2015 12th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE), Mexico, City.

Mexico

Invisible Digital Color Watermarking Technique In

Anaglyph 3D Images
David Octavio Munoz-Ramirez, Rogelio Reyes-Reyes, Volodymyr Ponomaryov, Clara Cruz-Ramos
Seccion de Estudios de Posgrado e Investigacion, Instituto Politecnico Nacional, ESIME Culhuacan
Mexico City, Mexico
[email protected], [email protected], [email protected]

Abstract- In recent years, 3D technology has grown image is revealed. The visual cortex of the brain fuses the images
significantly and over the time has become more accessible to a received from each eye and forms a 3D image.
wider group of people, however, there is so few research that seeks
to protect the intellectual property of 3D content, in specific There are schemes for 3D models [4,5],meshes [6,7],DIBR
anaglyph images, which are easier and cheaper to generate. This [S] and pair stereo images [9,10],but,there are only few articles
paper presents a method to embed a color image as an invisible about 3D anaglyph image watermarking. Bhatnagar, Wu and
watermark into a 3D anaglyph image, for the protection of Ranman, proposed a robust security framework [11]. In this
intellectual property of this type of digital material; the insertion paper, the 3D anaglyph images are transformed into RGB space
algorithm is performed in the DCT domain using the QIM-DM color and each channel is processed separately. The watermark
method in order to achieve the maximum imperceptibility of the embedding performs the Fractional Fourier Transform [12] and
watermark. Results demonstrate highest robustness of the Reversible Integer Transform with Singular Value
algorithm to JPEG compression, impulsive noise, Gaussian noise, Decomposition [13].
as well as high imperceptibility of the watermark obtaining
average values of about 40dB in the PSNR between the original Prathap and Anitha proposed in [14] a simple scheme for 3D
and marked anaglyph images. anaglyph image protection, based on 3D DWT and Jacket
Matrix. This scheme decomposes the host 3D anaglyph image
Keywords-3D image, Anaglyph, Invisible, Digital Color by performing 3D DWT. Next, the middle level sub bands are
Watermarking, copyright protection. divided into blocks and each block is masked using Jacket
matrix [15,16]. Jacket matrix is a special class with their inverse
I. INTRODUCTION being determined element-wise. The watermark is embedded by
Nowadays, 3D technology has become more accessible to a changing the diagonal elements of the masked blocks.
wider group of people, this technology can not only be
This article proposes an algorithm for the protection of
accessible from a room specially designed for this; now with the
copyright in anaglyph 3D images,by inserting an invisible color
necessary accessories it is possible to enjoy 3D content
watermark in the frequency domain, employing the Discrete
anywhere, even at home. However, this ease of access and
Cosine Transform (DCT) and Quantification Index Modulation
visualization of 3D content also provides the facility to illegal
(QIM) [17] with its variant Dither Modulation (DM) [IS].
copying and distribution, so a technique for the protection of
copyright in the digital content, which must preserve the quality First, the 3D anaglyph image is converted from RGB to
of the original 3D content is needed. YCbCr color space, then only the luminance channel Y is
selected; the watermark is converted to YCbCr and the Chroma
One of the techniques most widely used to solve the problem
subsampling (4:2:0) process is performed, with the objective to
of copyright of digital contents in transmitting, processing and
reduce the information of watermark to minimum, whereby only
storing, is digital watermarking; which consists in embedding
12 bpp (bits per pixel) are needed instead of 24 bpp. The next
information into the digital content to be protected, it is
step is to calculate the DCT to each SxS block, and the
important that the embedding process does not cause serious
watermark is embedded using the QIM-DM algorithm. Finally
degradation on the original digital content; additionally, the
the mCT is applied on each SxS watermarked block, and the
protected material must be robust to common degradations, i.e.,
originals Cb and Cr channels are jointed with the protected
digital compression, noise, copying. Digital watermarking
luminance channel Y'. Experimental results show that this
techniques are classified as visible or invisible, and may be used
scheme is quite robust against JPEG compression,impulsive and
to authenticate the author, who is the original owner of the
Gaussian noise; the main contribution of this work is the
digital content [1]. When inserting a digital watermark one
insertion and extraction of a color watermark in 3D images,
should care about certain characteristics as capacity, robustness
which remains a challenge to date.
and imperceptibility [2].
The rest of this paper is organized as follows: Section two
Anaglyph images are based on the binocular vision
provides information about DCT and Section three shows the
phenomenon which was patented by Louis Ducos in IS91 [3].
algorithm of Quantization Index Modulation. In section four we
Anaglyph contains two images filtered by color, one for each
explain the watermark embedding and extraction procedures.
eye. When viewed through special glasses, a three-dimensional
Section five presents the experimental results which illustrate

