WATERMARKING OF COMPRESSED VIDEO BASED ON DCT

COEFFICIENTS AND WATERMARK PREPROCESSING

Samira Bouchama

Research Center on Scientific and Technical Information,5 rue des trois freres Aissou, Ben Aknoun, Algiers, Algeria

Latifa Hamami

National Polytechnic School, 10 Hassen Badi Avenue, BP 182, El-Harrach, Algiers, Algeria

Hassina Aliane

Research Center on Scientific and Technical Information,5 rue des trois freres Aissou, Ben Aknoun, Algiers, Algeria

Keywords: Video watermarking, H.264/AVC standard, Quantized DCT coefficients, Intercorrelation function.

Abstract: Considering the importance of watermarking of compressed video, several watermarking methods have

been proposed for authentication, copyrights protection or simply for a secure data carrying through the

Internet. Applied to the H.264/AVC video standard, in most of cases, these methods are based on the use of

the quantized DCT coefficients often experimentally or randomly selected. In this paper, we introduce a

watermarking method based on the DCT coefficients using two steps: the first one consists in a watermark

pre-processing based on similarity measurement which can allow to adapt the best the watermark to the

carrying coefficients of low frequencies. A second step takes advantage from the coefficients of high

frequencies in order to maintain the video quality and reduce the bitrate. Results show that it is possible to

achieve a very good compromise between video quality, embedding capacity and bitrate.

1 INTRODUCTION

Today, with the progress in network services and

multimedia applications, the interest of

watermarking compressed documents is becoming

more and more important due to the easy

transmission of data and all the issues that follow.

With the availability of video sharing and

dissemination services through Internet, it becomes

important to develop appropriate tools for

authentication or copyright protection or just to offer

a secure data transmission service without using

additional channels.

Indeed, dealing with compressed video, several

watermarking methods have been proposed for the

H.264/AVC, the newest video coding standard

which is known for its coding

efficiency (Richardson, 2003), ( ITU-T, 2005).

These methods generally use the motion vectors or

the quantized discrete cosine transform (DCT)

coefficients in order to propose fragile or robust

video watermarking solutions.

Most of methods that use the DCT coefficients

exploit the human visual properties to reduce the

degradation caused by the watermark embedding.

This degradation is propagated by the inverse

quantization and the inverse DCT operations at the

decoding step, and also by the Intra and Inter

predictions. Thus, the video quality can be improved

by selecting the appropriate areas that respond to

some constraints such as embedding the watermark

according to the choice of the coding mode or other

DCT blocks’ properties.

In this context, we have worked on designing a

watermarking method for the video codec

H.264/AVC, useful for video authentication in order

to ensure the protection of data against unauthorized

modification, to check credibility and the origin

source of the video and respond this way to the need

for reliable information.

114

Bouchama S., Hamami L. and Aliane H..

WATERMARKING OF COMPRESSED VIDEO BASED ON DCT COEFFICIENTS AND WATERMARK PREPROCESSING.

DOI: 10.5220/0003321601140119

In Proceedings of the International Conference on Imaging Theory and Applications and International Conference on Information Visualization Theory

and Applications (IMAGAPP-2011), pages 114-119

ISBN: 978-989-8425-46-1

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

However, video watermarking methods cannot

be designed without thinking of three important

constraints: embedding capacity, video quality, and

video bitrate. Our main motivation, though, is to

arrive at a compromise between these constraints

and if possible reduce the bitrate for a better use

within the network applications.

The idea of decreasing the bitrate was suggested

in (Wang and Hsu, 2008) where the authors

proposed a fragile watermarking algorithm for JPEG

image, based on modifying the last nonzero

coefficient in each DCT quantized block. The

proposed algorithm provides an authentication

capability, and decreases the size of JPEG

compressed-domain images with no visual impact.

Our proposal is based on the quantized

significant DCT coefficients where both of the

coefficients of low and high frequencies are used in

order to obtain a sufficient embedding capacity and

maintain or reduce the bitrate. A watermark

preprocessing is performed by applying a rotation

several times in order to find the best position of the

watermark in the video and maintain the video

quality.

In the next section we describe the different

stages of our method. Results are presented in

section 3. Finally, we give the conclusion in

section 4.

2 RELATED WORK

Several watermarking methods based on the DCT

coefficients and watermark preprocessing were

proposed for the H.264/AVC video codec (figure 1).

