01377363
01377363
1, JANUARY 2005
Transactions Letters________________________________________________________________
Image Adaptive Watermarking Using Wavelet Domain Singular Value
Decomposition
Paul Bao and Xiaohu Ma
Abstract—In this letter, we propose a novel, yet simple, In 1997, Xia et al. [2] proposed a multiresolution water-
image-adaptive watermarking scheme for image authentication marking method by inserting pseudorandom codes to the
by applying a simple quantization-index-modulation process large coefficients at the high- and middle-frequency bands
on wavelet domain singular value decomposition. Unlike the
traditional wavelet-based watermarking schemes where the wa- of the dicrete wavelet transform (DWT) of an image. Their
termark bits are embedded directly on the wavelet coefficients, watermarking method is robust to some common image com-
the proposed scheme is based on bit embedding on the singular pressions and halftoning but the detection of the watermark
value (luminance) of the blocks within wavelet subband of the is dependent on the noise level in an image. Inoue et al.
original image. To improve the fidelity and the perceptual quality
of the watermarked image and to enhance the security of water-
[3] proposed a watermarking scheme by classifying wavelet
marking, we model the adaptive quantization parameters based coefficients as insignificant or significant using zerotree and
on the statistics of blocks within subbands. The scheme is robust then embedding a watermark in the location of insignificant
against JPEG compression but extremely sensitive to malicious coefficients or in the location of the thresholded significant
manipulation such as filtering and random noising. Watermark coefficients at the coarser scales. Xie et al. [4] proposed a
detection is efficient and blind in the sense only the quantization
parameters but not the original image are required. The quantiza- scheme combining watermarking with wavelet compression by
tion parameters adaptive to blocks are vector quantized to reduce engraving a watermark in the wavelet coefficients and encoding
the watermarking overhead. the watermarked coefficients using SPIHT compression algo-
Index Terms—Fragile watermarking, singular value decomposi- rithm. Their method is nonadaptive and may alter the original
tion (SVD), statistic modeling, wavelet. frequency correlations of the images. Tsai et al. [5] proposed
a watermarking scheme which utilizes the wavelet domain
I. INTRODUCTION image frequency components and the chaotic transformation to
select the location during the watermark embedding. Vehel et
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 12,2020 at 16:30:10 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 15, NO. 1, JANUARY 2005 97
may be able to preserve the visual perceptions of the original quantization, it is extremely robust to the JPEG compressions. On
images, it is vulnerable to compressions since the watermarked the other hand, the watermarks so embedded are very fragile to
coefficients may be subject to thresholding and quantization in any malicious manipulations such as filtering or random noising
the compression process, thus damaging the watermark bits. since any manipulation attempt would most probably damage
Aiming at the indirect watermarking embedding, Gorodetski et watermark bits in all the subbands, thus destroy the authenticity
al. [11] proposed a simple singular value decomposition (SVD) of the watermark. The watermark detection in our scheme is
domain watermarking scheme by embedding the watermark blind, i.e., only the quantization parameters but not the original
to the singular values (SVs) of the images, to achieve a better watermark image are needed. Furthermore, since the watermark
transparency and robustness. But the method is not image adap- embedding is based on the quantization of the SV of blocks where
tive and unable to offer a consistent perceptual transparency of the quantization parameters are modeled by the block statistics in
watermarking for different images. Liu et al. [12] presented a the wavelet domain, it is, thus, impossible to maliciously detect
robust SVD domain watermarking scheme where a watermark the watermark without the quantization parameters.
(in the form of a matrix) is added to the singular value matrix of
the watermarking image in spatial domain. Liu’s scheme offers II. WATERMARKING BASED ON SVD
good security against tampering and common manipulations for
A grayscale digital image is specified by an matrix
protecting rightful ownership. But since the scheme is designed
. If a color image is represented in RGB then it
for the rightful ownership protection, where the robustness
can be converted to the corresponding luminance matrix
against manipulations is desired, it is suitable for authentication.
. An arbitrary matrix can be repre-
The scheme is also nonadaptive thus unable to offer a consistent
sented by its SVD in the following form:
perceptual transparency of watermarking for different images.
