U.S. patent application number 11/474143 was filed with the patent office on 2006-10-26 for automated digital watermarking methods using neural networks.
Invention is credited to Kayvan Najarian.
Application Number | 20060239504 11/474143 |
Document ID | / |
Family ID | 26832124 |
Filed Date | 2006-10-26 |
United States Patent
Application |
20060239504 |
Kind Code |
A1 |
Najarian; Kayvan |
October 26, 2006 |
Automated digital watermarking methods using neural networks
Abstract
Embodiments of the present invention provide digital
watermarking methods that embed a digital watermark in both the low
and high frequencies of an image or other production, providing a
digital watermark that is resistant to a variety of attacks. The
digital watermarking methods of the present invention optimize the
strength of the embedded digital watermark such that it is as
powerful as possible without being perceptible to the human eye.
The digital watermarking methods of the present invention do this
relatively quickly, in real-time, and in an automated fashion using
an intelligent system, such as a neural network. The digital
watermarking methods of the present invention may also be used in a
variety of new applications, such as the digital watermarking of
sensitive aircraft parts and military equipment.
Inventors: |
Najarian; Kayvan; (Concord,
NC) |
Correspondence
Address: |
KILPATRICK STOCKTON LLP - 46872;J. STEVEN GARDNER
1001 WEST FOURTH STREET
WINSTON-SALEM
NC
27101
US
|
Family ID: |
26832124 |
Appl. No.: |
11/474143 |
Filed: |
June 23, 2006 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
10134255 |
Apr 29, 2002 |
7095872 |
|
|
11474143 |
Jun 23, 2006 |
|
|
|
60315223 |
Aug 28, 2001 |
|
|
|
Current U.S.
Class: |
382/100 ;
375/E7.018; 380/54; 382/232; 713/176 |
Current CPC
Class: |
H04N 1/3217 20130101;
H04N 21/23892 20130101; G06T 2201/0052 20130101; H04N 1/3216
20130101; H04N 2201/327 20130101; H04N 1/32154 20130101; G06T
1/0071 20130101; H04N 2201/3233 20130101; H04N 1/32288 20130101;
H04N 1/32192 20130101; G06T 1/0028 20130101; H04N 1/32165
20130101 |
Class at
Publication: |
382/100 ;
382/232; 713/176; 380/054 |
International
Class: |
G09C 5/00 20060101
G09C005/00; G06K 9/00 20060101 G06K009/00; H04L 9/00 20060101
H04L009/00; G06K 9/36 20060101 G06K009/36 |
Claims
1. A digital watermarking method, comprising: calculating a
discrete transform comprising a plurality of frequency bands; and
inserting a plurality of digital watermarks into the plurality of
frequency bands; wherein each of the plurality of digital
watermarks has a predetermined weight.
2. The digital watermarking method of claim 1, wherein the discrete
transform comprises a discrete transform selected from the group
consisting of a discrete wavelet transform (DWT), a discrete cosine
transform (DCT), and a discrete Fourier transform (DFT).
3. The digital watermarking method of claim 1, wherein the
plurality of frequency bands comprise a plurality of
high-frequency, median-frequency, and low-frequency bands.
4. The digital watermarking method of claim 1, further comprising
determining the predetermined weight of each of the plurality of
digital watermarks using the following equation:
c.sub.i'=c.sub.i(1+.alpha.m.sub.i), where .alpha. is a scaling
parameter, c.sub.i is a coefficient of an original production,
m.sub.i is a digital watermark to be added, and c.sub.i' is a
digitally watermarked coefficient.
5. The digital watermarking method of claim 1, wherein the digital
watermarking method is used to digitally watermark one of an image,
a picture, a video, a multimedia product, a CD, a DVD, biometrics,
an MRI, an FMRI, a CT, an ultrasound image, a non-digital VLSI
chip, a non-digital vital part, and a non-digital negative
film.
6. A computer-readable medium having executable commands operable
for digitally watermarking a production, the executable commands
comprising: calculating a discrete transform comprising a plurality
of frequency bands; and inserting a plurality of digital watermarks
into the plurality of frequency bands; wherein each of the
plurality of digital watermarks has a predetermined weight.
7. The computer-readable medium having executable commands of claim
6, wherein the discrete transform comprises a discrete transform
selected from the group consisting of a discrete wavelet transform
(DWT), a discrete cosine transform (DCT), and a discrete Fourier
transform (DFT).
