G Model

ARTICLE IN PRESS

AEUE 51503 18

Int. J. Electron. Commun. (AE) xxx (2015) xxxxxx

Contents lists available at ScienceDirect

International Journal of Electronics and

Communications (AE)
journal homepage: www.elsevier.com/locate/aeue

Binocular visual characteristics based fragile watermarking scheme

for tamper detection in stereoscopic images

Q1

Wujie Zhou a,b, , Lu Yu a , Zhongpeng Wang b , Mingwei Wu b , Ting Luo c , Lihui Sun b
a

Institute of Information and Communication Engineering, Zhejiang University, Hangzhou 310027, China
School of Information and Electronic Engineering, Zhejiang University of Science & Technology, Hangzhou 310023, China
College of Science & Technology, Ningbo University, Ningbo 315211, China

a r t i c l e

8
20

i n f o

a b s t r a c t

Article history:
Received 12 May 2015
Accepted 12 October 2015

10
11
12
13

19

Keywords:
Stereoscopic image
Watermarking
Binocular visual characteristics
Authentication
Tamper detection

21

1. Introduction

14
15
16
17
18

In the past decade, various fragile digital watermarking techniques have been proposed for monoscopic
image authentication and temper detection. In this paper, a novel binocular visual characteristics based
pixel-wise fragile watermarking scheme for stereoscopic image authentication and tamper detection is
proposed. The scheme consists of two processes: embedding of the stereoscopic image authentication
message and tamper detection. In the watermark embedding process, the binocular just noticeable difference (BJND) model is used for guiding watermark embedding, which is convenient for achieving a
balanced relationship between watermarking capacity and visual transparency. In the tamper detection
process, a probability theory and an algorithm of binocular disparity are employed to improve previously
obtained detection results and to enhance authentication accuracy. Moreover, to improve the security of
the proposed scheme, MD5 hash function and chaotic map are used. Experimental results reveal that the
proposed scheme is not only secure, but also achieves superior tamper detection for different attacks.
2015 Published by Elsevier GmbH.

Q3

Stereoscopic images are acquired from two slightly different

perspectives. One can perceive a stereoscopic image in three
dimensions (3D) when the left and right views of the stereoscopic image are viewed by the respective eye of the end-user
[1]. In recent times, stereoscopic images are being used in many
applications such as ight simulators for pilot training, robot
vision, medical surgery, military surveillance, virtual reality games,
and autonomous navigation [2]. With the rapid developments in
Internet and computer technology in recent years, it has become
possible to easily access, manipulate, and tamper with stereoscopic
images in open networks without the permission of the original
authors [3]. Precision and integrity of stereoscopic images is important, and in this regard, digital watermark technology is considered
a reliable and effective solution for authentication [4].
Several previous studies have made some progress in watermarking techniques for stereoscopic images especially in the
eld of copyright protection [57]. In [5], Lee et al. proposed a

22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

Q2

Corresponding author at: School of Information and Electronic Engineering, Zhejiang University of Science & Technology, Hangzhou 310023, China.
Tel.: +86 571 85070303.
E-mail address: wujiezhou@163.com (W. Zhou).

watermarking scheme based on discrete cosine transform (DCT) for

copyright protection of stereoscopic images. Hwang et al. used an
MPEG video watermarking scheme to embed watermarks into the
DCT [6] and discrete wavelet transform (DWT) [7] coefcients of the
decomposed right view. In [8], Coltuc et al. proposed a reversible
watermarking scheme for stereoscopic images for storage and
bandwidth reduction; however, inter-correlations between the
two views were missing. To improve the performance of a stereoscopic image watermarking scheme, the features of stereoscopic
images need to be excavated and utilized [912]. Campisi proposed
a proof of concept for a watermarking scheme for stereoscopic
images, wherein he makes use of the extracted depth information
[9]. Chammenm et al. embedded watermarks into the disparity computed from the left and right views while maintaining
an adequate trade-off between robustness and transparency [10].
Niu et al. proposed a visual sensitivity model guided stereoscopic
image watermarking scheme, wherein watermarks are embedded
in the DCT domain to make the watermarking scheme as robust
and invisible as possible [11]. Yu et al. took advantage of the
block inter-relationships between two views for embedding watermarks in order to improve robustness [12]. Recently, watermarking
schemes based on depth-image-based rendering have been proposed [13,14]. Lin et al. embedded watermarks into the DCT based
center view; this method protects the center view and rendered
views as well [13]. In addition, Kim et al. proposed a dual-tree

http://dx.doi.org/10.1016/j.aeue.2015.10.006
1434-8411/ 2015 Published by Elsevier GmbH.