978-1-4673-7839-0/151$31.00 ©20151EEE
2015 12th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE), Mexico, City.
Mexico
the imperceptibility and robustness of the scheme. Finally, ftN-1�
Section six concludes the paper and proposes some future work. 2 (2i + 1)krr
xCi) IV C(k)x(k)cos (2)
II. DISCRETE COSINE TRANSFORM (DCT)
=

2N [ ]
The Discrete Cosine Transform (DCT) is an operation based
where:
on the Discrete Fourier Transform (DFT), but it only acts on
k = 0,1,2, ...,N-1
periodic functions with symmetrical pair and the result is a
i = 0,1,2, ...,N-1
sequence of real numbers, in terms of the sum of the cosine
function; it is the most used one in image compression, because
it has a good energy compaction property. Another important The number of operations performed by the one-dimensional
aspect of the DCT is the property to quantifY the coefficients DCT is lower and the processing time and the computational
using quantization values. The coefficients of the resulting cost goes down considerably [19].
correlation are very important, since further processing may be III. QUANTIZATION INDEX MODULATION (QIM)
performed independently.
In 2001 Brian Chen and Gregory W. Wornel proposed a new
Because the functions of the two-dimensional DCT are method of watermarking, which they called "Quantization
products of two one-dimensional functions, the only constant Index Modulation (QIM)". From the test they prove the
feature (DC coefficient) is in the upper left comer of the matrix effectiveness of this algorithm, and demonstrate that this kind
(Fig. 1). of watermarking is quite robust against distortions, in addition
to have a very good performance [17], [20], [21]. The main
AC (28)
property of this method is to embed information by modulating
r--r---r----r---,---.--...,...., / a sequence of indexes with the watermark and then quantifY the
host signal with the quantifier or quantifier sequence index
associated with the watermark [IS].
A. Dither Modulation
It is important to consider that Dither quantifiers have the
property that their quantization values and points of
reconstruction of any existing quantifier in the set, are different
versions of the quantization values and points of reconstruction
of the other existing quantifiers, each quantization index has its
own point of reconstruction [IS].
Fig. I. DCT coefficients.
Dither quantifiers refer to a single quantifier base and
gradually replace the other quantifiers by pseudo-random noise.
An M x N image is usually divided into non-overlapping
For this you can use dither vectors and gradually replace these
blocks, each one of size L x L. For simplicity the rows and quantifiers, these vectors are generated pseudo-randomly. To
columns of an image, and also the block, are assumed to be the make the insertion into an image, it is necessary to modulate
same size. In general, they have been used in blocks of size SxS
each of the vectors with the watermark to be inserted. This
image codification. These blocks allow you to enter small
generates a dither vector d(m) for each of the values contained
adaptive features based on existing activity or detail in the block.
in the watermark. Finally, the host image is quantized by the
A. 2-DCTusing two i-DCT initial quantizer base q(.) and the marking function is given by
(3).
One of the properties of the two dimensional DCT is that it
is separable and it is possible to obtain the same result using two (3)
sex; m) q ( x + d(m)) - d(m).
=
one-dimensional DCTs. To get the same result, first the one­
dimensional DCT is applied to the rows and of the resulting To obtain the marked signal s the first thing to do is to add
matrix is applied again but now a one-dimensional DCT on the the dither vector d(m) to the sample x of the host image, then
columns getting the same result as if we use two-dimensional quantifY with the closest reconstruction point, finally the same
DCT. This same process applies to the mCT. dither vector d(m) is subtracted.

The two-dimensional discrete cosine transform is obtained IV. PROPOSED ALGORITHM

easily from one pair of one-dimensional DCT. For calculating The block diagram of the proposed data embedding method
the one dimensional DCT, equation (1) is used and for the is shown in Fig. 2.

fl
inverse of DCT (2).
A. insertion Method.

