An Efﬁcient Group of Pictures Decomposition based Watermarking for

Anaglyph 3D Video

Dorra Dhaou

, Saoussen Ben Jabra

1,2

and Ezzedine Zagrouba

Higher Institute of Computer Science, University Tunis El Manar, LR16ES06 Research Laboratory of Computer Science,

Modeling and Information and Knowledge Processing (LIMTIC), 2 Street Abou Raihane Bayrouni, 2080 Ariana, Tunisia

National Engineering School of Sousse (ENISo), University of Sousse, BP 264 Sousse Erriadh 4023, Tunisia

Keywords:

3D Video Watermarking, Anaglyph, Group of Pictures, Invisibility, Robustness.

Abstract:

Due to the rapid grow of 3D technology, 3D video consumption over the internet is proliferated. 3D Content

protection has then become an important challenging problem for many researchers. Watermarking allows

resolving this problem by embedding a signature into 3D video content. However, only a few works are pro-

posed for 3D anaglyph content protection. In this paper, a new approach of 3D video anaglyph watermarking

is proposed. In fact, the anaglyph 3D technique is considered as the best used technique for creating 3D per-

ception for both images and videos. The proposed approach is based on GOP decomposition where original

video is considered as a set of Group of pictures (GOP). Each GOP will be divided in three types of images:

only one reference image and several B and R images. Then, every type of images will be marked using a

different algorithm based on blue, red or depth channel. This allows to beneﬁt from advantage of every chan-

nel. Experimental results show a high level of invisibility of the proposed approach and a robustness against

several attacks such as compression, noise, ﬁltering, frame suppression, and geometric transformations.

1 INTRODUCTION

Thanks to the availability of 3D TVs, the high-speed

Internet access and the progression of 3D technology,

the popularity of 3D videos increases daily. Indeed,

3D video is getting an enormous public attention re-

cently because of vivid stereo visual experience over

2D video. There are several ways and techniques to

perceive videos in 3D such as the polarized light sy-

stem and the active shutter system. Despite the ad-

vantages of these two techniques, they present a main

drawback which is the hardware requirement. In fact,

the active shutter system requires a special display

with an alternate liquid crystal shutter glasses which

is too expensive because of the needed electronic de-

vice while in the case of the polarized system, a po-

larized display is needed such the pairs of polarized

ﬁlter glasses which are also very expensive. Anot-

her display mode for 3D videos that has taken the at-

tention of researchers is the anaglyph mode. In fact,

an anaglyph 3D image consists of two superimposed

images (called homologues) of complementary colors

representing the same scene but seen from slightly

offset points : usually the left view in red and the right

view in cyan. It is a printed image to be seen in relief,

using two ﬁlters of different colors (3D glasses) pla-

ced in front of each eye of the observer. Anaglyph 3D

images contain two ﬁltered colored images (red and

cyan), one for each eye. The anaglyphic processing is

the cheapest and the most simple way to make the 3D

visual experience accomplishable on ordinary moni-

tors without any special hardware needed, just only

colored glasses.

Due to the quick development of the 3D multi-

media technology, the transmission of 3D video con-

tent over the internet became very easy. However,

this content cannot be distributed illegally without

any protection. For produced views generated from a

depth image-based rendering technique, the left and

right views can be distributed as a 3D content and

also, the center, the left and right views, can be dis-

tributed separately as a 2D content. Protecting these

views from unauthorized distribution becomes a very

signiﬁcant and important issue. Watermarking techni-

que presents an effective solution for this problem be-

cause it provides security and copyright protection to

digital data. Indeed, it consists to insert a robust and

invisible signature into an original 3D video then to

try to detect the presence of the embedded signature

after any attack applied on marked content.

Several 2D video watermarking works have been

proposed in the literature but regarding 3D videos, the

Dhaou, D., Jabra, S. and Zagrouba, E.

An Efﬁcient Group of Pictures Decomposition based Watermarking for Anaglyph 3D Video.

DOI: 10.5220/0006619305010510

In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 4: VISAPP, pages

501-510

ISBN: 978-989-758-290-5

501

ﬁeld is still not mature given the diversity of display

models and the complexity of 3D data. Most of the

techniques that have been proposed for 3D video wa-

termarking consider 3D video as a set of video views

plus a rendered depth image (Multi view Video plus

Depth MVD). Hence, they insert the signature into

this Depth Image Based Rendering (DIBR) using a

still image embedding scheme (Asikuzzaman et al.,

2016) (Rana and Sur, 2015).