Please cite this article in press as: Zhou W, et al. Binocular visual characteristics based fragile watermarking scheme for tamper detection
in stereoscopic images. Int J Electron Commun (AE) (2015), http://dx.doi.org/10.1016/j.aeue.2015.10.006

39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63

G Model

64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129

AEUE 51503 18

ARTICLE IN PRESS

W. Zhou et al. / Int. J. Electron. Commun. (AE) xxx (2015) xxxxxx

complex wavelet transform based watermarking scheme that is

more robust than Lins [14]. Besides copyright protection, verication of the integrity and authenticity of the image content is another
important purpose of a watermark; however, stereoscopic image
watermarking schemes rarely focus on ensuring authenticity.
Most of the existing fragile watermarking schemes ensure
authenticity verication for monoscopic images captured from a
single camera. In general, fragile watermarking techniques can be
categorized into two major classes: pixel-wise and block-wise fragile watermarking schemes. The main idea in block-wise fragile
watermarking schemes is that the original image is divided into
non-overlapping blocks and each block contains its own watermarking information. For example, one of the rst block-wise
authentication scheme was proposed by Walton [15]. He divided
the image into 8 8 non-overlapping blocks and embedded the
checksum in the least signicant bit (LSB) of every block. Subsequently, many block-wise fragile watermarking schemes have been
proposed [1619]. If the target image is altered, the image content and the watermark extracted from the tampered blocks do
not match with the other blocks, therefore the tampered blocks can
be detected. In general, block-wise fragile watermarking schemes
are only capable of detecting a major alteration or replacement.
Moreover, these schemes can only identify tampered blocks and not
tampered pixels. Therefore, some pixel-wise fragile watermarking
schemes have been proposed to resolve this problem, wherein the
watermarking information derived from host pixels is embedded
into the host pixels themselves. Therefore, tampered pixels can be
identied by the absence of the watermarking information they are
expected to contain. For example, Zhang et al. proposed a statistical scheme for a pixel-wise fragile watermarking scheme for cases
when the tampered area is not too extensive [20]. Hsu et al. proposed a pixel-wise based scheme that decreases the probability of
mistakes in tamper detection, and enhances the accuracy of authentication. To improve the security of pixel-wise fragile watermarking
schemes [21], Rawat et al. proposed a chaos based watermarking scheme for image authentication and tamper detection [22].
Teng et al. analyzed the security of watermarking scheme [22] and
proposed an improved scheme to enhance the security of fragile watermarking scheme [23]. In addition, Zhang and Wang [24]
proposed a fragile watermarking scheme combining block-wise
and pixel-wise techniques. Their scheme can not only detect the
blocks containing tampered content, but also locate the tampered
pixels. For digital images, modication of content and modication of watermark are not the same. Content removal tampering
destroys the integrity and authenticity of an image, while watermark removal tampering does not affect the authenticity of the
image. Therefore, the verication process in a watermarking system should be able to detect and localize exactly where the content
is tampered. However, the aforementioned fragile watermarking
schemes are unable to distinguish between content removal tampering and watermark removal tampering.
In our previous work [4], block-wise fragile watermarking
scheme was proposed, which have some shortcomings: (1) only
identify tampered blocks and not tampered pixels, (2) unable to
distinguish between content removal tampering and watermark
removal tampering, and (3) the relationship between watermarking capacity and visual transparency have not been thoroughly
explored. In this paper, a novel binocular visual characteristics
based pixel-wise fragile watermarking scheme for stereoscopic
image tamper detection is proposed. In order to achieve a balanced
relationship between watermarking capacity and visual transparency, the binocular just noticeable difference (BJND) model is
applied, which helps guide watermark embedding without degrading the stereoscopic image. A stereoscopic image presents two
offset views; therefore, its total watermarking capacity is more than
that of a monoscopic image. Moreover, because most of the pixels

in the two views of a stereoscopic image match with binocular disparity, when the target image is tampered with, the left and right
views are often modied symmetrically with reasonable binocular
disparity; the inter-correlations in the modications, in such cases,
help improve the performance of tamper detection. This study aims
to integrate binocular visual characteristics and probability theory to improve image tamper detection accuracy and precision.
Further, MD5 hash function and chaotic map are employed for
improving the security of the proposed scheme. The effectiveness
of the proposed scheme is veried through a series of attacks.
The remainder of this paper is organized as follows. In Section
2, the proposed watermarking scheme, including the watermark
embedding and tamper extraction is explained. Experimental
results are presented in Section 3. Finally, conclusions are drawn in
Section 4.