However, they are different in the choice of the

block and the DCT coefficients for the watermark

embedding and in the appropriate watermark

processing. The aim is to respond to the application

needs in terms of robustness or fragility, video

quality and bitrate. Among these works, Quan and

Hong (2008) proposed a real time video

watermarking method based on the low frequencies

DCT coefficients features in I frame on H.264 to

generate the watermark information. Middle

frequencies of 4x4 DCT blocks are used to embed

the watermark, and the blocks with a DC coefficient

equal to zero are avoided; the aim is to maintain the

compromise between robustness and invisibility.

Chen, Lai and Chang (2009) proposed a video

watermarking method based on two algorithms for

low and high energy blocks of Intra frame in order

to consider the high frequency noise attack and the

low-pass filter attack. The watermark is

preprocessed by an encryption using a

Torus

Automorphisms algorithm and the Secret Image

Sharing technology to increase robustness.

Another robust video watermarking scheme is

based on the choice of one middle frequency in the

4x4 DCT blocks for the embedding process. In order

to accommodate the computational complexity of

the video codec, the 2 D 8 bits watermark undergo a

preprocessing based on five steps (DCT

decomposition and zigzag scanning, normalization,

frequency masking, transformation, reduction of the

level) transforming it to a simple 1D sequence

(

Zhang, Ho, Qiu and Marziliano, 2007).

Figure 1: General scheme of embedding watermark in

DCT coefficients of the H.264 video codec.

Our method is also based on the quantized

significant DCT coefficients where two coefficients

are used per block: the coefficients of low

frequencies would ensure the necessary embedding

capacity for the application, and the coefficients of

high frequencies would reduce the bitrate and

maintain the video quality. Embedding the

watermark is performed after an initial step which

consists in preprocessing the Watermark by applying

a rotation several times and calculating each time the

intercorrelation coefficient between the watermark

signal and the carrying coefficients transformed to a

binary signal. The aim is to find the best position of

the watermark in the video so that the introduced

distortion is reduced.

3 DESCRIPTION OF THE

METHOD

3.1 Embedding Position and Formula

In a first step, we choose among the significant

quantized DCT coefficients, the position of the

coefficients which would carry the watermark.

Choosing the low frequencies would allow to satisfy

the embedding capacity required for the application,

thus, to embed the watermark, the significant

coefficients in position k= 4 according to the zigzag

scan, are selected as illustrated in figure 2.

Input

signal

(Macroblock)

Watermark

pre-

processing

Prediction

(Intra/inter)

Transf.

(Integer DCT)

and Quant.

Entropy

coding

Bit-

stream

WATERMARKING OF COMPRESSED VIDEO BASED ON DCT COEFFICIENTS AND WATERMARK

PREPROCESSING

115

Figure 2: A zigzag scan in a block of DCT coefficients.

Coefficients equal to 1 are not used because of

the serious increase in bitrate they may cause. This

is because 1 have a special processing in the

entropy coding stage of the video codec

(Richardson, 2003). The DC coefficients are avoided

because of the high degradation they may introduce,

as well as the high frequencies coefficients because

they are mainly composed from values equal to 0

or 1.

In this case the majority of coefficients planned

for embedding the watermark would be even

coefficients. An example of a video histogram of

significant DCT coefficients of the position k = 4 is

given in figure 3.

Figure 3: Histogram of the quantized significant DCT

coefficients of position k=4 of Container sequence.

For the embedding process, simple methods are

preferred such as the Least Significant Bit (LSB)

method, in which the LSB of the coefficient is

replaced by the watermark bit. That means that the

bit 0 is preferebly embedded in an even value in

order not to change the original value of the

coefficient. However , the watermark may be

composed of bits “1” more than “0”, and because we

deal with more even coefficients than with odd ones

as mentionned previousely; it may be better to

change the embedding formula so that the LSB of

the coefficient is replaced, not by the watermark bit,

but by its complement. In this case we need to

reserve one bit during the embedding process to

know wich formula has been used.

3.2 Watermark Pre-processing

We suppose that our watermark has been encrypted

to ensure its security. Our objective is to find the

best transformation of the watermark in order to

correspond - as much as possible- the bits “0” to

even coefficients and the bits “1” to the odd ones.

To do so, we apply a rotation by one several

times to the watermark. In each rotation, the last bit

of the watermark becomes the first one and in each

time we calculate the intercorrelation coefficient

between the watermark signal (W) and the

coefficients (transformed to binary signal C).

