE AND COVER IMAGE . susceptible to cropping attack as only a portion of the embed-
ded image cropped will be affected and result in distorting
Embedded Image (w.r.t. Original Cover)
Noise the corresponding pixel values in the watermark. Because
PSNR NC Q-factor SSI
Original 34.21492 0.98 0.96 0.72 cropping attack is generally performed on small portions, and
Gaussian 31.33 0.97 0.93 0.51 that the watermark pixel information is spread out throughout
Poissons 32.99 0.97 0.94 0.51
Speckle 33.90 0.98 0.96 0.69
the cover image, only negligible portion of the watermark will
Salt & Pepper 33.49 0.98 0.96 0.70 be affected.
In our analysis, we computed the quality metrics between the
unaffected area of extracted watermark and the corresponding
The following are the different kinds of noise used: uncropped area in the original watermark. The modulation
1) Gaussian noise: The probability of this noise is a index chosen is as K = 10. The values are depicted in Table.V.
Gaussian-distributed density function. The relevance of this The result proves that the amount of cropping doesn’t
noise is in telecommunication & networks, where channels are affect the pixel values of the watermark corresponding to
affected by wide-band Gaussian noises coming from natural the uncropped area. This makes our algorithm very ideal for
sources. applications where robustness is a crucial factor expected.

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 03,2021 at 16:51:05 UTC from IEEE Xplore. Restrictions apply.
 Table V [8] Y. Himeur and B. Boudraa. “Secure and robust audio wa-
C ROPPING ATTACK R ESULTS . termarking system for copyright protection”. In: 2012 24th
International Conference on Microelectronics (ICM). Dec.
Cropping Ratio PSNR NC Q-factor SSI 2012, pp. 1–4. DOI: 10.1109/ICM.2012.6471449.
0% 34.723 0.996756 1 0.9315
[9] “Image Restoration”. In: Practical Image and Video Pro-
5% 34.431 0.940299 1 0.9964
10% 34.399 0.952176 1 0.9964
cessing Using MATLAB®. John Wiley Sons, Ltd, 2011.
25% 34.239 0.956386 1 0.9909 Chap. 12, pp. 265–298. ISBN: 9781118093467. DOI: 10.1002/
50% 34.464 0.972482 1 0.9906 9781118093467 . ch12. eprint: https : / / onlinelibrary . wiley .
com / doi / pdf / 10 . 1002 / 9781118093467 . ch12. URL: https :
//onlinelibrary.wiley.com/doi/abs/10.1002/9781118093467.
ch12.
IV. C ONCLUSIONS & F UTURE S COPE [10] Jan Eric Kyprianidis. “Artistic Stylization by Nonlinear Filter-
ing”. In: Image and Video-Based Artistic Stylisation. Ed. by
We have proposed a novel blind and invisible, semi-fragile Paul Rosin and John Collomosse. London: Springer London,
colour image watermarking scheme, where the possession of 2013, pp. 77–101. ISBN: 978-1-4471-4519-6. DOI: 10.1007/
a secret key (sequence) is required for extracting the blind 978-1-4471-4519-6 5. URL: https://doi.org/10.1007/978-1-
watermark from the embedded image. We have implemented 4471-4519-6 5.
[11] Phen Lan Lin, Chung-Kai Hsieh, and Po-Whei Huang. “A
the scheme using MATLAB programming, and have presented hierarchical digital watermarking method for image tamper
the results corresponding to the quality of the watermarked detection and recovery”. In: Pattern Recognition 38.12 (2005),
image, by computing various image quality metrics. After pp. 2519–2529. ISSN: 0031-3203. DOI: https://doi.org/10.1016/
performing different relevant analysis on the embedded image, j.patcog.2005.02.007. URL: http://www.sciencedirect.com/
such as noise attacks, cropping attacks, we are concluding that science/article/pii/S0031320305000890.
[12] Chun-Shien Lu. Multimedia Security: Steganography and Dig-
the proposed algorithm is ideal to be deployed in applications ital Watermarking Techniques for Protection of Intellectual
where computational resource and time is sparse, yet a highly Property. USA: IGI Global, 2004. ISBN: 1591401925.
robust yield is expected. [13] Bernard Marr. How Much Data Do We Create Every Day?