2. Proposed binocular visual characteristics based fragile

watermarking scheme

bn (0, 1)

(1)

where n = 0, 1, . . . is the map iteration index and  is a system parameter. If 3.5699456 <  4, then the logistic chaotic map
becomes chaotic. In this state, the sequence is non-periodic, nonconvergent, and very sensitive to the initial value of b0. Moreover,
all the orbits of the logistic map are dense in the range of the map
[0,1].


Step a2: An index ordered sequence A = a1 , . . ., ai , . . ., aN
such that ba ba ba
is obtained by sorting B =

133
134
135
136
137
138
139
140
141
142
143
144

147
148
149
150
151
152

153

In the self-embedding watermarking scheme, the watermark

embedding position of each pixel is randomly distributed over the
entire stereoscopic image. Therefore, this study adopts a random
sequence generated by a logistic chaotic map to obtain a one-toone pixel-mapping sequence. The steps of generating the pixelmapping sequence are as follows.
Step a1: A random sequence B = (b1 , b2 , . . ., bN ) of length N is
generated using the following logistic chaotic map [25].

132

146

2.1. Logistic chaotic map

131

145

In this section, we explain the proposed watermarking scheme.

Let us consider Il and Ir as the original left and right views of a
stereoscopic image of size M N, and Il and Ir as the tampered left
and right views of the stereoscopic image of size M N. The coordinates of the content pixel is (x,y) and the corresponding mapping
coordinates of the watermark pixel is (i,j).

bn+1 =  bn (1 bn ),

130

b1 , b2 , . . ., bN .
Step a3: The index of the content pixel is set as p(m) and its corresponding mapping index of watermark inserting pixel as p(am ).
Step a4: All pairs of p(m) and p(am ) are recorded to form the
one-to-one pixel-mapping sequence.
Step a5: p(m) and p(am ) are reset to two dimension pixelmapping coordinates.

2.2. Watermark embedding

The BJND model measures the perceptible minimum distortion
threshold of binocular vision for a stereoscopic image; this has
been veried through many psychophysical experiments [26]. In
the following, the derivation of the BJND model is summarized. By
incorporating the luminance and contrast masking effects, as well

Please cite this article in press as: Zhou W, et al. Binocular visual characteristics based fragile watermarking scheme for tamper detection
in stereoscopic images. Int J Electron Commun (AE) (2015), http://dx.doi.org/10.1016/j.aeue.2015.10.006

154
155
156
157
158
159
160
161

162

163
164
165
166
167
168
169
170
171
172
173
174
175
176
177

178

179
180
181
182
183

G Model

ARTICLE IN PRESS

AEUE 51503 18

W. Zhou et al. / Int. J. Electron. Commun. (AE) xxx (2015) xxxxxx

Fig. 1. Flowchart of watermark generation and embedding.

184
185

as considering binocular disparity between two views, the BJND of

the left view is dened as follows:
BJNDl = BJNDl (bgr (i, j + d), ehr (i, j + d), Ar (i, j + d))
= Ac,limit (bgr (i, j + d), ehl (i, j + d))

186

Ar (i, j + d)
Ac,limit (bgr (i, j + d), ehr (i, j + d))

 1/

(2)

195

where d is the horizontal binocular disparity value at pixel coordinate (i,j), bgr (i,j + d) denotes the background luminance level,
Ar (i,j + d) denotes the noise amplitude, ehr (i,j + d) denotes the edge
height, and Ac,limit denotes the contrast masking threshold. Parameter  controls the inuence of the noise in the right view and is
set to 1.25 (more details can be found in [26]). It should be noted
that BJNDl is dependent on the background luminance, the noise
amplitude, and the edge height of the corresponding pixel in the
right view. For a noise-free view, Ac,limit is dened by

196

Ac,limit (bg, eh) = Alimit (bg) + B(bg) eh

187
188
189
190
191
192
193
194

197

The Alimit and B functions are dened as follows:

0.0027 (bg 2 96bg) + 8,

if 0 bg < 8

0.0007(bg 2 32bg) + 1.7,

if 48 bg 255

198

Alimit (bg) =

199

B(bg) = 106 (0.7bg 2 + 32 bg) + 0.07

200
201
202
203
204
205
206
207
208
209

210

211
212
213
214
215
216
217
218
219
220

221
222

(3)