Considering W and C two binary sequences of

the same length N:

W={W1=1, W2, .. , WN}, C={C1=1,C2,.. , CN}.

The intercorrelation function for positive delays

is given by:





(

, 

)

∑













,  = 0,1,2, … ,  − 1 (1)

In our case d=0.

To clarify this point an example is given:

Let’s consider the sequences of dimension N= 8.

is composed from coefficients of position k=4 in

the block and W is the Watermark.

={2, -3, 6, 4, 5, 2,-2,3}, W={0,0,1,0,1,1,0,1}.

The preprocessing consists in the following

steps:

- C

is replaced by the binary sequence :

C’

= {1,-1,1,1,-1,1,1,-1}, 1 for even coefficients

and -1 for odd coefficients.

- If we consider that the bit 0 is preferably

embedded in an even coefficient, we transform

the watermark W to W1={1,1,-1,1,-1,-1, 1,-1} ,

where the bit 1 replaces 0 and -1 replaces 1, then

we apply a rotation (N-1) times to obtain:

W2={-1,1,1,-1,1,-1,-1, 1 },

W3={1,-1,1, 1,-1,1, -1, -1},

…,

W8= {1,-1,1,-1,-1, 1,-1,1}.

The highest value of the correlation coefficient is

given for W3 or W6 (Cor = 6). The watermark

corresponding to W3 is: W’={0,1,0,0,1,0,1,1}.

After the embedding process (using the LSB

method) the watermarked DCT coefficients become

={2, -3, 6,4,5,2,-3,3} where only one coefficient

is modified. Besides, we can check that if we had

k=0

k=4

k=11

IMAGAPP 2011 - International Conference on Imaging Theory and Applications

116

embedded the signature without preprocessing we

would have about four coefficients modified in the

sequence

During the embedding process we need to

reserve coefficients to save the rotation dimension

(three, in the example) so that we can deduce W

from W’ at the detection stage. Of course, the

reserved coefficients won’t be involved in the

intercorrelation calculation. In our case, their

number is experimentally estimated to 11

coefficients.

3.3 Video Quality Improvement

and Bitrate Reduction

After the watermark preprocessing and the

embedding process as described in the previous

section, the coefficients of C

for which the

watermark bits are causing a modification, are

restored to their initial value after modifying the

coefficients of position k=11 (C

) instead of C

The modification consists of transforming all the

coefficients of the position k=11 to even values for

those corresponding to the unchanged coefficients

, and C

become odd for those corresponding to

the coefficients of C

supposed to be modified by the

bit embedding. The idea is to exploit C

to make the

difference between the unchanged coefficients and

the restored ones, and because of the position of C

comparing to C

in the block (figure 2), we expect

that the modification of C

would introduce a

smaller degradation.

The only coefficients of C

that won’t be

restored are those corresponding to the coefficients

=0. The embedding and extracting processes are

presented in the next section.

3.4 Watermark Embedding/ Extracting

Embedding and extracting the watermark may be

summarized as follows:

- Extract the sequence of M significant DCT

quantized coefficients used for the

watermarking, and transform it to a binary

signal composed of 1.

- Reserve about 12 bits for the additional

information.

- Apply the watermark preprocessing:

 Transform the watermark signal W (composed of 0

and 1) to binary signal composed of 1.

 Apply a rotation (N-1) times, with N= M-12, and

calculate the correlation coefficient between the

watermark and the binary signal representing the

coefficients, using the formula (1).

 Deduce the new watermark W’ to embed from the

rotation dimension giving the highest

intercorrelation coefficient.

The embedding and extracting algorithms are

given bellow:

(i) is the i

watermarked coefficient and 0≤i<N.

The variables:

Restore1, Restore2, Restore3, Restore4,

NoRestore, initialized to 0, are introduced to identify

the cases where the coefficients of C

are changed or

not. “abs” is the absolute value and “mod” is the

modulo operation.