8. The computer-readable medium having executable commands of claim
6, wherein the plurality of frequency bands comprise a plurality of
high-frequency, median-frequency, and low-frequency bands.
9. The computer-readable medium having executable commands of claim
6, further comprising determining the predetermined weight of each
of the plurality of digital watermarks using the following
equation: c.sub.i'=c.sub.i(1+.alpha.m.sub.i), where .alpha. is a
scaling parameter, c.sub.i is a coefficient of an original
production, m.sub.i is a digital watermark to be added, and
c.sub.i' is a digitally watermarked coefficient.
10. The computer-readable medium having executable commands of
claim 6, wherein the computer-readable medium having executable
commands is used to digitally watermark one of an image, a picture,
a video, a multimedia product, a CD, a DVD, biometrics, an MRI, an
FMRI, a CT, an ultrasound image, a non-digital VLSI chip, a
non-digital vital part, and a non-digital negative film.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This non-provisional patent application claims the benefit
of U.S. Provisional Patent Application No. 60/315,223, entitled
"Maximum-Strength Multi-Resolution Watermarking of Digital Products
Using Neural Networks," filed Aug. 28, 2001 and U.S.
Non-Provisional patent application Ser. No. 10/134,255, entitled
"Automated Digital Watermarking Methods Using Neural Networks,"
filed Apr. 29, 2002.
FIELD OF THE INVENTION
[0002] The present invention relates generally to the field of
digital information security. Specifically, the present invention
relates to copyright protection for digital products, or digital
watermarking. More specifically, the present invention relates to
automated digital watermarking methods using neural networks. The
methods of the present invention may also be applied to non-digital
applications.
BACKGROUND OF THE INVENTION
[0003] The explosive growth of globally-distributed computer
networks, such as the Internet and broadband networks, the presence
of relatively large storage devices and data archives, and the
development of efficient compression standards have provided
digital image producers and others with a variety of means for
distributing their images and other productions. However, these
advancements have also made it relatively simple for copyright
violators and others to create illegal copies of these images and
other productions. As a result, a variety of conventional digital
watermarking methods have been proposed. A digital watermark
encodes an owner's copyright and/or tracking information and embeds
it into an image or other production. Preferably, this digital
watermark is invisible to the human eye and may only be perceived
by a computer or the like. Digital watermarking methods may be used
to identify the owner of an image or other production, to track
illegal copies of the image or other production, and to facilitate
the licensing of the image or other production. For a digital
watermarking method to be successful, the digital watermark must be
unobtrusive and must not degrade the perceptual quality of the
image or other production. The digital watermark must also be
resistant to attacks from a variety of image and signal processing
tools and techniques, both unintentional and intentional. Such
attacks may include, for example, image compression, smoothing,
low-pass filtering, cropping, geometric transformation (including
translation, rotation, and scaling), noise addition, printing and
scanning, and collusion.
[0004] In general, a digital watermark may be applied in either the
spatial (pixel) domain or the transform domain. Transform domain
digital watermarking methods, such as discrete cosine transform
(DCT) and discrete wavelet transform (DWT) digital watermarking
methods, typically provide relatively high image fidelity and are
resistant to image manipulations. Wavelet-based digital
watermarking techniques have multi-resolution hierarchical
characteristics that mimic human visual (and audio) perception and
allow for the independent processing of the resulting components.
With wavelet-based digital watermarking techniques, digital
watermark detection may be achieved at relatively low image
resolutions, saving computational load. In addition, the
high-frequency sub-bands of a wavelet transform include the edges
and textures of an image and the human eye is typically insensitive
to changes in such sub-bands. This allows digital watermark to be
added to the sub-bands without being perceived by the human eye.
Wavelets are used in a variety of emerging image and video
compression standards, such as JPEG2000 and MPEG4, encouraging the
use of wavelet-based digital watermarking. Although embodiments and
examples of the present invention deal with transform-based
imperceptible digital watermarking methods, the principles and
techniques of the present invention are also applicable to spatial
(time)-based imperceptible digital watermarking methods as
well.
[0005] A variety of conventional wavelet-based digital watermarking
methods have been proposed. All of these methods are designed to be
resistant to a variety of attacks. Typically, these digital
watermarking methods insert a digital watermark into the
high-frequency contents of an image or other product. Although
marginally effective with respect to some types of attacks, the
conventional digital watermarking methods are vulnerable to attack
if techniques such as low-pass filtering are used. In addition, the
conventional digital watermarking methods tend to be
labor-intensive, relatively slow, and lack the ability to be
adequately automated.