(4)

(5)

Eqs. (4) and (5) are derived using psychophysical experiments,

and the detailed representation of eh and bg can be referred to [26].
The BJND of the right view can be obtained in a similar manner.
The following algorithm briey describes how the watermark
of each pixel is generated and embedded as shown in Fig. 1.
Step b1: Extract ve most signicant bits (MSBs) of the original
left view of the stereoscopic image Il to obtain its residual view Ilr .
Step b2: Compute the BJNDl from the residual view Ilr (i, j) using
Eq. (2); the number of adaptive LSBs k(i,j) in pixel Ir (i,j) for the
modular addition operation can be derived as

k(i, j) =

k max

BJNDl (i, j)

max(BJNDl (i, j)

min(BJNDl (i, j)

) min(BJNDl (i, j)

(6)

where kmax = 3 is the maximum number of LSBs, which will be discussed in Section 3, and max() and min() return the maximum
and minimum of the array, respectively. The symbol   represents
the round up integer operation.
 The watermarking capacity of the
left view is dened as Cw =
k(i, j). The watermarking capacity
i,j
calculation of the right view is the same as that for the left view.
Step b3: Then, employ the values of the residual view I rl (x, y)
and BJNDl (x,y), the pixel coordinates (x,y), the user ID, and the user
secret key k2 to produce a 128-bit hash value, through the following
MD5 hash function:
H(Ilr (x, y)||BJNDl (x, y)||x||y||ID||k2 )
= (d0 (x, y), d1 (x, y), . . ., d127 (x, y))

(7)

where the symbol || denotes a concatenation operation, while dm

denotes the m-th bit of the hash value.
Step b4: Then, the feature needs to be folded, that is, the bit string
of the feature is converted into another bit string whose length
is shorter, because the embedded pixel only has k(i,j) LSBs to be
embedded. This bit string must be concatenated with 0 0 0 0 to
obtain a new bit string of length 132 bits because it must be folded
into k(i,j) bits authentication message for the corresponding pixel,
where 1 k(i, j) 3. The folding equation is shown as follows:
fl (x, y) =

223
224
225
226
227
228
229
230
231

(132/k(i,j)
 )1

(dik(i,j)+1 (x, y)||dik(i,j)+2

232

i=0

(x, y)|| ||dik(i,j)+r (x, y))

(8)

where fl (x,y) is the folding result.

Step b5: The folded bit string is inserted into the k(i,j) LSBs of
the corresponding mapping pixel.
Step b6: Repeat Step b2 to Step b4 until all pixels are processed.
Step b7: The watermark generation and embedding for the original right view of the stereoscopic image is performed in a similar
manner as the original left view.
2.3. Watermark extraction
When there is doubt whether a stereoscopic image has been
tampered with, then the tampered area, if any, can be located
through an authentication message. In this paper, tamper detection
is implemented by a two-stage method, wherein a prime authentication result is obtained in the rst stage, and then the result is
improved in the second stage.
The steps of stage 1 authentication are as follows:
Step c1: Dene the left and right views identity matrices
FL (m,n) = 0 and FR (m,n) = 0, respectively.
Step c2: Step b1 to Step b4 of Section 2.2, as dened above, are
performed, to obtain the folding result f l (x, y).
Step c3: Compare the folding result f l (x, y) with the extracted
k(i,j) LSB of the corresponding mapping pixel of the tampered left
view of the stereoscopic image. If there is a mismatch, we mark
those pixels as invalid and dene FL (x,y) = 1.
Step c4: The stage 1 authentication of the tempered right view
of the stereoscopic image is performed in a similar manner as that
for the tempered left view.
The owchart of stage 1 authentication is shown in Fig. 2.
The proposed scheme, through mapping and the BJND model,
embeds the watermark into the LSBs of the views pixels. As a
result, the FL and FR display randomly distributed isolated regions
if the watermark is altered. If FL and FR include clustered regions,
then the contents of the stereoscopic image in those regions have
been tampered with. Using the predened threshold T, we can
localize the tampered content of the stereoscopic image with high
probability and reduce the probability of false acceptance. Moreover, two views are often modied symmetrically with reasonable
binocular disparity if a stereoscopic image is tampered with, and

Please cite this article in press as: Zhou W, et al. Binocular visual characteristics based fragile watermarking scheme for tamper detection
in stereoscopic images. Int J Electron Commun (AE) (2015), http://dx.doi.org/10.1016/j.aeue.2015.10.006

233

234
235
236
237
238
239
240

241

242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270

G Model
AEUE 51503 18

ARTICLE IN PRESS

W. Zhou et al. / Int. J. Electron. Commun. (AE) xxx (2015) xxxxxx

Fig. 2. Flowchart of stage 1 authentication.