if (k=4 and abs(C

(i))≠ 1) then

{

if (C

(i) > 0 ) then { if (C

(i) mod 2=1)

then {if (W’(i)=0) then { Cw

(i)= C

(i) – 1,

Restore1=1}

else {Cw

(i) = C

(i) ,

NoRestore=1}}

else {if (W’(i)=1) then {C w

(i)= C

(i) + 1,

Restore2=1}

else {Cw

(i)= C

(i) , NoRestore=1

}}

}

if (C

(i) < 0 ) then { If (abs(C(i)) mod 2=1)

then {if (W’(i)=0) then {Cw

(i) = C (i) + 1,

Restore3=1}

else {Cw

(i)= C

(i) , NoRestore=1

}}

else {if (W’(i)=1) then { Cw

(i) = C(i) – 1,

Restore4=1}

else {Cw

(i) = C

(i) , NoRestore=1

}}

}

Concerning the values of C

, they are

transformed to even or odd values depending on the

modifications that the coefficients of C

have

undergone. The coefficient C

(i) is then restored if

it has been changed unless C

(i) was initially equal

to 0. The corresponding algorithm is the following:

if (k=11) then

{

if (NoRestore = 1 and C

(i)> 0) then

{ if (C

(i)

mod

2 =0 ) then Cw

(i)

= C

(i),

else Cw

(i)

= C

(i) - 1 }

if (NoRestore = 1 and C

(i)< 0) then

{if (abs(C

(i)) mod 2 =0) then Cw

(i) = C

(i) ,

else Cw

(i)

= C

(i) + 1}

if (Restore1 = 1 or Restore2 = 1 or Restore3 = 1

or Restore4 = 1) then

{

if ((C

(i)) > 0) then

{ if (C

(i)

mod 2 =1 ) then Cw

(i) = C

(i),

else Cw

(i)= C

(i)- 1 }

if ((C

(i)) < 0) then

WATERMARKING OF COMPRESSED VIDEO BASED ON DCT COEFFICIENTS AND WATERMARK

PREPROCESSING

117

{ if (abs(C

(i)) mod 2 =1) then Cw

(i) = C

(i) ,

else Cw

(i)= C

(i)+ 1}

}

if (Restore1=1 or Restore4=1) then Cw

(i) = Cw

(i) +1

if (Restore2=1 or Restore3=1) then Cw

(i)= Cw

(i)-1

}

For the detection stage, the coefficients C

(i) and

(i) are first extracted. The even values of C

correspond to the cases where the watermark bits

didn’t introduce a modification in C

or they may be

a modification in case C

(i)=0, in this case the

watermark bit corresponds to the LSB of the

coefficient C

(i). For odd values of C

, the

watermark bit corresponds to the complement of the

LSB of the coefficient C

(i). The corresponding

algorithm is:

for i=0 to N do

{

if ( abs(Cw

(i)) mod 2 = 0)) then

{if (abs(Cw

(i)) mod 2 = 0) then W’(i)=0 else W’(i)=1 }

if ((abs(Cw

(i)) mod 2 = 1)) then

{if (abs(Cw

(i)) mod 2 = 1) then W’(i)=0 else W’(i)=1 }

}

W is then deduced from W’. The 12 bits

containing the additional information used for

reconstructing the watermark (see section 3.1 and

3.2) are embedded in 12 blocks reserved for that

aim, in low frequencies using the LSB method, and

extracted using the expression:

if (abs(C

(i)) mod 2 = 0) then W(i)=0 else W(i)=1.

4 RESULTS AND ANALYSIS

In this section we present the experimental results

obtained by applying our watermarking proposal to

various CIF video sequences including Walk,

Bridge, Flower, Coast guard, Foreman (with

Siemens logo) and Container. Using the base-line

profile, the first 50 frames of the video clips are

coded at the quantizer parameter (Qp=28), at a frame

rate of 30 pictures/s. The proposed watermarking

algorithm has been integrated into the H.264 JM

reference software.

The watermark used for tests is a pseudo random

binary sequence which is embedded only in the first

frame (Intra coded frame) of the video clip. The

objective is to assess the impact of the watermark

embedding on the video quality and the bitrate.

Results corresponding to the embedding capacity,

the average of PSNR of 50 frames and the bitrate are

tabulated in table 1.

The embedding capacity depends on the picture

content and the choice of the coefficients used for

embedding the watermark (C

in our case). It varies

between 174 and 1484 bits/ I frame corresponding to

sequences of Foreman and Flower respectively. We

recall that in this embedding capacity there are about

12 bits reserved to carry information about the

watermark (such as the rotation dimension, the

embedding formula) and are necessary for its

recovery; this is why it is important to ensure that

the remaining embedding capacity is enough for the

application.

