1
Color Image Watermarking
Using Multidimensional
Fourier Transforms
Received : December 7, 2006
Revised : October 21, 2007
Adviser ： Chih-Hung Lin
Speaker ：Ming-Rui Chen
Date： 2009/10/27
2016/7/15
Authors:
Tsz Kin Tsui, Student Member, IEEE
Xiao-Ping Zhang, Senior Member, IEEE
Dimitrios Androutsos, Senior Member, IEEE
2
2016/7/15
Outline
1. Author
2. Introduction
3. Background and Motivation
4. Multidimensional FOURIER Transforms
5. Simulation results
6. Conclusions
3
2016/7/15
1.Author
School :Ryerson University in Canada
• Tsz-Kin Tsui
▫ The M.A.Sc. degree in electrical and computer engineering, in 2006.
▫ Currently, he is with Research In Motion (RMI), Waterloo.
▫ interests are in image watermarking, multimedia data hiding, and image
segmentation.
• Xiao-Ping Zhang
▫
The B.S. and Ph.D. degrees in electronic engineering from Tsinghua University, Beijing, China,
in 1992 and 1996 .
▫
An Associate Professor Director of CASPAL.
Communications And Signal Processing Applications Laboratory
▫
Received the Science and Technology Progress Award.
▫ DSP and SAM Engineer .
• Dimitrios Androutsos
▫ Associate Professor.
▫ DSP Engineer and SOMA Networks.
4
2016/7/15
2.Introduction
There are many studies showing that the music and
video industry loses billions of dollars per year due to
illegal copying and downloading of copyrighted
materials from the Internet.
Pirate version of software
Produce
Digital Watermarking
illegal copy
5
2016/7/15
2.Introduction
[ Insertion ]
[ Extraction ]
6
2016/7/15
2.Introduction
• Spatial-Domain Techniques
▫ Random Seed
• Transform-Domain Techniques
▫ Spatial-Domain to Frequency-domain
▫ Change Frequency-domain’s coefficients
7
2016/7/15
2.Introduction-Spatial-Domain
• 1)Least-Significant Bit (LSB) 1990
▫ The earliest work of digital image watermarking schemes.
▫ Easy to implement and dos not generate serious distortion.
▫ It is not very robust against attacks .
• 2) Patchwork
▫ Does this image contain a watermark ?
▫ Using statistics method .
• 3)SSM-modulation-based techniques
▫ Spread-spectrum techniques spread or distributed in time or frequency domains .
▫ embed information Using linearly combining.
▫ modulated small pseudo noise signal.
8
2016/7/15
2.Introduction-Transform-Domain
Techniques
The aim is to Embed the watermarks in the spectral
coefficients.
HVS are better captured by the spectral coefficients.
Frequency-domain methods are more widely applied.
•
•
•
•
•
discrete
discrete
discrete
discrete
discrete
cosine
Fourier
wavelet
Laguerre
Hadamard
transform
transform
transform
transform
transform
(DCT )
(DFT)
(DWT)
(DLT)
(DHT)
9
2016/7/15
2.Introduction-Transform-Domain
Techniques
• DCT
▫ It has been shown that the watermarks are very hard to detect because
they consist of relatively weak noise signals.
• DFT
▫ Fourier coefficients have two components—phase and magnitude.
• DWT
▫ It searches the significant wavelet coefficients to embed the watermarks
in order to increase robustness.
10
2016/7/15
2.Introduction-Transform-Domain
Techniques
• DLT
▫ using the DLT and DFT together.
• DHT
▫ The advantage is that it offers much shorter processing time and easier
hardware implementation. Thus, it is optimal for realtime applications.
11
2016/7/15
3.BACKGROUND AND MOTIVATION
1.
Primary Challenges =>Tradeoff between robustness and perceptivity.
2.
Chromatic information to find appropriate transforms to bring the
information to the frequency domain.
3.
Embed watermarks as vectors into the chromatic as well as the
luminance channels.
This paper uses the theory of spatiochromatic image processing((SCIP)).
proposed by McCabe in 2000
12
2016/7/15
3.BACKGROUND AND MOTIVATION
Spatiochromatic image processing (SCIP) encodes only the color
information of an image to complex numbers a + j b
SCDFT (spatiochromatic discrete Fourier transform)
•
•
•
•
M * N is the size of the image
P and Q are the frequency coordinates
A and B are the real and imaginary values occurring at the spatial frequency
p and q are the image spatial coordinates
13
2016/7/15
3.BACKGROUND AND MOTIVATION
• One color pattern can have different oscillation frequencies
Conditions
HVS is not sensitive to these colors
Two different Conditions demonstrates
Yellow–blue
Color is, undoubtedly, an important feature in color images.
Red–green
14
2016/7/15
3.BACKGROUND AND MOTIVATION
• Colors can be added or subtracted to produce other
colors
▫ JPEG (easily destroy)
▫ Embedded in the highly sensitive area (strength must be low)
How can a better tradeoff be achieved?