271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287

288

inter-correlations between the modications can help improve the

performance of tamper detection. The steps of stage 2 authentication are as follows:
Step d1: Let Ql (m,n) denote the non-zero number in the b b
(b = 5) pixel-neighborhood of FL (m,n).
Step d2: The identity matrix F L (m, n) of the left view is as follows:
FL (m, n) = 0,

when Ql (m, n) < T

(9)

where T is the predened threshold (details will be discussed in the

following subsection).
Step d3: The calculation process for the identity matrix F R (m, n)
of the right view is performed similar to that for the left view.

n) of the tampered stereoStep d4: The binocular disparity d(m,

scopic image is calculated by using the stereoscopic matching

n)) = 1, and if
algorithm. If FL (m, n) = 1, dene F R (m, n d(m,

F  (m, n) = 1, dene F  (m, n + d(m,

n)) = 1.
R

The owchart of stage 2 authentication is shown in Fig. 3.

297
298

Pfa (T ) = P{Ql (m, n) < T |H0 }

(10)

299

Pfr (T ) = P{Ql (m, n) T |H1 }

(11)

290
291
292
293
294
295
296

300
301

302

The probability of detecting a modication of the content is

dened as:

l1 = P(Ql (m, n) =/ 0|FL (m, n) H0 ) = 1 2w

 

M
N

k(i, j)

i=1
j=1
w =

(12)

MN

303
304
305
306

307

Suppose the image content is tampered with at random, then the

probability of detection can be concluded as Ql (m,n) B (b b, l1 )
through the theory of probability. The probability of false acceptance Pfa (T) is dened as:
Pfa (T ) = P{Ql (m, n) < T |H0 } =

l2 = P(Ql (m, n) =
/ 0|FL (m, m) H1 ) = l1 

T 1

i=0

i
Cbb
(l1 ) (1 l1 )

bbi

(13)

(14)

Similarly, suppose the image content is untampered, then the

detection probability can be concluded as Ql B(b b, l2 ) through
the theory of probability. The probability of false rejection Pfr (T) is
dened as:

Pfr (T ) = P{Ql (m, n) T |H1 } = 1

T 1


i
Cbb
(l2 ) (1 l2 )

bbi

(15)

308
309
310

311

312
313
314
315

316

i=0

For a given threshold T, Pfd (T) and Pfd (t) should satisfy the equation:
Pfd (T ) Pfd (t)

2.4. Threshold T
In this subsection, we will discuss how the threshold T is selected
using the theory of probability [27]. Tamper discrimination can be
formulated as a binary hypothesis test as follows: H0 : the tampered
content of the image pixel, H1 : the content of the image pixel that
is not tampered with. In order to determine a valid hypothesis,
Ql (m,n) is compared with a suitably selected threshold T. For a given
threshold T, the system performance can be measured in terms of
the probability of false acceptance Pfa (T) and the probability of false
rejection Pfr (T):

289

Suppose  is the ratio of the tampered regions and the image

content that is untampered, then the probability of detecting a
modication of the watermark is dened as:

(16)

where Pfd (t) =  Pfa (t) + Pfr (t) (1 ) when the threshold is t [0,
b b].

3. Experimental results
In this section, ve stereoscopic images of size 640 480 are
tested to assess the accuracy and effectiveness of the proposed
scheme, as illustrated in Fig. 4. Since watermarks are embedded
into the stereoscopic images through BJND guiding, the difference between original pixels and watermarked pixels are within
BJND allowable ranges and watermarked stereoscopic images are
imperceptible. The less LSB (the smaller parameter kmax ) are used
for embedding watermark, the higher visual quality of the watermarked stereoscopic image can be achieved. However, the smaller
parameter kmax is chosen, the less watermark capacity can be
obtained. In order to achieve a balanced relationship between
watermarking capacity and visual transparency, the parameter
kmax is set to 3 based on our experimental study.
Smooth and shallow depth stereoscopic images have low capacities, and the PSNR of the watermarked view is relatively higher;
for example, the left and right watermarked views of Newspaper
are 39.49 dB and 39.51 dB, respectively. On the contrary, a stereoscopic image with more texture and deeper depth can hide more
information, and the PSNR of the watermarked view is lower; for
example, the left and right watermarked views of Altmoabit are
37.66 dB and 37.77 dB, respectively. The parameters of logistic map
are chosen as  = 3.845 and b0 = 0.654 (the same value as the one
used in [23]).