2
N-l
(2i + l)krr The host anaglyph image Ah is converted from RGB (A�GB)
X(k) e(k) b x(i)cos (1) to YCbCr (A�CbCT). From YCbCr color space (A�CbCT) only the
=

N [ 2N ] luminance channel Y (A�) is used; the embedding process of

watermark W is performed in this channel. A� is divided into
non-overlapping SxS pixel blocks to apply DCT, with the i-th
2015 12th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE), Mexico, City.
Mexico
block denoted as A�i' From A�i the first twelve AC coefficients
are modified according to the QIM-DM algorithm and the W d[k. 1] =
{
d[k.O] +!J./z. d[k. 0] < 0
k 1 L (6) = • ... •

values, the AC coefficient is denoted as A�i' d[k.O] _ !J./z, d[k. 0] 2: 1

where R is a random number generator, this factor gives

more strength to the embedding process. The distance between
d[k,O] and d[k, 1] is tJ.j2, these dither values are used for
embedding the bit 0 and bit 1 from the watermark Wsj' The
length L depends on the number of coefficients that are selected
from A�i'
After the color watermark is embedded, the mCT is applied
to obtain a new block A�i in the spatial domain. This whole
process is repeated until the watermark Ws has been inserted
completely. Then it is necessary to join all the channels �A�b ,
A �r and A� ) to obtain a watermarked anaglyph image A� CbCr.
Finally, if necessary, A� cbcr is returned to RGB color space
Non overlapping
8x8 block A�i
Non overlapping
block of 12 bits of
(A�GB).
i = {1.2,3 . .. n}

B. Extraction Method.
The proposed color watermark extraction algorithm is
totally blind, this means that the extraction process does not
require any additional information as: original anaglyph image,
original color watermark, nor additional data. The block
Chane Is CbCr diagram for watermark extraction is show in Fig. 4.
Anaglyph A�b A�T

Fig. 2. Color watermark insertion block diagram.

Color image is chosen as watermark W, and it is converted

from RGB ( W R GB ) to YCbCr (wYCbCr), then a subsampling
(4:2:0) over the WYCbCr is performed to obtain a new
watermark (Ws), in which each pixel is represented with twelve Luminance
bits per pixel (bpp) instead of twenty four bpp, allowing to ChannelY
reduce the watermark in the chroma channels a quarter of the A�
original size (Fig. 3).

Non overlapping QIM-DM

C!:EJ 8x8 block A�i Extracction
C!:EJ i = {1.2.3 ... n} per block

Fig. 3. 4:2:0 Chroma Subsampling.

Fig. 4. Color watermark extraction block diagram..
For each Wsj it is necessary one A�i to embed the watermark
inside the first twelve AC coefficients. First the watermarked anaglyph (Aw) is converted from
Both A�i and Wsj are used by QIM algorithm (4). color space RGB (A�GB) to YCbCr (A� cbcr). The watermarked
(4) luminance channel Y (A�) is separated to extract the color
watermark.

Using two different dither vectors d[k. 0] and d[k. 1] of length

The luminance channel Y (A�) is divided into non­
L, pseudo randomly generated using a key with the step sizes !:J.
overlapping 8x8 blocks and DCT is applied,with the i-th block
and is generated as follow (5, 6).
denoted as A�i' only the first twelve AC coefficients are
necessary to extract the color watermark. A watermark bit Wsj
d[k.O] = round(!J.j R) * !J. (5)
2015 12th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE), Mexico, City.
Mexico
is recovered by evaluating the AC coefficients A�i using the Gaussian Noise
following rule (7).

w5J. = {O1 if dmin1 < dmin2

Otherwised
(7) 40

35
Where the parameters dmin1 and dmin2 are calculated as
iD
follow:
�
z
30
(f)
0-

dmin1 =
(( d[k, 0]
-
AWC i + tJ, -
) ) *
tJ,
- d[k,O]
(8)
25

_
dmin2 -
(( AWC i +
d[k, 1]
-
tJ,
-
) ) * tJ, - d[k,l] (9) 20

k = 1,2,3 ... . L 1 5 ����--��������--����

o 0.1 0.2 03 0.4 0.5 0.6 01 0.8 0.9
i
where A�i is i th coefficient of DCT block, d[k, 0] and d[k, 1]
Variance

are two dither values obtained by (6) and 11 is the distance Fig. 5. Robustness against Gaussian noise.
between those dither values. The length L depends on the
number of coefficients that are selected from A�i' Impulsive Noise

These steps should be repeated until all the input values of

the watermark W5j are obtained. Then the watermark Ws is
resampled to get WYCbCT which is converted from YCbCr color
space to RGB, so finally the recovered color watermark is
obtained WRGB• 34
iD
os.
V. EXPERIMENTAL RESULTS '" 32
z
C/J
0-
The proposed invisible digital color watermarking scheme 30
for anaglyph 3D images, was developed using .Net Framework
2.0 platform and C # language; and tested on a PC with the 28

following technical features:

26

• CPU Intel (R) Pentium (R) D 2.79GHz 2.80 GHz -�--����-�

24 L-�-
o 0.1 0.2 0.3 0.4 0.5 0.6 07 0.8 0.9
• RAM memory 2.00GB Noise Density
)( 1030
• Windows 7 64-bit OS
Fig. 6. Robustness againts Impulsive noise.
Dataset of 3D anaglyph images used for test the proposed
algorithm were taken from Middlebury [22]; the results shown

il \ �, I'
in the next section were obtained from the following images:
Backpack, Classroom1, Flowers, Motorcycle, Pipes and

. . ( ,4f-,� j. \ ( '�
Playroom. All images were resized to a standard size of

�)
3648x2736 pixels, and the watermark (Lena) has a resolution
.. I
of 456x342. The average time of watermark insertion was 12.31
seconds and the average watermark extraction time was 6.23
seconds. a)
�' .

b)

To evaluate the effectiveness against JPEG compression

attack, the following evaluation criteria were used: Peak signal­
to-noise ratio (PSNR), Structural Similarity Index Measure
(SSIM) and Normalized Color Difference (NCD), results are
shown in Tables I and II. Fig. 5 and 6 show the robustness
against Gaussian and Impulsive noise attacks respectively; Fig.
7 and 8 show recovered color watermark after noise attacks with c) d)
different densities and variances. Finally, Fig. 9 shows the
recovered color watermark with different JPEG compression Fig. 7. Watermark recovered after applying gaussian noise attack
ratios. a)O' = 0.0001, b) 0' = 0.0003, c) 0' = 0.0005, d) 0' = 0.0007
2015 12th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE), Mexico, City.
Mexico

TABLEL PSNR AND SSIM VALUES OF WATERMARKED 3D ANAGLYPH IMAGES AT DIFFERENT JPEG COMPRESSION RATIOS

Quality Backpack Classroml Flowers Motorcycle Pipes Playroom

100 464/0.9996 43.57/0.9989 45.94/0.9993 43.6/0.9991 45.51/0.9971 41.5/0.9945

90 43.76/0.993 42.12/0.9983 44.08/0.9989 41.57/0.9987 41.78/0.9989 41.15/0.9967

85 42.97/0.9991 41.57/0.9979 43.5/0.9986 40.88/0.9984 41.04/0.9986 40.58/0.9963

80 42.28/0.9985 4107/0.9972 42.93/0.998 40.28/0.9977 4043/0.9976 40.06/0.9962

70 41.52/0.9978 4048/0.9964 42.16/0.9963 39.56/0.9968 39.69/0.9965 394/0.9938

60 40.81/0.9966 39.93/0.9952 41.47/0.9937 38.91/0.995 39.01/0.9938 38.82/0.9919

50 40.36/0.9956 39.61/0.9944 41.14/0.9951 38.54/0.9942 38.64/0.9931 38.5/0.9919

40 39.89/0.9938 39.27/0.9925 40.58/0.9907 38.11/0.9915 38.21/0.9894 381/0.989

30 39.37 /0.991 38.76/0.9884 39.87/0.981 37.61/0.9872 37.68/0.9818 37.61/0.9819

TABLE n. PSNR, SSIM AND NCD VALUES OF RECOVERED WATERMARK LENA FROM DIFFERENT JPEG COMPRESSION RATIOS

Quality Backpack Classroml Flowers Motorcycle Pipes Playroom

100 48.09/0.9730/0.006 45.01/0.9717/0.006 46.03/0.9725/0.006 45.01/0.9713/0.006 47.26/0.9728/0.006 40.24/0.9073/0.006

90 47.98/0.9729/0.006 43.35/0.9695/0.006 44.59/0.9718/0.006 43.42/0.9686/0.006 46.51/0.9727/0.006 38.78/0.9033/0.006

85 47.80/0.9728/0.006 43.20/0.9691/0.006 44.22/0.9714/0.006 42.79/0.9662/0.006 4648/0.9725/0.006 3843/0.9016/0.006

80 47.29/0.9723/0.006 42.58/0.9674/0.006 44.02/0.9711/0.006 41.69/0.9623/0.006 4642/0.9724/0.006 37.66/0.897/0.006

70 45.38/0.9461/0.103 40.29/0.9401/0.103 41.64/0.9443/0.103 3945/0.9316/0.103 43.74/0.9468/0.103 364/0.8676/0.103

