here - UBC ECE - The University of British Columbia

A ROBUST DCT ENERGY BASED WATERMARKING SCHEME FOR IMAGES
Adarsh Golikeri, Panos Nasiopoulos
Department of Electrical and Computer Engineering
The University of British Columbia, Vancouver, BC, Canada
{adarshg, panos}@ece.ubc.ca
ABSTRACT
We propose a new, robust and simple DCT energy based
watermarking scheme for images. We first divide the image
into blocks and compute the amount of DCT energy in
each block. Blocks containing significant DCT energy,
which represent areas of significant detail, are chosen for
insertion of a Gaussian pseudorandom watermark. The
watermark is inserted in a spread-spectrum fashion in the
spectral domain, thereby making it robust against signal
processing operations. Our scheme is very easy to
implement and is shown to be robust against cropping,
scaling, compression distortion and multiple document
attacks. We show that our algorithm performs equally well
for JPEG compression and Median filtering, when
compared existing DCT-based methods, which are
computationally more comp lex.
1. INTRODUCTION
The need for watermarking of images has gained
importance in the past few years, owing to the rapid
growth of such digitized media over the internet. Images
can now easily be copied and distributed, with little or no
control of ownership. Traditional encryption systems exist,
which allow only valid key-holders to access data.
However, once decrypted, this data is again susceptible to
unauthorized reproduction. Therefore, digital watermarking
schemes are needed in order to serve as standalone or
complementary copyright protection systems.
A digital watermark is a secret code carrying
identification information about the copyright owner or
creator. Generally, a watermark is embedded such that it is
invisible. In order to be effective, a watermark should
satisfy certain basic criteria:
Unobtrusiveness: The watermark should not degrade or
affect the image quality in any perceptible manner.
Robustness: It should be resistant to attacks, both
intentional and unintentional, specifically the following
types of attacks:
i) Geometric distortions: Operations such as cropping,
scaling, rotation and translation.
ii) Collusion attacks: Attempts to destroy the
watermark by making use of multiple watermarked images.
Also, it should not be possible to generate a new
watermarked image by combining several watermarked
images.
iii) Other operations: This includes digital-to-analog
conversion, analog-to-digital-conversion, resampling,
requantization and compression distortion (e.g. JPEG).
Detection Accuracy: The detection algorithm should have
a low false-positive and false-negative rate. Detection of
the watermark should be able to prove ownership
unambiguously.
Two major applications of digital watermarking are
copyright protection (proving ownership of data) and data
authentication (for use as evidence against crimes). In
such cases the data needs to be proved reliable and
unmodified. Current techniques for watermarking
concentrate mainly on images and can be classified into
two groups. The first group is based on spatial domain
techniques, which embed the watermark by directly
modifying the pixel values in the image. The second group
comprises of transform domain methods, which embed the
watermark by modulating the transform domain
coefficients of the data. The transform methods are more
complex, but more robust than the spatial methods.
In this paper, we propose a new, simple and robust
DCT-based scheme for images. We show through
experiments that our scheme satisfies the basic criteria of a
watermark mentioned above. Our method is more robust
against compression distortion than the existing DCTbased methods [2]. The rest of the paper is organized as
follows: Section 2 provides an overview of existing DCTbased watermarking schemes. Section 3 details our
proposed scheme. We then discuss experimental results in
Section 4 and provide concluding remarks in Section 5.
2. EXISTING DCT-BASED WATERMARKING
Cox pioneered the frequency-domain watermarking scheme
based on DCT [1]. Cox’s proposed method
computes the N x N DCT coefficients for an N x N image.
The watermark of length n is placed into the n highest
magnitude coefficients of the transform matrix, excluding
the DC component. The motivation for choosing the
higher value coefficients is that they represent the low
frequency regions of an image, which contain most of the
perceptually significant image information. Also, the
human visual system attaches more resolution to the lowfrequency spectral components. Further, it has been
observed that common signal processing operations and
distortions affect the perceptually insignificant regions of
an image, which correspond to high-frequency
components. So, the watermark has to be inserted in the
low-frequency components.