Please cite this article in press as: Zhou W, et al. Binocular visual characteristics based fragile watermarking scheme for tamper detection
in stereoscopic images. Int J Electron Commun (AE) (2015), http://dx.doi.org/10.1016/j.aeue.2015.10.006

317
318

319

320
321

322

323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345

G Model

ARTICLE IN PRESS

AEUE 51503 18

W. Zhou et al. / Int. J. Electron. Commun. (AE) xxx (2015) xxxxxx

Fig. 3. Flowchart of stage 2 authentication.

348

For measuring the accuracy of stereoscopic image tamper

detection and localization, two evaluation indicators, i.e., false
acceptance rate (Rfa ) and false rejection rate (Rfr ), are dened as

349

Rfa =

350

Rfr =

346
347

1 Ntd
100%
Nt
N fd
N Nt

100%

(18)

356

357

3.1. Performance evaluation

352
353
354
355

358
359
360
361
362

363
364
365
366

(17)

where N denotes the number of image pixels in a stereoscopic

image, Nt is the actual number of tampered pixels, Ntd is the number of tampered pixels detected correctly, and Nfd is the number
of valid pixels detected incorrectly. The smaller the values of Rfa
and Rfr , the better the performance of the tamper localization and
detection scheme.

351

of the watermarked stereoscopic image. The tampered stereoscopic

image is shown in Fig. 5(a). The stage 1 tamper detection result is
shown in Fig. 5(b). The stage 2 tamper detection result is shown in
Fig. 5(c).

3.1.1. Copy and paste tampering

In the copy and paste attack experiment, the watermarked
stereoscopic image, Balloons, is modied by inserting Balloon in
the left and right views of the stereoscopic image with reasonable
binocular disparity, where Balloon is copied from the same view

3.1.2. Text addition tampering

In this experiment, the watermarked stereoscopic image, shown
in Fig. 6(a), is modied by adding the text Altmoabit at the bottom of the stereoscopic image with reasonable binocular disparity.
Detected tampered region through stage 1 detection is shown in
Fig. 6(b). The stage 2 tamper detection result is shown in Fig. 6(c).

3.1.3. Content removal tampering

In this experiment, some content of the watermarked stereoscopic image is removed without degrading the stereoscopic image
quality. We have removed stone lion from the watermarked
stereoscopic image. The tampered stereoscopic image is shown in
Fig. 7(a). The tamper detection result through stage 1 detection is
shown in Fig. 7(b). The stage 2 tamper detection result is shown in
Fig. 7(c).

Fig. 4. Original stereoscopic images. (a) Balloons, (b) altmoabit, (c) doorower, (d) art, and (e) newspaper.

Please cite this article in press as: Zhou W, et al. Binocular visual characteristics based fragile watermarking scheme for tamper detection
in stereoscopic images. Int J Electron Commun (AE) (2015), http://dx.doi.org/10.1016/j.aeue.2015.10.006

367
368
369
370
371
372

373
374
375
376
377
378
379
380

G Model
AEUE 51503 18

ARTICLE IN PRESS

W. Zhou et al. / Int. J. Electron. Commun. (AE) xxx (2015) xxxxxx

Fig. 5. Copy and paste attack. (a) The tampered stereoscopic image, (b) stage 1 detection, and (c) stage 2 detection.

Fig. 6. Text insertion tampering. (a) The tampered stereoscopic image, (b) stage 1 detection, and (c) stage 2 detection.

Fig. 7. Content removal tampering. (a) The tampered stereoscopic image, (b) stage 1 detection, and (c) stage 2 detection.

Please cite this article in press as: Zhou W, et al. Binocular visual characteristics based fragile watermarking scheme for tamper detection
in stereoscopic images. Int J Electron Commun (AE) (2015), http://dx.doi.org/10.1016/j.aeue.2015.10.006

G Model

ARTICLE IN PRESS

AEUE 51503 18

W. Zhou et al. / Int. J. Electron. Commun. (AE) xxx (2015) xxxxxx

Fig. 8. Collage attack. (a) The tampered stereoscopic image, (b) stage 1 detection, and (c) stage 2 detection.

Fig. 9. Watermark removal tampering. (a) The tampered stereoscopic image, (b) stage 1 detection, (c) stage 2 detection, and (d) authentication result of Rawats scheme [22].