60 40.84/0.9088/0.155 38.23/0.9054/0.164 36.54/0.9123/0.164 36.52/0.8789/0.155 40.81/0.9104/0.164 34.03/0.8049/0163

50

� :-"' " c ,
31.99/0.7217/0.146 31.72/0.7257/0146 32.06/0.7294/0136 31.00/0.6965/0.146 32.04/0.7205/0.146 30.37/0.6648/0136

40 23.88/0.3862/0.133 23.79/0.3822/0.133 23.86/0.3849/0.132 23.70/0.3747/0.152 23.98/0.3898/0.133 24.93/04404/0.119

�
30 20.28/0.2525/0.153 20.20/0.2461/0.210 20.23/0.2492/0.153 20.23/0.2408/0.152 20.33/0.2511/0.210 21.06/0.2863/0153

--,

",'" . �.J \.' ",' I" J .

. ..�
.
�' ! I :-"" '

,IJ
bo' )
, ., ..

\.' . .� .

� ..·�
\ .
\ " ", /3
.�. .

'
. .'
. ' .
. :�.
" �I '

� -
�
i
'
",

(
/3; �c - f
' '

' . '
.
,
. (' �" ( ,

' ;.�.
, - a) b)
a) bl

I ��. , ! . �,

� \ iJ.'·"
I 4J )�
\ .�
�'(' . �. _ , �" '" "
. ; :'; '::'

•
• ,. r·
r • ' , .. ,
. . . -
c) d)
c) d)