In table 1 is also reported the average of PSNR

of the 50 coded frame of the test sequences. The

reduction in PSNR varies between 0.73 and 3.38 dB

corresponding to Flower and foreman respectively.

However, for all the test sequences, the

subjective quality is preserved and maintained

throughout the frames as presented in an example of

the PSNR graph of Bridge sequence in figure 4.

Figure 5 shows three examples of unwatermarked I

frames and the corresponding watermarked frames.

Figure 4: PSNR of unwatermarked and watermarked

Bridge sequence.

Regarding the bitrate, it is maintained or reduced

in most of cases and the most important reduction in

bitrate is 0.91 kbit/s observed for Flower sequence.

This was expected because of the formula used

to transform the odd values of C

to even values.

Indeed, the reduction of the coefficients “C

(i)= 1”

to the value “0” is responsible of this reduction of

the bitrate. Fortunately, it was possible to maintain

the video quality because the number of the

significant coefficients of C

remains limited and

both of the coefficients of low and high frequencies

were involved in the embedding process.

The proposed scheme may be considered as a semi

fragile one because the watermark is being

embedded during compression it is not easy to erase

IMAGAPP 2011 - International Conference on Imaging Theory and Applications

118

it from the video stream and preserve the file

content. Indeed any attack of the stream would cause

an error in the decoding stage. However, if the

watermarked video stream is decoded and re-

encoded the watermark is lost which

wouldconstitute a proof of the file manipulation

while checking the authenticity.

Table 1 : The PSNR, the bitrate and the embedding

capacity of the test sequences.

Sequences / Embedding

Capacity (Bits /I frame)

a. Unwatermarked seq.

b. watermarked seq.

Average

PSNR

(dB)

of 50 frames

Bitrate

(kbits/s)

Walk / 295

a 39.05 1111.15

b 36.41 1111.15

Bridge / 301

a 37.04 515.65

b 35.90 515.44

Flower / 1484

a 36.63 1457.63

b 34.78 1456.67

Coast guard /

315

a 38.33 1339.66

b 37.54 1339.65

Foreman / 174

a 39.65 365.02

b 36.27 364.84

Container / 457

a 39.02 186.43

b 36.91 186.30

5 CONCLUSIONS

This paper describes a video watermarking method

applied to the video codec H.264/AVC. The

objective was to embed data in the video with the

necessary amount of information for the video

authentication. The watermarking process should

preserve the video quality and reduce the bitrate

which would constitute a boon for network

applications.

The proposed method combines between the use

of low and high frequencies to achieve a good

compromise between the embedding capacity, the

video quality and the bitrate. A watermark

preprocessing helps considerably in improving the

video quality.

Promising results have been obtained, however,

while preserving this bitrate and quality, it would be

interesting to increase the embedding capacity

exploiting additional DCT coefficients and involving

the P and B frames in the watermarking process. It is

in this sense that our future work will focus.

Figure 5: (a), (b) and (c) are the I frames of the

unwatermarked sequences of Foreman, Container and

Flower respectively and (a’), (b’) and (c’) are the

corresponding watermarked frames respectively.

REFERENCES

Chen, W-M., Lai, C-J., Chang, C-C., 2009, H.264 video

watermarking with secret image sharing. In BMSB

'09, IEEE International Symposium on Broadband

Multimedia Systems and Broadcasting Bilbao, Spain,

pp. 1-7 .

ITU-T, 2005, Recommendation H.264: Advanced video

coding for generic audiovisual services.

Quan, L., Hong, L., 2008, Robust video watermarking

algorithm based on H.264. In WICOM’08, IEEE 4

International Conference on Wireless

communications, Networking and Mobile Computing,

Dalian, China, pp. 1-3.

Richardson, I. E. G., 2003, H.264 and Mpeg-4 video

compression: video coding for next-generation

multimedia, Wiley.

Wang, C-C., Hsu, Y-C. , 2008, New watermarking

algorithm with data authentication and reduction for

JPEG image. J.Electron. Imaging, vol. 17, pp.1-8.

Zhang, J., Ho, A. T. S., Qiu, G., Marziliano,P., 2007,

Robust video watermarking of H.264/AVC. In IEEE

transactions on circuits and systems II: express briefs,

vol. 54, no. 2, pp. 205-209.

(a)

(b)

(a’)

(b’)

(c)

(c’)

WATERMARKING OF COMPRESSED VIDEO BASED ON DCT COEFFICIENTS AND WATERMARK

PREPROCESSING

119