Concerning 3D anaglyph video, only few works

are proposed and they didn’t maximize the compro-

mise invisibility / robustness. In order to resolve this

problem, we propose in this paper a new robust ap-

proach of 3D anaglyph video based on GOP decom-

position in which, the original video is considered as

a set of GOP. Each GOP will be divided into three

image types : one reference image I, R, and B images.

The reference image will be marked using a depth ba-

sed embedding, R image will be marked using a red

channel based embedding and B image using a blue

channel based scheme. This allows beneﬁting from

the robustness of every emebedding scheme. In order

to maximize the invisibility of the proposed approach,

the GOP will contain B images more than R images

because the eye is less sensitive to the blue channel.

The remaining of this paper is organized as fol-

lows: the next section presents a state of the art of

anaglyph 3D images and 3D videos watermarking

schemes. Section 3 deals with the different steps of

the proposed approach of anaglyph 3D video water-

marking based on GOP. Section 4 evaluates the per-

formance of the proposed method by giving the expe-

rimental results. Finally, a conclusion and some per-

spectives are drawn.

2 STATE OF THE ART

An anaglyph 3D video can be considered as a se-

quence of anaglyph 3D images. Hence, an anaglyph

3D video watermarking scheme can apply an anag-

lyph 3D image technique into images which com-

pose the 3D video. For this reason, we begun by stu-

dying the existing watermarking techniques proposed

for 3D anaglyph images.

2.1 Overview of Anaglyph 3D Image

Watermarking

The existing approaches proposed for anaglyph 3D

images can be classiﬁed according to two criteria: the

chosen channel for the embedding and the used trans-

formation.

Based on the ﬁrst criteria, signature can be embed-

ded into one of the two pairs of stereo images using

a 2D embedding scheme. The watermarked image

will then be combined with the other views in order

to obtain the marked anaglyph 3D image. We have

noted that in the most of the proposed works, authors

insert the signature in the blue image because the hu-

man eye is less sensitive to the blue color. This allows

obtainning a high level of invisibility but it decreases

the robustness (Prathap and Anitha, 2014).

(Patel and Bhatt, 2015) propose a watermarking

scheme based on Wavelets where the signature is in-

serted into the blue image (right view). The obtained

marked blue image is then combined with the origi-

nal red image (left view) to obtain a marked anaglyph

3D image. For signature extraction, an inverse de-

anaglyph process is used to separate the two stereo-

scopic images and the signature is extracted from blue

image. This method has proven a high level of invi-

sibility but it is not robust to the majority of attacks.

Another watermarking technique is proposed in (Za-

dokar and Rathod, 2015) and (Zadokar et al., 2013)

where the mark is embedded into the three images

composing the anaglyph 3D image : red (left view),

blue (right view), and depth image. This technique

offers a better robustness against different attacks and

high invisibility.

Concerning the used transformation, we have no-

ted that the most of the proposed works are mainly

based on the Wavelet Transformation to embed sig-

nature in the anaglyph 3D images by modifying the

high frequencies in order to maximize the invisibility

and robustness compromise (Patel and Bhatt, 2015),

(Zadokar et al., 2013).

In (Prathap and Anitha, 2014) a simple scheme for

the protection of 3D red-cyan anaglyph images based

on 3D-DWT and the Jacket matrix is proposed. In

this approach, the original image is transformed using

multi-level 3D-DWT and the middle level sub-bands

are divided into blocks. Next, the Jacket matrix is

applied to the middle level sub-band blocks and the

signature is embedded by modifying the diagonal ele-

ments of each block. This method is robust against

several attacks thanks to the use of the decomposition

level of 3D-DWT, the block size, the minimum value

of middle level sub-bands and the watermark strength

factor. In addition, it is a blind watermarking scheme

that does not require the original image during the ex-

traction process. Therefore, this method has a good

invisibility with a PSNR value greater than 51 dB.

(Devi et al., 2016) proposed a robust and optimi-

zed blind encrypted three dimensional red-cyan anag-

lyph image watermarking system. In fact, the propo-

sed approach comprises two phases: trainning phase

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

502

and testing phase. In the training phase, the origi-

nal red-cyan anaglyph cover image is decomposed ap-

plying Discrete Wavelet Transform (DWT) to get the

low and high pass ﬁlter values. The DWT transfor-

med bits (LH and HL) are optimized using genetic al-

gorithm (GA) and are trained using Back Propagation

neural Network (BPN). During the testing phase, a bi-

nary image with a size of 32 × 32 is embedded after

applying the Advanced Encryption Standard (AES)

encryption method into the three dimensional anag-

lyph image by adding it to the optimized transformed

bits. Finally, the extraction of the embedded image

is done by ﬁnding the Eigen feature vectors that give

the characteristic values for the extraction of the wa-

termarked image.