[0006] Thus, what is needed is a digital watermarking method that
embeds a digital watermark in both the low and high frequencies of
an image or other production, providing a digital watermark that is
resistant to a variety of attacks. This digital watermarking method
should optimize the strength of the embedded digital watermark such
that it is as powerful as possible without being perceptible to the
human eye. The digital watermarking method should do this
relatively quickly, in real-time, and in an automated fashion using
an intelligent system, such as a neural network. The watermarking
method should also be able to be used in a variety of new
applications, such as the watermarking of sensitive aircraft parts,
military equipment, and machines.
BRIEF SUMMARY OF THE INVENTION
[0007] Embodiments of the present invention provide digital
watermarking methods that embed a digital watermark in both the low
and high frequencies of an image or other production, providing a
digital watermark that is resistant to a variety of attacks. The
digital watermarking methods of the present invention optimize the
strength of the embedded digital watermark such that it is as
powerful as possible without being perceptible to the human eye.
The digital watermarking methods of the present invention do this
relatively quickly, in real-time, and in an automated fashion using
an intelligent system, such as a neural network. The watermarking
methods of the present invention may also be used in a variety of
new applications, such as the watermarking of sensitive aircraft
parts, military equipment, and machines such that the destruction
of a watermark associated with a given machine compromises the
function of the machine.
[0008] In one embodiment of the present invention, a digital
watermarking method includes calculating a discrete transform
including a plurality of frequency bands and inserting a plurality
of digital watermarks into the plurality of frequency bands,
wherein each of the plurality of digital watermarks has a
predetermined weight.
[0009] In another embodiment of the present invention, a
computer-readable medium having executable commands operable for
digitally watermarking a production includes executable commands
operable for calculating a discrete transform including a plurality
of frequency bands and inserting a plurality of digital watermarks
into the plurality of frequency bands, wherein each of the
plurality of digital watermarks has a predetermined weight.
[0010] In a further embodiment of the present invention, an
automated watermarking method includes creating a database
including a first plurality of productions and receiving a first
plurality of scores associated with the quality of the first
plurality of productions from a plurality of human subjects. The
automated watermarking method also includes providing information
associated with the first plurality of productions and the first
plurality of scores to an intelligent system and training the
intelligent system to generate a second plurality of scores
associated with the quality of the first plurality of productions,
wherein the second plurality of scores is substantially the same as
the first plurality of scores. The automated watermarking method
further includes providing information associated with a second
plurality of productions to the intelligent system, receiving a
third plurality of scores associated with the quality of the second
plurality of productions from the intelligent system, and
optimizing the strength of a watermark using the third plurality of
scores.
[0011] In a further embodiment of the present invention, a
computer-readable medium having executable commands operable for
digitally watermarking a production in an automated fashion
includes executable commands operable for creating a database
including a first plurality of productions and receiving a first
plurality of scores associated with the quality of the first
plurality of productions from a plurality of human subjects. The
executable commands are also operable for providing information
associated with the first plurality of productions and the first
plurality of scores to an intelligent system and training the
intelligent system to generate a second plurality of scores
associated with the quality of the first plurality of productions,
wherein the second plurality of scores is substantially the same as
the first plurality of scores. The executable commands are further
operable for providing information associated with a second
plurality of productions to the intelligent system, receiving a
third plurality of scores associated with the quality of the second
plurality of productions from the intelligent system, and
optimizing the strength of a digital watermark using the third
plurality of scores.
BRIEF DESCRIPTION OF THE DRAWINGS
[0012] FIG. 1 is a representation of a photograph illustrating the
decomposition of an image using an automated digital watermarking
method of the present invention;
[0013] FIG. 2 is a flow chart of one embodiment of the automated
digital watermarking method of the present invention, the digital
watermarking method using a neural network;
[0014] FIG. 3 is a plot of a training curve for the neural network
used in accordance with the automated digital watermarking method
of the present invention;
[0015] FIG. 4 is a functional block diagram of one embodiment of an
automated digital watermarking system of the present invention;
and
[0016] FIG. 5 is a functional block diagram of another embodiment
of the automated watermarking system of the present invention.