Fig. 8. Watermark recovered after applying impulsive noise attack Fig. 9. Watermark recovered after applying different JPEG compression
a) density = 0.0001, b) density = 0.0003, c) density = 0.0005, d) density = 0.001 qualities a)QF = 85, b) QF = 70, c)QF = 50, d)QF = 30
2015 12th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE), Mexico, City.
Mexico
VI. CONCLUSIONS [8] Yu-Hsun Lin; Ja-Ling Wu, "A Digital Blind Watermarking for Depth­
Image-Based Rendering 3D Images", Broadcasting, IEEE Transactions,
In this paper an algorithm for copyright protection in 2011, Volume: 57, Issue: 2, Pages: 602 611, DOl:
anaglyph 3D images based on the invisible insertion of a color 10.1109ffBC.2011.2131470
image as a watermark is presented; the proposed method is based [9] Zadokar, S.R.; Raskar, V.B.; Shinde, S.V., "A digital watermarking for
mainly on the QIM-DM technique that has proven to be efficient anaglyph 3D images", Advances in Computing, Communications and
Informatics (ICACCI), 2013 International Conference, Pages: 483 - 488,
when applied to several algorithms of watermarking; the use of DOl: 10.1I09/1CACCI.20I3.6637219
a color image as a watermark, is a significant challenge to the
[10] Mei Yu, Aihong Wu, Ting Luo, Gangyi Jiang, Wujie Zhou, Songyin Fu,
methods of invisible watermarking, due to the large amount of "A New Stereo Image Watermarking Method for 3D Media", Original
information to be embedded into the host image, this challenge Research Article Procedia Engineering, Volume 29, 2012, Pages 2399-
is mitigated by chroma subsampling of the color image, which 2404, doi:I 0.1 016/j.proeng.2012.01.322
reduces by half the number of bits required to represent each [II] Bhatnagar G, Wu J, Raman B. "A robust security framework for 3D
color pixel. The proposed system achieves high imperceptibility images", J Visual 2011;14(1):85-93.
with average values of PSNR and SSIM of 41.36 dB and 0.9966 [12] Yury F. Luchko, Hector Martinez, Juan J. Trujillo, "Fractional Fourier
respectively; as well as high robustness against ] PEG Transform and some of its applications", Fractional Calculus & applied
Analysis, vol. II, number 4, pages 457-470, 2008, ISSN 1311-0454.
compression attacks, impulsive noise and Gaussian noise; even
[13] B.Chandra Mohan, S. Srinivas Kumar, "A Robust Image Watermarking
the color watermark could be recovered up to a JPEG Scheme using Singular Value Decomposition", Journal of Multimedia,
compression of 70%, obtaining average values of PSNR, SSIM vol. 3, no. I, pages 7-15, May. 2008.
and NCD 41 dB,0.9090 and 0.061 respectively for the recovered [14] Ivy Prathap, R. Anitha, "Robust and blind watermarking scheme for three
color watermark. Implementing two one-dimensional DCTs dimensional anaglyph images", Computers & Electrical Engineering,
instead of a bi-dimensional DCT significantly reduces the Volume 40, Issue I, January 2014, Pages 51-58,
insertion and extraction time,up to a factor of 10. Moreover,this doi:10.1016/j.compeleceng.2013.11.005
proposal uses a totally blind extraction scheme, where original [15] Lee MH, Hou J. Fast block inverse jacket transform. IEEE Signal Process
Lett 2006; 13(8):461-4.
anaglyph image, original color watermark, or additional data,
are not required for the correct extraction of the color watermark [16] Lee MH, Finlayson K. A simple element inverse Jacket transform coding.
IEEE Signal Process Lett 2007;14(5):325-8.
image. Also it is important to mention,that there is not a similar
[17] Brian Chen, Gregory W. Wornell, "Quantization Index Modulation: A
proposal in the literature, which embeds a color image as a Class of Provably Good Methods for Digital Watermarking and
watermark into a 3D anaglyph image, and the experimental Information Embedding", IEEE Transaction on Information Theory,
results clearly demonstrate that the proposed scheme achieves VOL. 47, NO. 4, MAY 2001.
excellent image quality and robustness. [18] Brian Chen and Gregory W. Wornell, "Dither modulation: a new
approach to digital watermarking and information embedding", Research
Laboratory of Electronics, arid Department of Electrical Engineering and
Computer Science Massachusetts Institute of Technology Cambridge,
ACKNOWLEDGMENT MA 02139, January 1999.
The authors are thankful to the Instituto Politecnico [19] Dr. Naim Dahnoun, "Chapter 18 Discrete Cosine Transform", Bristol
University, Texas Instruments, 2004
Nacional,BEIFI Projects 20151241 and 20151522,CONACYT
[20] Mendoza-Noriega, J.A., Kurkoski, B.M., Nakano-Miyatake, M., Perez­
560086 and COFAA, for providing financial support for this
Meana, H., "Halftoning-based self-embedding watermarking for image
work. authentication and recovery", 53rd IEEE International Midwest
Symposium on Circuits and Systems (MWSCAS), pp. 612-615, Aug.
2010
[21] Avila, C.S. and Miyatake, M.N., "Multipurpose Image Watermarking
REFERENCES
Scheme Based on Self-Embedding and Data Hiding into Halftone Image",
[I] Podi1chuk, c.i.; Delp, E.J., "Digital watermarking: algorithms and Electronics, Robotics and Automotive Mechanics Conference (CERMA),
applications", Signal Processing Magazine, IEEE 2001, Volume: 18, 2010, pp. 394-398, Oct. 2010.
Issue: 4 Pages: 33 - 46, DOl: 10.1109179.939835 [22] D. Scharstein, H. HirschmOller, Y. Kitajima, G. Krathwohl, N. Nesic, X.
[2] M.L. Miller, I.J. Cox, M. G. Linnartz, T. Kalker, "A review of Wang, and P. Westling. High-resolution stereo datasets with subpixel­
watermarking principles and practices". 1999. accurate ground truth. In German Conference on Pattern Recognition
[3] Ian P. Howard,Brian J. Rogers, "Binocular Visiond and Stereopsis", (GCPR 2014), MOnster, Germany, September 2014.
Oxford University Press, 1996, Print ISBN-13: 9780195084764,
DOl: 10.1 093/acprof:oso/9780195084764.001.0001
[4] Ming Luo; Bors, A.G., "Surface-Preserving Robust Watennarking of 3-D
Shapes", Image Processing, 2011, Volume: 20, Issue: 10 Pages: 2813 -
2826, DOl: 1O.1I09ffIP.2011.2142004
[5] Hua Deng; Lei Chen; Jifu Zhang; Rui Wang, "A 3D Model Watermarking
Algorithm Resistant to Affine Transformation", Multimedia Information
Networking and Security (MINES), 2012 Fourth International
Conference, Pages: 549 - 551, DOl: 10.1109/MINES.2012.11
[6] Ben Jabra, S.; Zagrouba, E., "A Robust and Blind 3D Mesh Watermarking
Driven by Stable Regions", Computer Technology and Development,
2009. ICCTD '09. International Conference, Pages: 381 - 385, DOl:
10.1 I09/ICCTD.2009.31
[7] Vasic, B., "Blind QIM-LDPC watermarking of 3D-meshes",
Communications Workshops (ICC), 2013 IEEE International Conference,
Pages: 702 - 706, DOl: 10.1 I 09/1CCW.201 3.6649324