A more recent DCT-based method, which uses feature
extraction points and the Voronoi diagram can be found in
[2]. The authors segment the given image based on the
Voronoi diagram and feature extraction points. The
complexity of this step is of the order of O(n log n)
computations. Then, they embed the watermark into the
DCT domain of each image segment. Our scheme has the
advantage of computational simplicity, when compared to
[2], without compromising on the level of robustness.
3. PROPOSED WATERMARKING SCHEME
Our algorithm makes use of perceptually significant DCT
coefficients of the image and uses them to carry the
watermark information. So, we utilize blocks with high DCT
energy. These blocks represent regions in the image with
lot of texture and detail. This serves two purposes. First,
these regions have a great perceptual capacity, in that
modifying their DCT coefficients will not cause visible
artifacts. Second, any attack on these regions (malicious or
otherwise), will render the image unusable. This serves as
a security measure against tampering.
Watermark insertion:
1) Read input image to be watermarked (Figure 1(a)).
2) Compute high DCT energy blocks (Figure 1(b), defined
as those having at least 30% sum of squared AC energy of
the highest energy block ), by first dividing the image into
16x16 blocks and computing DCT for each block. Let n be
the number of selected blocks.
3) Compute the watermark sequence W, from an N(0,1)
distribution. This type of sequence is particularly robust
against collusion attacks. W is an array of length n. We
denote the elements of W as wi (where i = 1 to n).
4) For i = 1 to n:
i) Select the perceptually significant DCT coefficients of
this block – we choose the first 8 AC coefficients of the
16x16 block, in JPEG zig-zag scan order.
(a)
(b)
(c)
Figure 1. (a) Original image (b) Selected high energy blocks
(c) Watermarked image
ii) Alter these coefficients according to :
Vi' = Vi ( 1 + a wi)
where, Vi' = adjusted coefficient, Vi = original coefficient,
a = scaling factor (determines watermark strength).
8) Generate the watermarked image by applying the inverse
DCT and merging all blocks (Figure 1(c)).
Watermark detection:
Given a possibly distorted version of the image, we
attempt to extract the watermark and compare it with the
original watermarked data.
1) Read the image to be tested.
2) Compute the high energy DCT blocks as in Step (2) of
the insertion scheme.
3) For i = 1 to n, extract the watermark estimate W* (which
is also an array of length n.
4) Apply the following transformations to W and W* for
robustness:
x = x - mean(x) (reduce x by its average value)
x = sign(x) (convert x into an array of -1, 0 or 1s)
5) Compare W and W* using the correlation coefficient as
the similarity measure.
6) Detection is successful if the correlation exceeds the
experimentally determined threshold of 0.2. Performance
evaluations have shown that this threshold value
minimizes the probability of false positives.
(a)
(b)
Figure 2. Uniqueness of watermark – detector response to 1000
random watermarks, only one of which matches the watermark
present in Figure 1. Dashed line indicates threshold value = 0.2
4. EXPERIMENTAL RESULTS
To test our proposed scheme, we used several standard
test images such as Cameraman, Baboon, Lena, etc. We
display results from the Cameraman image. Figure 1(a)
shows the original Cameraman image, while 1(b) shows the
high energy DCT block. For all our experiments, we use a
scaling factor a of 0.1. Figure 1(c) shows the watermarked
Cameraman image. Subjectively, it is indistinguishable from
the original image. We conducted the following tests:
Uniqueness of watermark: In this test, 1000 random
watermarks were generated, out of which only one
matched the original watermark. The output of the detector
is shown in Figure 2. The response to the correct
watermark is 0.9204, which is much stronger than the
threshold value of 0.2. This indicates that the proposed
algorithm has very low false positive response rates.
Geometric manipulations (scaling and cropping):
The watermarked image was resized to 50% of its original
size (Figure 3(a)) and then rescaled to its original
dimensions (Figure 3(b)). The loss of detail in the rescaled
image is clearly visible. The detector response to Figure
3(b) was 0.3065, which is well above the threshold. Next,
about 40% of the watermarked image was cropped (portion
shown in Figure 3(c)) and the remaining pixels were
replaced from the unwatermarked image. The detector
response to Figure 3(d) was 0.3692.