(Munoz-Ramirez et al., 2015) propose to embed

the signature into the frequency domain using the Dis-

crete Cosine Transform (DCT) and quantization of

index modulation (QIM) with its variant Dither Mo-

dulation (DM). (Rakesh and Dr.K.Sri Rama, 2016)

proposed a watermarking algorithm for anaglyph 3D

images based on the Fourier Fractional Transform

(FrFT). In this approach, the watermark is embedded

into the right image (blue channel) of the anaglyph

3D image after applying the 2-D FrFT transformation.

Then, the watermarked right image and left image are

combined to obtain the marked anaglyph 3D image.

At extraction step, the signature is detected by cal-

culating the threshold and comparing it with the de-

tection value.

(Wang et al., 2015) proposed two 3D watermar-

king scheme for 3D anaglyph image. The ﬁrst one

uses Spread Spectrum (SS) technology to embed sig-

nature while the second is based on adaptive Dither

Modulation (DM) with Watsons improved perception

model. The two schemes present a robustness against

several attacks and a high level of invisibility.

An other blind three-dimensional anaglyph image

watermarking scheme for copyright protection is pro-

posed in (Devi and Singh, 2017) where the N-level

nonsubsampled contourlet transform and the Princi-

pal Component Analysis (PCA) are used to embedd

signature. In fact, The binary watermark is encrypted

using Arnold transform. Then, it is embedded into the

selected sub-band of the nonsubsampled contourlet-

transformed anaglyph 3D image using the principal

component analysis.

2.2 Overview of 3D Anaglyph Video

Watermarking

Based on our knowledge, there are only two existing

works proposed for anaglyph 3D videos watermar-

king until now. Indeed, the ﬁrst work (Waleed et al.,

2013) is based on RGB color analysis where the sig-

nature is inserted in all the blue channels of all frames

composing the original anaglyph 3D video using the

Discret Wavelet Transform (DWT). The blue channel

is chosen because the blue color variation is hardly

perceived by the human visual system. For this appro-

ach, a 4-level Discrete Wavelet Transform is perfor-

med to the blue channel and the signature is embed-

ded in the high frequency bands (HH) of that chan-

nel. The detection step is blind and didn’t require the

original signature or the original video during the ex-

traction process.

The second scheme is a blind and invisible water-

marking technique based on scenes change detection.

Since human eye is less sensitive to blue color, the

signature bits are inserted into the high frequency

bands (HH) of the blue channel of the anaglyph fra-

mes where a scene change is detected (Salih et al.,

2015).

Despite the advantages of these two proposed ap-

proaches, they are not robust against the most impor-

tant video attacks. In fact, the insertion of mark only

in the blue images forming the video causes a lack of

robustness if the attack targets the blue images of the

video then the signature will be easily lost. Hence,

an insertion in the different views of the anaglyph

image is necessary to obtain a high level of robust-

ness. Moreover, if the signature is inserted only in

scenes change frames, a fragile watermarking can be

obtained in the case of videos presenting a single or a

small number of scenes.

Table 1: Comparison of the existing approaches.

Proposed

works

Method Invisivility Robustness

(Prathap and

Anitha, 2014)

image + (51 dB) JPEG compression, Filtering,

Noise, Geometric attacks, Contrast

adjustment, Color quantization.

(Devi et al.,

2016)

image + (53 dB) Filtering, Noise, Geometric at-

tacks, Blurring, Histogram Equali-

zation, Color quatization, Gamma

correction.

(Munoz-

Ramirez

et al., 2015)

image + (40 dB) JPEG compression, Noise.

(Rakesh and

Dr.K.Sri Rama,

2016)

image - (27 dB) Not mentioned

(Wang et al.,

2015)

image + (41 dB) Noise, Shrinking, Filtering, cut-

ting, JPEG compression, Lumi-

nance change, Volumetric scaling.

(Devi and

Singh, 2017)

image ++ (69 dB) Filtering, Noise, Geometric at-

tacks, Blurring, Histogram Equali-

zation, Color quantization, Gamma

correction, Contrast adjustment.

(Waleed et al.,

2013)

video ++ (65 dB) Noise, Cropping, Filtering,

Gamma correlation, Intensity

Adjustment,

(Salih et al.,

2015)

video ++ (69 dB) Histogram equalization, JPEG

compression.

An Efﬁcient Group of Pictures Decomposition based Watermarking for Anaglyph 3D Video

503

In table 1, the existing approaches are compared

based on the most important criteria: invisibility and

robustness. We can observe that the majority of the

existing approaches guarantee a high level of invisi-

bility however they don’t resist against the most im-

portant attacks such as MPEG compression and fra-

mes suppression. Moreover, All the existing approa-

ches didn’t evaluate the combination of manipulations

which is an important attack to assess any video wa-

termarking approach.