381
382
383
384
385
386
387

388
389
390
391
392

3.1.4. Collage attack

To evaluate the performance under collage attack, a counterfeit
stereoscopic image, as shown in Fig. 8(a) is constructed by copying
stone lions from Fig. 7(a) and inserting it in the relative spatial
location in the foreground. Fig. 8(b) shows the stage 1 detected
tampered region. Fig. 8(c) shows the stage 2 detected tampered
region.
3.1.5. Watermark removal tampering
In this experiment, the watermark of the watermarked stereoscopic image is removed without degrading the stereoscopic image
quality. We have removed the watermark of the stereoscopic
image, ower. The tampered stereoscopic image is shown in

Fig. 9(a). The tamper detection result through stage 1 detection

is shown in Fig. 9(b). The stage 2 tamper detection result is shown
in Fig. 9(c). Fig. 9(d) shows the result when Rawats scheme [22] is
applied directly to each view separately. Although, Rawats scheme
can locate where the tampered part of the stereoscopic image is, it
cannot determine what kind of manipulation is performed on the
watermarked stereoscopic image.
In order to further demonstrate the effectiveness of the proposed scheme, we compare the proposed scheme with Rawats
scheme. Table 1 shows that, for the same attacks, the proposed
scheme performs better than the scheme proposed by Rawat et al.
in terms of Rfa and Rfr . Respective detection results before and after
stage 2 detection are also shown in Table 1.

Table 1
Performance comparisons between the proposed scheme and Rawats scheme.
Schemes

Rawats scheme
Stage-1 authentication
Stage-2 authentication

Indicators

Rfa
Rfr
Rfa
Rfr
Rfa
Rfr

Tamper
Copy and paste tamper

Text addition tamper

Content removal tamper

Collage attack

50.72%
0.00%
18.66%
0.94%
8.42%
0.00%

45.32%
0.00%
5.37%
2.37%
0.36%
0.04%

49.08%
0.00%
37.61%
1.43%
14.49%
0.00%

50.36%
0.00%
51.56%
0.86%
28.84%
0.00%

Please cite this article in press as: Zhou W, et al. Binocular visual characteristics based fragile watermarking scheme for tamper detection
in stereoscopic images. Int J Electron Commun (AE) (2015), http://dx.doi.org/10.1016/j.aeue.2015.10.006

393
394
395
396
397
398
399
400
401
402
403
404
405

G Model
AEUE 51503 18

ARTICLE IN PRESS

W. Zhou et al. / Int. J. Electron. Commun. (AE) xxx (2015) xxxxxx

409

According to Figs. 59 and Table 1, the tampered parts of the

embedded stereoscopic image are located correctly. A two-stage
tamper detection algorithm improves stereoscopic image tamper
detection accuracy and precision.

410

4. Conclusion

406
407
408

429

In this paper, we have proposed an efcient fragile watermarking scheme for stereoscopic image authentication and locating
tampered regions. The main features of the scheme are summarized below: (1) the binocular just noticeable difference (BJND)
model helps guide watermark embedding without affecting perceptual quality, while also increasing watermarking capacity. (2)
MD5 hash function and chaotic map are employed in our scheme
to make the scheme highly secure. (3) Probability theory and
binocular disparity are used to improve image tamper detection accuracy and precision. Experimental results show that our
scheme achieves superior tamper detection and localization accuracy for different attacks; furthermore, it can distinguish between
content removal tampering and watermark removal tampering
attacks.
In future research, we plan to focus more on stereoscopic visual
saliency for fragile watermarking of stereoscopic images to achieve
a balanced relationship between watermarking capacity and visual
transparency, and also to improve the accuracy of tamper localization.

430

Acknowledgements

411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428

431
Q4
432
433
434

435

436
437
438
439
440
441

This work was supported by the National Natural Science

Foundation of China (Grant Nos. 61431015, 61371162, 61302112,
61551346) and the Natural Science Foundation of Zhejiang
Province (Grant Nos. LQ15F020010, LY13F050005, LY14F010004).
References
[1] Zhou WJ, Jiang GY, Yu M, Shao F, Peng ZJ. PMFS: a perceptual modulated feature
similarity metric for stereoscopic image quality assessment. IEEE Signal Process
Lett 2014;21(8):10036.
[2] Tang Z, Zhang Z, Dai X, Yang J, Wu T. Robust image hash function using local
color features. AEU Int J Electron Commun 2013;67(8):71722.
[3] Liu Z, Wang H. A novel speech content authentication algorithm based on
BesselFourier moments. Digit Signal Process 2014;24:197208.