Compression distortion: The Cameraman image was
encoded as different JPEG images with parameters of 65%,
45%, 25% and 10% quality (Figure 4(a)-(d)). The detector
gave outputs of 0.3571, 0.2576, 0.2440 and 0.2356
respectively. This is in spite of severe JPEG compression
(c)
(d)
Figure 3. Image scaling – (a) 50% scaled version (b) Rescaled
version showing loss of detail (c) cropped portion from
watermarked image (d) reconstructed image using the remaining
portion of the unwatermarked image
artifacts in the JPEGs with lower quality factors.
Collusion attacks: Any good watermarking scheme needs
to be sufficiently resistant to collusion (multiple document)
attacks, wherein multiple watermarked images maybe be
combined to destroy the watermark. In this test, we first
generated 5 different watermarked images and then
generated a 6th image by averaging them. When the
resulting average image (Figure 5(a)) is compared with 1000
random watermarks, the detector gives good correlation
values of 0.2861, 0.3978, 0.3423, 0.3247 and 0.2832 for the 5
original watermarks and low values (mean = 0.0039) for the
remaining watermarks (Figure 5(b)). This shows that this
collusion attack has failed to destroy the uniqueness of
the 5 individual watermarks.
Comparison of proposed scheme and existing DCT-based
method: The Cameraman image was watermarked using our
scheme and the method by Suhail and Obaidat in [2], then
encoded as JPEGs of different compression ratios. As can
be seen in Figure 6, our algorithm performs equally well,
without introducing the computational complexity for
computing the Voronoi diagram [2]. Also, our scheme gave
an output of 0.2345, for a compression ratio as low as 10%.
We also compared performance after Median filtering,
using different kernel sizes, from 3x3 to 9x9. Both schemes
performed similarly, with our scheme giving a slightly
better result (0.5122 vs 0.47) for a filter size of 9x9.
(a)
(b)
(a)
(c)
(d)
Figure 4. JPEG distortion – with (a) 65% (b) 45% (c) 25% and
(d) 10% quality factors
5. CONCLUSION
We have presented a new, robust and simple DCT energy
based watermarking scheme, based on spread spectrum
approach. First, the image is divided into 16x16 blocks and
the DCT of each block is computed. Then, high energy
DCT blocks, which represent areas of significant detail, are
selected and a Gaussian pseudorandom watermark is
inserted into perceptually significant DCT coefficients. Our
scheme is shown to be resistant against cropping, scaling,
compression distortion and multiple document attacks. We
have shown that our algorithm performs equally well,
compared to computationally more expensive DCT-based
methods, while maintaining the same level of robustness.
(b)
Figure 5. Collusion attacks (a) Image generated by averaging 5
independently watermarked images (b) Detector output for 1000
random watermarks, 5 of which were used to create (a).
6. REFERENCES
[1] I. Cox, J. Kilian, F. T. Leighton and T. Shamoon, “Secure
spread spectrum watermarking for multimedia,” ( IEEE Trans.
Image Processing, vol.6, pp. 1673–1687, Dec. 1997.)
[2] M. A. Suhail and M. S. Obaidat, "Digital Watermarking-based
DCT and JPEG model" (IEEE Trans. on Instrumentation and
Measurement, Vol 52, No. 5, October 2003)
[3] I. Cox, M. Miller and J. Bloom, Digital Watermarking. New
York: Morgan Kaufmann, 2001.
[4]
F. Hartung and B. Girod, "Watermarking of
Uncompressed and Compressed Video", in Signal
Processing, pp. 283-301, 1998.
Figure 6. Comparison of proposed scheme and Suhail's method
for JPEG Compression
[5] P. Wolfgang and E. Delp, "A watermark for digital images",
in Proc. IEEE Int. Conf. Image Processing, Sept 1996, pp. 219222.
[6] A. Bors and I. Pitas, "Image watermarking using DCT
domain constraints", in Proc. ICIP, Sept. 1996, pp. 231-234.