When used for practical scenarios, it is necessary to design a URL: https : / / www . forbes . com / sites / bernardmarr / 2018 /
dedicated VLSI hardware system which is capable of perform- 05 / 21 / how - much - data - do - we - create - every - day - the -
mind - blowing - stats - everyone - should - read / #17cd71af60ba.
ing this algorithm. Due to the low computational requirements (accessed: 09.01.2020).
of this scheme, the ASIC development is preferred as com- [14] V. M. Potdar, S. Han, and E. Chang. “A survey of digital image
pared to other feasible blind image watermarking algorithms. watermarking techniques”. In: INDIN ’05. 2005 3rd IEEE In-
ternational Conference on Industrial Informatics, 2005. Aug.
R EFERENCES 2005, pp. 709–716. DOI: 10.1109/INDIN.2005.1560462.
[1] Jobin Abraham and Varghese Paul. “An Imperceptible Spatial [15] Ayesha Shaik and V. Masilamani. “Secure Video Watermark-
Domain Color Image Watermarking Scheme”. In: Journal of ing Technique Using Cohen-Daubechies-Feauveau Wavelet
King Saud University - Computer and Information Sciences and Gollman Cascade Filtered Feedback Carry Shift Register
31 (Dec. 2016). DOI: 10.1016/j.jksuci.2016.12.004. (F-FCSR)”. In: Proceedings of the 3rd International Sympo-
[2] Namita Agarwal, Amit Singh, and Pradeep Singh. “Survey of sium on Big Data and Cloud Computing Challenges (ISBCC
robust and imperceptible watermarking”. In: Multimedia Tools – 16’). Ed. by V. Vijayakumar and V. Neelanarayanan. Cham:
and Applications 78 (Jan. 2019). DOI: 10.1007/s11042-018- Springer International Publishing, 2016, pp. 111–118. ISBN:
7128-5. 978-3-319-30348-2.
[3] Umar Albalawi, Saraju Mohanty, and Elias Kougianos. “A new [16] Anastasios Tefas, Nikos Nikolaidis, and Ioannis Pitas. “Chap-
region aware invisible robust blind watermarking approach”. ter 22 - Image Watermarking: Techniques and Applications”.
In: Multimedia Tools and Applications (Sept. 2017). DOI: 10. In: The Essential Guide to Image Processing. Ed. by Al Bovik.
1007/s11042-016-4063-1. Boston: Academic Press, 2009, pp. 597–648. DOI: https://doi.
[4] F. Argenti et al. “A Tutorial on Speckle Reduction in Synthetic org/10.1016/B978-0-12-374457-9.00022-6.
Aperture Radar Images”. In: IEEE Geoscience and Remote [17] Zhou Wang and A. C. Bovik. “A universal image quality
Sensing Magazine 1.3 (Sept. 2013), pp. 6–35. ISSN: 2373- index”. In: IEEE Signal Processing Letters 9.3 (Mar. 2002),
7468. DOI: 10.1109/MGRS.2013.2277512. pp. 81–84. ISSN: 1558-2361. DOI: 10.1109/97.995823.
[5] D. E. Bakken et al. “Data obfuscation: anonymity and desen- [18] Zhou Wang et al. “Image quality assessment: from error
sitization of usable data sets”. In: IEEE Security Privacy 2.6 visibility to structural similarity”. In: IEEE Transactions on
(Nov. 2004), pp. 34–41. ISSN: 1558-4046. DOI: 10.1109/MSP. Image Processing 13.4 (Apr. 2004), pp. 600–612. ISSN: 1941-
2004.97. 0042. DOI: 10.1109/TIP.2003.819861.
[6] Enping Li, Huaqing Liang, and Xinxin Niu. “Blind Image
Watermarking Scheme Based on Wavelet Tree Quantization
Robust to Geometric Attacks”. In: 2006 6th World Congress
on Intelligent Control and Automation. Vol. 2. June 2006,
pp. 10256–10260. DOI: 10.1109/WCICA.2006.1714009.
[7] Mohamad Forouzanfar and Hamid Moghaddam. “Ultrasound
Speckle Reduction in the Complex Wavelet Domain”. In:
Jan. 2010, pp. 558–577. ISBN: 1891121952. DOI: 10 . 1049 /
SBRA023E ch33.

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 03,2021 at 16:51:05 UTC from IEEE Xplore. Restrictions apply.