Chen et al. [13], [14] introduced a framework for charactering (1)
the inherent tradeoffs between the robustness of the embedding,
the degradation to the host image, and the amount of data em- where and are orthogonal and matrices,
bedded and designed a framework of information embedding respectively, and is a diagonal matrix with nonnegative ele-
systems, namely quantization index modulation (QIM), aiming at ments. Diagonal terms of matrix are SVs of
optimizing the rate-distortion-robustness tradeoffs. They devel- matrix and is the rank of matrix . SVD possesses sev-
oped a method, the dither modulation, to realize and demonstrate eral attractive mathematical properties, one of which is that each
the QIM framework, where the embedded information would SV specifies the luminance of the SVD image layer, whereas
modulate the dither signal of a dithered quantizer. The QIM the respective pair of singular vectors specifies intrinsic geom-
schemes can achieve an optimal tradeoff between the embedding etry properties of images. It was discovered that slight varia-
rate, the distortion and the robustness by freely adjusting the tions of SVs do not affect the visual perception of the cover
number of quantizers, the quantization steps and the minimum image, which motivates the watermarking embedding through
distance, respectively, for different requirements such as robust- slight modifications of SVs in the segmented images.
ness to a known manipulation. But the QIM schemes manipulate The proposed scheme is briefly described as follows. An
the quantization-index modulation directly in the original signal image represented in matrix format is segmented into blocks
domain and fail to take the advantages offered by wavelet-based of size (in our experiment, is generally set to 4) and
or other indirect watermarking schemes in achieving a better the SVD for each of the blocks is performed. Then, one bit of
performance in all the characteristics. data is embedded through a slight modification of the SV of
Synthetically, it seems that none of the schemes could si- the block. Let be the current bit of the watermark image to
multaneously offer the characteristics of adaptive transparency, be embedded into this block . The embedding algorithm is
robustness to compressions, sensitivity to malicious manipula- described as follows.
tion and blind detection of watermarks. To address this issue, we 1) Segment the image into blocks of size ,
design a novel image-adaptive watermarking scheme for image , where is the number of the blocks.
authentication by applying a quantization-index-modulation 2) Compute , where ,
process on the SVs of the images in wavelet-domain in this is a vector formed by the SVs of each block .
letter. In this scheme, the watermark bits are embedded on the 3) Compute integer number , where is the
SVs in the SVD layers for each of the wavelet domain blocks, quantization step for corresponding to the block .
with the quantization steps adaptive to the statistical model 4) Embed one bit of watermark image as follows.
of the block, which ensures that the overall luminance change If , then
be minimized and perceptually unnoticeable. The block size, If is odd number, then
together with the weighting parameters and the minimum and ELSE remains unchanged
maximum quantization steps, would determine how well a block If , then
could be modeled statistically leading to a unified framework for If is even number, then
controlling the transparency, robustness against compressions, Else remains unchanged.
fragility to malicious manipulations, and embedding rate. Since 5) Compute the value and the modified
the watermark is distributed at the luminance (the SVs) in all the SV
subbands where the relative luminance (thus,the watermark) in
wavelet domain would be well preserved during the coefficient
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 12,2020 at 16:30:10 UTC from IEEE Xplore. Restrictions apply.
98 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 15, NO. 1, JANUARY 2005
TABLE I
CONSTANT QUANTIZATION STEP d: THE QUALITY OF WATERMARKED IMAGE (PSNR) AND BER OF THE EXTRACTED WATERMARKS CORRESPONDING TO
SEVERAL JPEG COMPRESSIONS, RESPECTIVELY
6) Compute the matrix of the block using the modified SV quantization scheme, a watermarking scheme with high trans-
parency and robustness to JPEG compressions can be obtained.
The method can be described as follows.
1) Calculate the standard deviation for each block .
2) Calculate the maximum value and minimum value
7) Reconstruct the watermarked image from all the blocks for all the .
. 3) Compute the quantization step for the block as
It should be noticed that quantization step would be served
as a secret key to ensure the authorized access to the watermark. (2)
The extraction of the watermark is straightforward. Let be a
block with an embedded watermark bit. where is the number of all blocks, and are
1) Segment watermarked image into blocks of size minimum and maximum quantization step values, respec-
, , where is the number of the blocks. tively, specifiable by user.
2) Compute the value , , Table II gives some comparison results with the basic algorithm
where is a vector formed by the SV of in Section III and Fig. 2 gives the visual examples of the water-
each block . marked images, where 9 and 36 or 45.