Using the on SCIP the effect is the same as adding a rainbow grating to the image
This paper proposed algorithm adds one more step:
to add a compensation pattern in the negative frequency F(-u,-v)
15
2016/7/15
3.BACKGROUND AND MOTIVATION
Human Visual System
To satisfy the invisibility requirement in this paper
• the HVS is particularly sensitive to changes in image hue;
• the eye is less sensitive to the yellow–blue component [7];
• the JND model proposed by Chou in 2003 [30].
16
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
SCDFT Image Watermarking
(spatiochromatic discrete Fourier transform,SCDFT)
8
SCDFT scheme
8
W(k2,k2)
α is strength to be maximized
W(-k2,-k2)
17
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
SCDFT Image Watermarking
HVS characteristics
1)Keep the Hue (Angle) Unchanged:
Goal
Lat p=1 can get
r=real
i=imaginary
p=positive
n=negative
18
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
SCDFT Image Watermarking
2)Make the Overall Effect be Yellow–Blue
Lat
cancels out the real part of the watermarks,
only the imaginary part remains.
19
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
SCDFT Image Watermarking
To ensure invisibility of the watermarks
Defined distortion metric
the difference in the spatial domain generated by modifying the coefficients
Defined
20
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
SCDFT Image Watermarking
Depending on the perceptual requirement of the applications,
two methods are available to determine
Approach 1:
Make the average distortions of a block less NUJNCD (loose condition).
Approach 2:
Make the distortion of every pixel less NUJNCD (tight condition).
PS:nonuniform justnoticeable color difference(NUJNCD)
Approach 1 should be used when the robustness of the watermarks is
critical while the invisibility constraint can be relaxed.
Approach 2 is good for cases where the invisibility constraint is more
important.
21
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
SCDFT Image Watermarking
To extract the embedded watermarks
followed by a distance comparison
to the coefficient extracted from the
watermarked image
If coefficient =>
bit 1 was embedded
else
bit 0 was embedded
JPEG compression and color to grayscale conversion, destroy the watermark.
22
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
QFT Image Watermarking (quaternion Fourier transform)
1)Quaternion Image Processing (QOP)
Exploit:
Further information inherited among the channels of color images
Objective:
generalize different techniques from signal and image processing
ps: handle color images as vector images
SCIP => [ two components]
QIP => [four components]
most color images have three components (RGB, Lab, etc.)
1.spectral points must be known
2. McCabe SCDFT coefficients interpretation is not available
23
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
QFT Image Watermarking
2) New Adaptive Color Watermarking Algorithm
the goal is to design a watermark W(u,v) with the scaling factor
α as big as possible, such that the visible distortion to the image
is perceptually minimal.
Lat
b3 => luminance
c3 and d3 => chromatic
24
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
QFT Image Watermarking
2) New Adaptive Color Watermarking Algorithm
If b3=c3 the overall effect would vary along the 45 axis
Since the sensitivity of the HVS to the yellow–blue component
is approximately 1/5 compared to the luminance component
d3=(1/5)b3
25
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
QFT Image Watermarking
To ensure invisibility of the watermarks
Defined distortion metric
the difference in the spatial domain generated by modifying the coefficients
Defined
26
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
QFT Image Watermarking
watermark invisibility is achieved
if the distortion of the watermarked image is smaller than the
quantity NUJNCD
Approach 1:
Make the average distortions of a block less NUJNCD (loose condition).
Ans: Calculate α such that the average distortion of a block is less than NUJNCD
Approach 2:
Make the distortion of every pixel less NUJNCD (tight condition).
Ans: Recursively decrease α until the distortions of all pixels are smaller than
NUJNCD
27
2016/7/15
4.
MULTIDIMENSIONAL FOURIER TRANSFORMS
QFT Image Watermarking
To extract the embedded watermarks
F(u’,v’)+ α w(u’,v’)
F(u’,v’)- α w(u’,v’)
followed by a distance comparison
to the coefficient extracted from the
watermarked image
If coefficient =>
F(u’,v’)+ α w(u’,v’)
bit 1 was embedded
else
bit 0 was embedded
28
2016/7/15
original
5.SIMULATION RESULTS
• A series of experiments to Test!
• This algorithms increase the robustness against attacks ?
Test Image Data : Lenna image of size 512 X 480
The attacks include
•
•
•
•
•
Image compression—JPEG.
Geometric transformations—resizing, rotation, shearing, cropping.
Image enhancement —sharpening, and histogram equalization.
Color manipulation—color to grayscale conversion.
Information reduction—Gaussian noise.
29
2016/7/15
5.SIMULATION RESULTS
BERs =>bit-error rates
30
2016/7/15
5.SIMULATION RESULTS
31
2016/7/15
5.SIMULATION RESULTS
32
2016/7/15
6.CONCLUSION
• Color is a very important content in images
• Most existing watermarking schemes only utilize the
luminance content for watermarking
• The algorithms use the adaptive approach to obtain the
optimal strength of the watermark based on the local
characteristics of different images
33
2016/7/15
•Thank you