DETAILED DESCRIPTION OF THE INVENTION
[0017] In a variety of embodiments and examples, the present
invention provides automated methods, systems, and
computer-readable media for digital watermarking using neural
networks. In the description that follows, specific details are set
forth in order to provide a thorough understanding of the present
invention. It will be obvious, however, to one of ordinary skill in
the art that the methods, systems, and computer-readable media of
the present invention may be practiced with or without the
inclusion of some or all of these specific details. For example,
although the description that follows focuses on using a discrete
wavelet transform (DWT) for the space-frequency decomposition of an
image or other production, any other suitable frequency
transformation may be used, including a discrete Fourier transform
(DFT) or a discrete cosine transform (DCT). Likewise, although the
description that follows focuses on the digital watermarking of
"images," any other suitable production may be digitally
watermarked, including audio and video productions. Familiar tools
and techniques have not been described in detail in order to avoid
obscuring the novel features of the present invention.
[0018] A DWT divides a signal into low and high scales (bands). The
high-scale component is split again into low and high frequencies.
This process is repeated a plurality of times. The original signal
may be reconstructed using an inverse discrete wavelet transform
(IDWT). The DWT and IDWT are defined for one-dimensional (1-D) and
two-dimensional (2-D) signals, such as images. The DWT (or IDWT)
for a 2-D image x[m,n] is implemented by applying the 1-D DWT (or
IDWT) separately for each dimension, as follows:
DWT(x[m,n])=DWT.sub.n[DWT.sub.m(x[m,n]) ]. (1)
[0019] An image may be decomposed into a pyramid structure with a
(such as a low-low band, a high-low band, a low-high band, and a
high-high band) using a Mallat pyramid synthesis algorithm. This is
illustrated in FIG. 1, which shows the low-low band 11, the
high-low band 13, the low-high band 15, and the high-high band 17,
and a low-low/low-low band 19, a low-low/low-high band 21, etc. The
DWT for a 1-D signal is mathematically stated as follows: Let
H(w)=.SIGMA..sub.kh.sub.ke.sup.-jkw and
G(w)=.SIGMA..sub.kg.sub.ke.sup.-jkw be the low-pass and high-pass
filters, respectively. A signal x[n] may be decomposed recursively
as: c.sub.j-1,k=.SIGMA..sub.nh.sub.n-2kc.sub.j,n, and (2)
d.sub.j-1,k=.SIGMA..sub.ng.sub.n-2kc.sub.j,n, for j=J+1, J, . . . ,
J.sub.0, (3) where c.sub.J+1,k=x[k], k.epsilon.Z,J+1 is the
high-resolution level index and J.sub.0 is the low-resolution level
index. The coefficients c.sub.J0,k, d.sub.J0,k, d.sub.J0+1,k, . . .
d.sub.J,k are called the DWT of signal x[n], where C.sub.J0,k is
the lowest-resolution component of x[n] and d.sub.j,k are the
details of x[n] at a variety of scales. Furthermore, the signal
x[n] may be reconstructed from its DWT coefficients recursively, as
follows:
c.sub.j,n=.SIGMA..sub.nh.sub.n-2kc.sub.j-1,k+.SIGMA..sub.ng.sub.n-2kd.sub-
.j-1,k. (4) To ensure the DWT and IDWT relationships described
above, the following orthogonality condition on H(w) and G(w) is
needed: |H(w)|.sup.2+|G(w)|.sup.2=1.
[0020] A plurality of DWT-based digital watermarking methods have
been proposed. For a majority of these methods, a multi-resolution
digital watermarking technique is used to add a digital watermark
to the high-frequency (low-scale) bands of the DWT of an image.
Because the high-frequency components represent the edges of the
image, the digital watermark is not visible to the human eye.
Unavoidably, this makes the digital watermark vulnerable to attacks
such as low-pass and median filtering. These techniques have been
tested using regular attacks, such as wavelet-based image
compression, rotation and transformation, low-pass filtering,
cropping, additive noise corruption, image rescaling/stretching,
and image half-sizing. All of the algorithms described above either
add a digital watermark to the low bands (high frequencies) or
apply a threshold-based selection to add a digital watermark to
significant DWT coefficients.
[0021] Referring to the methods of the present invention, given an
image, a DWT is calculated and a digital watermark is inserted in
all frequencies. The digital watermark is added to every band with
a different weight. The insertion is based upon the following
equation: c.sub.i'=c.sub.i(1+.alpha.m.sub.i), (5) where .alpha. is
a scaling parameter, c.sub.i is the coefficient of the original
image, m.sub.i is the digital watermark to be added, and c.sub.i'
is the watermarked coefficient. It should be noted that m.sub.i
follows a normal distribution.