3 PROPOSED ANAGLYPH 3D

VIDEO WATERMARKING

APPROACH

In order to avoid all the disadvantages of the existing

techniques, a new approach of 3D anaglyph video wa-

termarking based on GOP decomposition is proposed

in this paper. There are two main steps in this scheme:

GOP decomposition and signature embedding.

Giving an original 3D anaglyph video, it will be

decomposed in different Groups of Pictures (GOP).

Then, each GOP will be divided in three types of ima-

ges : reference image that we called I, Blue and Red

images that we called respectively B and R. Every

type of images will be marked using a different em-

bedding algorithm in order to maximize the robust-

ness of the proposed approach. Indeed, B images will

be marked using a scheme which modify only blue

channel, R images will be marked by modifying only

red channel and ﬁnally I image will be marked by

modifying all of red, blue channels in addition of the

depth image. Hence, if any attack removes one type

of images, the detection algorithm will be able to ex-

tract signature from the other types of images. Mo-

reover, in order to increase the invisibility of the pro-

posed scheme, the GOP must contain a larger number

of B images because the human eye is less sensitive

to the blue channel. Embedding process will be based

on Discrete Wavelet Transform (DWT), which allows

obtainning robustness against usual attacks.

The general watermarking scheme consists then of

the following steps :

1. Divide the original anaglyph 3D video into a set

of GOP.

2. Divide each GOP into three types of images : I, B

and R images.

3. Repeat the embedding stages for every GOP:

(a) Embed the signature into the blue and red chan-

nels in addition of the depth image generated

from I image.

(b) Embed the signature into the blue channel of B

images.

4. Combine the marked channels of each type of

images with their complementary original chan-

nels to obtain the marked anaglyph 3D images.

5. Combine all marked GOP in order to generate the

marked anaglyph 3D video.

Basic building block of the proposed approach is

shown in the Figure 1.

Figure 1: The general layout of the proposed approach.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

504

3.1 GOP Decomposition

In order to guarantee a high level of robustness against

attacks and especially against MPEG compression

and frames suppression attacks, the ﬁrst step of the

proposed approach consists in decomposing the ori-

ginal anaglyph 3D video into a set of Group of Pic-

tures (GOP). Indeed, several existing video water-

marking techniques are based on GOP decomposi-

tion which is used in MPEG compression process. In

(Liu and Zhao, 2010), a watermarking algorithm ba-

sed on GOP decomposition was proposed. In fact,

the original 2D video sequence is divided into GOP.

Then, the signature is embedding in the coefﬁcients

obatained after applying the 1D DFT transformation

on each GOP in order to improve the robustness of

the algorithm. In (Liu et al., 2008) a real-time video

watermarking scheme is presented by embedding the

watermark bits both in I and P frames based on GOP

decomposition. An uncompressed video watermar-

king system based on Hidden Markov Model (HMM)

and Artiﬁcial Neural Network (ANN) is proposed in

(ELBAS¸I, 2010). In fact, the proposed scheme splits

the video sequences into GOP with HMM. Then, Por-

tions of the binary watermark are embedded into each

GOP with a wavelet domain watermarking algorithm.

The embedding process is the standard additive al-

gorithm in low pass (LL) and high pass (HH) bands

in the wavelet domain. This proposed system increa-

ses the robustness against geometric and temporal at-

tacks, and increases the quality of the marked video.

In order to choice the GOP size for the proposed

approach, we consider that it is difﬁcult to have large

changes in a video sequence after one second. Hence,

we choose one GOP per one second. Then, each GOP

is divided into three types of images : I, B and R ima-

ges. As the I image will be marked with a depth based

scheme, one only reference image is selected per one

GOP in order to guarantee a high invisibility. In ad-

dition, the number of B images should be superior to

R images because B images will be marked by mo-

difying blue channel which is the less sensitive for a

human eye. Each GOP is then composed of a single

reference image I, two R images and the rest will be

considered as B images.

Figure 2: GOP decomposition.

For the selected test videos, we have about 25 fra-

mes per second. Hence, each GOP will begin by one I

image followed by two identical parts each containing

one R image and 11 successive B images as shown in

Figure 2.

3.2 Signature Embedding

In order to increase the robustness and the invisibi-

lity of the proposed approach, a DWT based scheme

is chosen. Figure 3 shows the general architecture of

the signature embedding process which is decompo-

sed in several steps. Indeed, to embed the signature,

a three levels of wavelet decomposition are applied

on the different images of each GOP using the Haar

ﬁlter wavelet which is known for its simplicity and

speed of computation (Zhang, 2009) (Zheng et al.,

2007). DWT transformation will be applied on re-

spectively blue and red channels for B and R images

while for the reference image, DWT will be applied

on its depth image and both blue and red channels.