[4] Zhou WJ, Jiang GY, Luo T, Yu M, Shao F, Pneg ZJ. Stereoscopic image tamper
detection and self-recovery using hierarchical detection and stereoscopic
matching. J Electron Imaging 2014;23(2):023022.
[5] Lee MK, Bae KH, Kim JJ, Kim ES. New watermark scheme for copyright protection of the stereo vision system using adaptive disparity matching algorithm.
SPIE Proc Vis Inf Process XII 2003;5108:2109.
[6] Hwang DC, Bae KH, Lee MH, Kim ES. Real-time stereo image watermarking
using discrete cosine transform and adaptive disparity maps. SPIE Proc Multimed Syst Appl VI 2003;5241:23342.
[7] Hwang DC, Ko JH, Park JS, Kim ES. Stereo watermarking scheme based on discrete wavelet transform and feature-based window matching algorithm. Proc
SPIE 2004;5600:18291.
[8] Coltuc D, Caciul Ia, Coanda H. Color stereo embedding by reversible watermarking. IEEE Int Symp Electr Electron Eng (ISEEE) 2010:2569.
[9] Campisi P. Object-oriented stereo-image digital watermarking. J Electron Imaging 2008;17(4):043024.
[10] Chammem A, Mitrea M, Preteux F. DWT-based stereoscopic image watermarking. Proc SPIE 2011;7863:786326.
[11] Niu Y, Souidene W, Beghdadi A. A visual sensitivity model based stereo image
watermarking scheme. In: 3rd European Workshop on Visual Information
Processing (EUVIP). 2011. p. 2115.
[12] Yu M, Wu A, Luo T, Jiang G, Li F, Fu S. New block-relationships based stereo
image watermarking algorithm. In: ICSNC. 2011. p. 1714.
[13] Lin YH, Wu JL. A digital blind watermarking for depth-image-based rendering
3D images. IEEE Trans Broadcast 2011;57(2):60211.
[14] Kim HD, Lee JW, Oh TW, Lee HK. Robust DT-CWT watermarking for DIBR 3D
Images. IEEE Trans Broadcast 2012;58(4):53343.
[15] Walton S. Information authentication for a slippery new age. Dr Dobbs J
1995;20(4):1826.
[16] Suthaharan S. Fragile image watermarking using a gradient image for improved
localization and security. Pattern Recogn Lett 2004;25(16):1893903.
[17] He H, Zhang J, Chen F. Adjacent-block based statistical detection
method for self-embedding watermarking techniques. Signal Process
2008;89(8):155766.
[18] Chuang J, Hu Y. An adaptive image authentication scheme for vector quantization compressed image. J Vis Commun Image Represent 2011;22(5):4409.
[19] Gao G. Composite chaos-based lossless image authentication and tamper localization. Multimed Tools Appl 2013;63(3):94764.
[20] Zhang X, Wang S. Statistical fragile watermarking capable of locating individual
tampered pixels. IEEE Signal Process Lett 2007;14(10):72730.
[21] Hsu CS, Tu SF. Probability-based tampering detection scheme for digital images.
Opt Commun 2010;283(9):173743.
[22] Rawat S, Raman B. A chaotic system based fragile watermarking scheme for
image tamper detection. AEU Int J Electron Commun 2011;65(10):8407.
[23] Teng L, Wang X. Cryptanalysis and improvement of a chaotic system based
fragile watermarking scheme. AEU Int J Electron Commun 2013;67(6):5407.
[24] Zhang X, Wang S. Fragile watermarking scheme using a hierarchical mechanism. Signal Process 2009;89(4):6759.
[25] Kanso A, Smaoui N. Logistic chaotic maps for binary numbers generations.
Chaos Solitons Fractals 2009;40(5):255768.
[26] Zhao Y, Chen Z, Zhu C, Tan YP, Yu L. Binocular just-noticeable-difference model
for stereoscopic images. IEEE Signal Process Lett 2011;18(1):1922.
[27] He HJ, Zhang JS, Tai HM. A wavelet-based fragile watermarking scheme for
secure image authentication. Lect Notes Comput Sci 2006;4283:42232.

Please cite this article in press as: Zhou W, et al. Binocular visual characteristics based fragile watermarking scheme for tamper detection
in stereoscopic images. Int J Electron Commun (AE) (2015), http://dx.doi.org/10.1016/j.aeue.2015.10.006

442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495