3) Compute . Note from Table II that the BERs increase rapidly as the JPEG
4) If is even number, then the embedded bit is 1. Other- compression ratios. This is because that the watermarking distri-
wise, it is 0. bution is only adaptive to the standard deviation of each block in
the spatial domain. Thus, the SVs of a more spatially structured
block (larger deviation) would be significantly modified which
III. MODELING OF THE QUANTIZATION PARAMETERS may be insignificant in wavelet domain and thus be thresholded.
A. Quantization Parameters Based on Spatial Domain On the other hand, the SVs of a less spatially structural block
Modeling (small deviation) would be slightly modified, which, however,
could be significant in wavelet domain allowing the embedding
For a watermarking scheme, it is desired that the water-
of more watermark information. Thus, the aforementioned wa-
marked images achieve a high PSNR while the watermark can
termarking distribution scheme would result in a less satisfying
be detected and extracted robustly against any compression
transparency and a poor robustness to compressions.
schemes, i.e., possessing a lower bit-error ratio (BER) against
compressions. In the aforementioned SVD-based watermarking
B. Adaptive Quantization Parameters Based on Wavelet
scheme, it is obvious that the increase of the value leads
Domain Modeling
to more robust watermark embedding (lower BER) but poorer
transparency of the watermarked image (lower PSNR) and vice In view of this potential drawback, we propose a set of quanti-
versa. Table I gives the PSNRs of the watermarked images using zation parameters modeled adaptively to both the deviation and
several constant quantization steps and their corresponding the expectation of each block within the wavelet domain where
BERs against JPEG compressions. Fig. 1 shows the visual the SVs for the real structures will be largely modified with wa-
perceptions of the watermarked images for various quantization termarking whereas that of the nonstructural background will be
steps, respectively. slightly modified to ensure a high perceptual quality of water-
From Table I, it can be observed that the perceptual quality of marked image and a low BER of the detected watermark. The
watermarked image is reversely proportional to the robustness scheme is detailed as follows.
against JPEG compression. In order to preserve the high percep- 1) An image is transformed into wavelet subbands. In each of
tual quality of the watermarked image while retaining a good the subbands, the coefficients are segmented into blocks
robustness to JPEG compression, we propose an image-adap- of size and SVDs for each of the blocks are com-
tive quantization for image watermarking. With this adaptive puted.
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 12,2020 at 16:30:10 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 15, NO. 1, JANUARY 2005 99
Fig. 1. Visual appearances of the watermarked images obtained by constant quantization steps. (a) Watermark image. (b)–(c) Two original images. (b1)–(b3)
Watermarked images corresponding to the original image (b) when quantization steps d are set to 9,27, and 54, respectively. (c1)–(c3) Watermarked images
corresponding to (c) when d = 9, 27, and 54, respectively.
TABLE II
IMAGE ADAPTIVE QUANTIZATION STEP d : THE QUALITY OF WATERMARKED IMAGE (PSNR) AND BER OF THE EXTRACTED WATERMARK CORRESPONDING TO
SEVERAL JPEG COMPRESSION RATIOS, RESPECTIVELY
2) Calculate the standard deviation and average value where is the number of the blocks, and and are the
for DWT coefficients of each block . minimum and maximum quantization step values, respectively,
3) Calculate the value for each block specifiable by user.
(3) We model the quantization parameters based on both the ex-
pectation and deviation of the block to capture both the struc-
where and are the weight parameters for and tural and background statistics in quantizing the modification
. of the SV. In our experiment, we set and
4) Calculate the maximum value and minimum value or , and let , . The value is formu-
from all the . lated so that both the baseband and the high-pass bands can be
5) Compute the quantization step for block as follows: watermarked and the modifications of the SVs can be adjusted
(4) using parameters and . Parameter controls the quanti-
zation weighting for the background while for the structures.
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 12,2020 at 16:30:10 UTC from IEEE Xplore. Restrictions apply.
100 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 15, NO. 1, JANUARY 2005
TABLE III
THE QUALITY OF WATERMARKED IMAGE (PSNR) AND BER OF THE EXTRACTED WATERMARK CORRESPONDING TO 40% JPEG COMPRESSION FOR DIFFERENT
VALUES OF PARAMETER c AND c , RESPECTIVELY
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 12,2020 at 16:30:10 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 15, NO. 1, JANUARY 2005 101
TABLE IV
2
COMPARISONS BETWEEN THE PROPOSED METHOD AND WAVELET BASED WATERMARKING SCHEMES ON 512 512 IMAGES: BER OF THE EXTRACTED
WATERMARK CORRESPONDING TO SEVERAL JPEG COMPRESSION RATIOS, RESPECTIVELY
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 12,2020 at 16:30:10 UTC from IEEE Xplore. Restrictions apply.
102 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 15, NO. 1, JANUARY 2005
Fig. 4. (a1)–(a2) Modified watermarked images. (b1)–(b2) Detected watermark images corresponding to the LL band. (c1)–(c2) Enlarged detected watermark
images for detecting the malicious modification operation.
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 12,2020 at 16:30:10 UTC from IEEE Xplore. Restrictions apply.