For each image, the watermark information is em-

bedded by adding the signature to the low frequency

sub-band (LL3) of each image in order to increase the

robustness of the signature, since it is the most sig-

niﬁcant band, which contains more information and

includes the most energy of the image rather than the

others bands which include edge components of hori-

zontal, vertical and diagonal directions. The marked

images will be obtained by applying the inverse DWT

to the marked sub-bands. Finally, the marked B and

R images are obtained by combining respectively the

corresponding original red and blue channels with the

marked blue and red channels. The marked reference

image will be obtained by combining marked red and

blue channels with the depth image.

Given an original anaglyph 3D video, the signa-

ture embedding procedure consists then of the follo-

wing steps :

1. Apply a three levels of wavelet decomposition to

the original channel.

2. Apply a three levels of wavelet decomposition to

the signature.

3. Add the coefﬁcients of the low frequency band of

the signature (CW

) to those of the original chan-

nel (C

) using an invisibility factor (α) to obtain

the marked coefﬁcients (C

)0 according to the fol-

lowing equation :

0 = C

+ α ∗ CW

(1)

4. Apply the inverse decomposition wavelet trans-

form (I-DWT) to generate the marked channel.

An Efﬁcient Group of Pictures Decomposition based Watermarking for Anaglyph 3D Video

505

5. Combine the marked channel to the other chan-

nels in order to obtain anaglyph 3D marked

image.

6. Combine all marked anaglyph 3D images to ge-

nerate the marked anaglyph 3D video.

Figure 3: Layout of the embedding process.

3.3 Extraction Process

The general detection scheme is decomposed in se-

veral steps where the two ﬁrst ones are identical to

those of the embedding scheme.

In fact, given a marked anaglyph 3D video, it will

be decomposed in several GOP with the same frame

number chosen at embedding phase. Then, R, B and

reference images will be detected for every GOP. For

each type of image, the inverse embedding process

will be applied on the corresponding channel in which

the signature was embedded.

Given an input anaglyph 3D video, the extraction

algorithm consists of the following steps:

1. Divide the given video into several GOP.

2. Apply the 3-levels DWT to the different channels

as done in embedding process.

3. Extract the coefﬁcients of the signature from the

low frequency band (LL3) channels.

4. Apply the inverse decomposition wavelet trans-

form (I-DWT) to recover the watermark.

4 EXPERIMENTAL RESULTS

In order to evaluate the proposed approach, a set of

robustness and invisibility tests are applied on a data-

base containing ﬁve original anaglyph 3D videos ha-

ving different characteristics of background texture,

movement, resolution and frame number.

Table 2 shows the characteristics of the chosen videos.

Moreover, for a same video, we have tested different

sizes in order to evaluate the GOP size choice.

Table 2: Video test characteristics.

Videos

Frame

number

Resolution

Background

Texture

Movement

Video 1 250 480 × 360 Uniform Slow

Video 2 750 480 × 360 Uniform Slow

Video 3 1500 640 × 180 Textured Medium

Video 4 2250 1280 × 720 Textured Rapid

Video 5 3000 480 × 360 Uniform Slow

Three different binary images (Lab Logo, Lab

name and Author name) with different sizes (245 ×

140, 160 × 50 and 80 × 25) are used as signatures

(Figure 4).

Figure 4: Original watermarks: (a) Lab logo, (b) Lab name,

4.1 Invisibility Evaluation

Figure 5 presents embedding results. No differences

can be observed between the original and marked vi-

deos.

In order to prove this high level of invisibility, the

average of the PSNR (Peak Signal to Noise Ratio) be-

tween original and marked frames is calculated. Ge-

nerally, a good invisibility is marked with a PSNR su-

perior to 40 dB.

Table 3 presents the obtained PSNR values for test

videos using the different signatures. These values

show that the proposed watermarking has a high invi-

sibility. In fact, the minimum PSNR is about 69 dB

and its maximum is about 72 dB.

Moreover, we evaluated the impact of video size

variation on the PSNR value in order to test the invisi-

bility of the proposed approach. Obtained results are

shown in Figure 6 which prove that even by increa-

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

506

Figure 5: a- Original anaglyph frame. b- Watermarked

anaglyph frame.

Table 3: PSNR values of test videos.

Watermarked

video sequences

Watermark PSNR

Video 1

(a)

(b)

(c)

70.99

71.90

72.91

Video 2

(a)

(b)

(c)

70.93

71.82

71.89

Video 3

(a)

(b)

(c)

70.89

71.32

71.64

Video 4

(a)

(b)

(c)

70.77

70.68

70.56

Video 5

(a)

(b)

(c)

70.65

70.52

69.95

sing the size of the test video, the visual quality of the

marked video does not degrade.

Figure 6: Inﬂuence of invisibility based on video size.

4.2 Robustness Evaluation

To evaluate the robustness of the proposed approach,

different unintentional and malicious attacks are tes-

ted. These attacks were applied directly to the mar-

ked video. The Normalized Correlation value (NC)

and the Bit Error Rate (BER) are measured between

the original and the extracted marks in order to decide

the result of robustness against an attack.

In order to evaluate robustness against unintenti-

onal manipulations, geometric attacks are ﬁrstly tes-

ted as; cropping, rotation with various angles, scaling,

and resizing. Also some noises and ﬁlters were ap-

plied to the marked video and then we try to detect

the signature and calculate the BER and the Correla-

tion between extracted and original mark. Obtained

results of the Correlations values of the different test

anaglyph 3D videos without attacks and after attacks

are shown in Table 4.

The proposed approach is robust to cropping,

thanks to the repetition of the mark in all the frames.

The robustness against rotation, scaling and cropping

is obtained thanks to the use of DWT transformation,

which is invariant to these attacks.

The proposed technique is also robust against dif-

ferent noises such as salt and pepper, Gaussian, Pois-

son and Speckle, ﬁlters (Average and Gaussian ﬁlter)

and blurring where the BER was almost about 0.001

with a correlation about 0.9 almost times. This ro-

bustness was enhanced thanks to the invariance of the

used DWT transformation.

In second part, temporal attacks are tested by ap-

plying given transformation to the marked video as

deleting some frames and changing the frames orders

from the marked video.

In addition, we have tested frames transposition

and dropping until 70% into the same scene. We have

noted that the detection succeeds after all these at-

An Efﬁcient Group of Pictures Decomposition based Watermarking for Anaglyph 3D Video

507

Table 4: Correlation values of different test videos after at-

tacks.

Attacks Video 1 Video 2 Video 3 Video 4 Video 5

Salt & Pepper

noise 0.1

0.9997 0.9995 0.9998 0.9996 0.9994

Gaussian

noise

0.9999 0.9998 0.9999 0.9997 0.9996

Poisson

noise

0.9997 0.9997 0.9996 0.9998 0.9995

Speckle

noise

0.9997 0.9998 0.9997 0.9995 0.9997

Average

Filter

0.9992 0.9994 0.9993 0.9994 0.9991

Gaussian

Filter 9x9

0.9987 0.9979 0.9969 0.9990 0.9989

Blurring 0.9998 0.9999 0.9996 0.9997 0.9995

Rotation

(50%)

09996 0.9995 0.9997 0.9998 0.9993

Rotation

(180%)

0.9958 0.9945 0.9961 0.9939 0.9986

Cropping

(50%)

0.9999 0.9996 0.9994 0.9996 0.9999

Scaling

0.9998 0.9996 0.9999 0.9997 0.9999

Frame

suppression

0.9993 0.9994 0.9995 0.9994 0.9992

Frame

swapping

0.9998 0.9997 0.9998 0.9994 0.9996

Resizing 0.9995 0.9993 0.9997 0.9996 0.9994

Histogram

equalization

0.9998 0.9995 0.9996 0.9997 0.9995

MPEG

compression

0.9999 0.9998 0.9999 0.9999 0.9997

Intensity

adjustment

1 0.9978 0.9987 0.9991 0.9999

No attacks 1 1 1 1 1

tacks. This is obtained thanks to the repetition of the

signature embedding in different GOP. This fact in-

creases the percentage of detection success even if a

part of the sequence is missed or destroyed.

Moreover, Our scheme can withstand various ot-

her attacks such as histogram equalization and Inten-

sity adjustment.

Then, the proposed scheme is evaluated against

the common video compression standards with bit

rate changing. To simulate real cases the mediaco-

der (http://www.mediacoderhq.com) software appli-

cation was adapted to transcode the marked video.

In fact, this transcoder uses different video codecs to

transcode the media to various formats. We have tes-

ted the MPEG-4 codec with a variable bit rate. Obtai-

ned results show the robustness against compression

where the BER was close to 0.0002 and the corre-

sponding correlation close to 0.99. The mark de-

tection is achieved by using the GOP decomposition

on which is based the MPEG compression. We have

also tested some combination of attacks such as rota-

tion with noise, ﬁltering with geometric attacks and

frame suppression combined with geometric attacks.

We observe that the signature can be detected after

these manipulations which are considered as an im-

portant attacks for any video watermarking.

Finally, we noted that the proposed approach is ro-

bust against any manipulation which can attack either

only blue or red channel. This is obtained thanks to

embedding the signature in different channels of the

frames.

Figure 7 illustrates the mean BER value of the dif-

ferent anaglyph 3D video samples without attacks and

after attacks.

Figure 7: The mean BER value of different test videos.

4.3 Comparative Study

In order to assess the proposed approach, obtained re-

sults are compared with those of existing methods.

Since there exist only two anaglyph 3D video wa-

termarking schemes (Salih et al., 2015) and (Waleed

et al., 2013), we use these two techniques for the com-

parison. This last one is based on the two main fac-

tors: invisibility and robustness.

4.3.1 Invisibility Comparison

Figure 8 illustrates the obtained PSNR values by the

proposed approach and the two existing techniques in

(Salih et al., 2015) and (Waleed et al., 2013) and it

shows the high quality of our proposal.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

508

Figure 8: Invisibility comparison based on PSNR criteria.

In spite of the high visual quality of the existing

approaches, the proposed scheme allows the best in-

visibility rate where the PSNR is around 73 dB.

4.3.2 Robustness Comparison

To compare the robustness of the proposed technique

with existing approaches, we have distinguished two

categories of attacks: usual attacks as geometrics, noi-

ses, ﬁltering, scaling and developed attacks as com-

pression, frame suppression, and combined attacks.

The proposed technique shows a good robustness

against these two types of attacks. However, a com-

parison of obtained results with previous techniques

is important to assess the proposed approach.

Table 5 resumes the robustness of each technique.

The proposed technique shows robustness to almost

of attacks whereas other techniques have some limits.

Considering usual attacks, the proposed technique has

the best robustness to ﬁlter attacks. In fact, it is robust

to Gaussian, average ﬁlter whereas (Salih et al., 2015)

and (Waleed et al., 2013) resist only to Gaussian ﬁlter.

Proposed approach has the best robustness against

noises such as Salt and Pepper , Gaussian, Poisson

and Speckle noise where the NC and the BER were

almost equal to 1 and 0 respectively. For (Salih et al.,

2015) and (Waleed et al., 2013), the NC is close to

0.9.

For geometric attacks, the proposed scheme and

the two existing schemes are robust to rotation, crop-

ping and resizing attacks with NC value equal to 0.9.

The proposed technique is robust to frame ba-

sed attacks as frame suppression and frame swapping

thanks to the embedding in all the video frames. The

proposed technique shows the best robustness against

MPEG compression whereas (Salih et al., 2015) and

(Waleed et al., 2013) are robust only to JPEG com-

pression.

Table 5: Robustness comparison between the proposed ap-

proach and the two existing techniques.

Attacks

Proposed

approach

(Salih et al., 2015) (Waleed et al., 2013)

Salt & Pepper

noise 0.1

Robust Robust Robust

Gaussian noise Robust Robust Robust

Poisson noise Robust Not Robust Not Robust

Speckle noise Robust Not Robust Not Robust

Average Filter Robust Robust Robust

Gaussian Filter

9x9

Robust Robust Robust

Blurring Robust Robust Robust

Rotation

(50%)

Robust Robust Robust

Rotation

(180%)

Robust Robust Robust

Cropping

(50%)

Robust Robust Robust

Scaling

Robust Robust Robust

Frame

suppression

Robust Not Robust Not Robust

Frame

swapping

Robust Not Robust Not Robust

Resizing Robust Robust Robust

Histogram

equalization

Robust Robust Robust

JPEG

compression

Robust Robust Robust

MPEG

compression

Robust Not Robust Not Robust

Intensity

adjustment

Robust Robust Robust

Combined

attacks

Robust Not tested Not tested

No attacks Robust Robust Robust

5 CONCLUSION

With the development of digital image and video pro-

cessing, anaglyph 3D images and videos consumption

over the internet is proliferated. To protect this type

of 3D video, a robust and invisible anaglyph 3D video

watermarking approach based on GOP decomposition

and a Discrete Wavelet Transformation is proposed in

this paper. In this approach, the original video is de-

An Efﬁcient Group of Pictures Decomposition based Watermarking for Anaglyph 3D Video

509

composed to several GOP which are then divided in

reference, B and R images. Every type of images was

marked using a different embedding scheme. In fact,

reference image was marked by modifying its corre-

sponding depth image and both red and blue channels;

R images were marked by modifying its red channel

and ﬁnally, signature was embedded in B images by

modifying its blue channel. This allows maximizing

the robustness of our approach and obtaining a high

level of invisibility. In order to maximize the per-

formance of our approach, we choose 25 frames by

GOP (1 second = 1 GOP) and for each GOP we se-

lected the ﬁrst image as a reference image, 2 R ima-

ges and the rest will be considered as B images be-

cause the human eye is less sensitive to the blue co-

lor. Experimentations show that the proposed appro-

ach is robust against several attacks such as geometric

attacks, frame suppression, compression, noises and

ﬁltering with a high visual quality level. In addition,

The comparison of the proposed approach with ex-

isting techniques shows the good performance of the

proposed scheme to resist almost attacks types with a

good invisibility.

As future work, we will try to enhance the propo-

sed approach in order to be robust against other in-

tentional attacks and especially against collusion and

camcording.

REFERENCES

Asikuzzaman, M., Alam, M. J., Lambert, A. J., and Picke-

ring, M. R. (2016). Robust dt cwt-based dibr 3d video

watermarking using chrominance embedding. IEEE

Transactions on Multimedia, 18(9):1733–1748.

Devi, H. S., Imphal, T., and Singh, K. M. (2016). A ro-

bust and optimized 3d red-cyan anaglyph blind image

watermarking in the dwt domain.

Devi, H. S. and Singh, K. M. (2017). A novel, efﬁcient, ro-

bust, and blind imperceptible 3d anaglyph image wa-

termarking. Arabian Journal for Science and Engi-

neering, pages 1–13.

ELBAS¸I, E. (2010). Robust multimedia watermarking: hid-

den markov model approach for video sequences. Tur-

kish Journal of Electrical Engineering & Computer

Sciences, 18(2):159–170.

Liu, S., Chen, T., Yao, H., and Gao, W. (2008). A real-time

video watermarking using adjacent luminance blocks

correlation based on compressed domain. In Intelli-

gent Information Hiding and Multimedia Signal Pro-

cessing, 2008. IIHMSP’08 International Conference

on, pages 833–836. IEEE.

Liu, Y. and Zhao, J. (2010). A new video watermarking

algorithm based on 1d dft and radon transform. Signal

Processing, 90(2):626–639.

Munoz-Ramirez, D. O., Reyes-Reyes, R., Ponomaryov, V.,

and Cruz-Ramos, C. (2015). Invisible digital color

watermarking technique in anaglyph 3d images. In

Electrical Engineering, Computing Science and Auto-

matic Control (CCE), 2015 12th International Confe-

rence on, pages 1–6. IEEE.

Patel, R. and Bhatt, P. (2015). Robust watermarking for

anaglyph 3d images using dwt techniques. Internati-

onal Journal of Engineering and Technical Research

(IJETR), 3(6).

Prathap, I. and Anitha, R. (2014). Robust and blind water-

marking scheme for three dimensional anaglyph ima-

ges. Computers & Electrical Engineering, 40(1):51–

58.

Rakesh, Y. and Dr.K.Sri Rama, K. (2016). Digital water-

marked anaglyph 3d images using frft. International

Journal of Computer Trends and Technology (IJCTT),

41(2):77–80.

Rana, S. and Sur, A. (2015). 3d video watermarking

using dt-dwt to resist synthesis view attack. In Signal

Processing Conference (EUSIPCO), 2015 23rd Euro-

pean, pages 46–50. IEEE.

Salih, J. W., Abid, S. H., and Hasan, T. M. (2015). Imper-

ceptible 3d video watermarking technique based on

scene change detection. International Journal of Ad-

vanced Science and Technology, 82:11–22.

Waleed, J., Jun, H. D., Hameed, S., Hatem, H., and Majeed,

R. (2013). Integral algorithm to embed impercepti-

ble watermark into anaglyph 3d video. International

Journal of Advancements in Computing Technology,

5(13):163.

Wang, C., Han, F., and Zhuang, X. (2015). Robust digital

watermarking scheme of anaglyphic 3d for rgb color

images. Int. J. Image Process.(IJIP), 9(3):156.

Zadokar, S. R., Raskar, V. B., and Shinde, S. V. (2013).

A digital watermarking for anaglyph 3d images. In

Advances in Computing, Communications and Infor-

matics (ICACCI), 2013 International Conference on,

pages 483–488. IEEE.

Zadokar, S. R. and Rathod, R. B. (2015). A robust dwt

watermarking for 3d images. IJETT, 2(1).

Zhang, Y. (2009). Blind watermark algorithm based on hvs

and rbf neural network in dwt domain. WSEAS Tran-

sactions on Computers, 8(1):174–183.

Zheng, D., Liu, Y., Zhao, J., and Saddik, A. E. (2007).

A survey of rst invariant image watermarking algo-

rithms. ACM Computing Surveys (CSUR), 39(2):5.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

510