[0022] The digital watermark recovery procedure requires the
original coefficients where the digital watermark has been added
and the digital watermark is extracted from all frequency bands.
Digital watermark detection is based upon the following equation:
d.sub.i=(c.sub.i'-c.sub.i)/.alpha.c.sub.i. (6) The correlation
between the extracted numbers and the digital watermarks of each
band is then calculated. A relatively high correlation indicates
the existence of the digital watermark in a given band. The
correlation is calculated as follows: Correlation: sim(X,X')=(X'X)/
{square root over (X'X')} (7) where ( X ' .times. X ) = i = 1 n
.times. d i * m i . ##EQU1## When the correlation calculated above
is large, the existence of a digital watermark is detected.
Preferably, in order for the methods of the present invention to be
more resistant to translation attack, the correlation is calculated
for several positive and negative shifts of the digital watermark
and a relatively large spike in such a function detects the digital
watermark.
[0023] Using, for example, a Daubechies 1 (db1) wavelet, a digital
watermark length of 1,000 in both the low and high bands, a scaling
parameter of 0.01 for the low-frequency bands, and a scaling
parameter of 0.1 for the high-frequency bands, no noticeable
difference is observed between an original image and a digitally
watermarked image. The methods of the present invention were
compared with two conventional digital watermarking methods. This
comparison indicated that human subjects cannot detect the presence
of digital watermarks created using the methods of the present
invention.
[0024] Tables 1a, 1b, and 1c illustrate the resistance of the three
digital watermarking methods to a variety of attacks. Table la
incorporates the methods of the present invention. TABLE-US-00001
TABLE 1a Method of the Present Invention (Digital Watermark on
Approximation Coefficients: Length = 1,000, Scaling Parameter =
0.01; Digital Watermark on Detail Coefficients: Length = 1,000,
Scaling Parameter = 0.1) Low-Pass Median No Filter Filter Gaussian
JPEG Rotate Translate Image Degradation (3 .times. 3) (3 .times. 3)
Noise Quality = 30 (Center, 1.degree.) (2 pixels right) Half-Sizing
Teapot 13.4225 2.4924 3.6129 2.9725 4.5604 0.9918 0.9491 3.3602
Peppers 12.9801 2.5339 9.2016 7.1644 6.9962 0.6524 0.8605 7.8442
Woman 12.5057 3.0409 8.7305 5.7235 6.8542 0.8516 0.5831 6.4926
[0025] TABLE-US-00002 TABLE 1b Conventional Method A (Scaling
Parameter = 0.4, Threshold for Adding Digital Watermark = 40,
Threshold for Detection = 50) Low-Pass Median No Filter Filter
Gaussian JPEG Rotate Translate Image Degradation (3 .times. 3) (3
.times. 3) Noise Quality = 30 (Center, 1.degree.) (2 pixels right)
Half-Sizing Teapot 1.8431 0.2488 0.8301 1.6103 1.0918 0.6551 0.2602
0.6818 Peppers 2.8755 0.9313 1.3323 2.5259 1.9602 0.3849 0.7402
1.1969 Woman 2.3295 0.6083 1.3853 2.2077 1.0957 0.2346 0.0223
1.5723
[0026] TABLE-US-00003 TABLE 1c Conventional Method B (Digital
Watermark on Detail Coefficients: Length = 500, Scaling Parameter =
0.001) Low-Pass Median No Filter Filter Gaussian JPEG Rotate
Translate Image Degradation (3 .times. 3) (3 .times. 3) Noise
Quality = 30 (Center, 1.degree.) (2 pixels right) Half-Sizing
Teapot 9.1822 0.5905 0.8798 1.1026 0.6928 0.7331 0.5480 0.5937
Peppers 9.6499 1.1804 1.9448 3.6405 4.0427 0.6871 0.6730 1.7478
Woman 9.4054 1.0896 1.4389 2.6512 3.8851 0.5076 0.5891 2.8291
[0027] The value in each cell of Tables 1a, 1b, and 1c is the ratio
of the correlation value of the added digital watermark and the
maximum correlation value from 300 randomly generated marks. As is
illustrated, the digital watermark may be relatively easily
detected in the digitally watermarked image (with no degradation)
using a correlation test, i.e. the largest peak in the correlation
is about 12-14 times larger than the remainder of the peaks,
proving the existence of the digital watermark. Tables 1a, 1b, and
1c illustrate that digital watermarks generated using the methods
of the present invention are resistant to low-pass filtering,
median filtering, JPEG compression (quality=30), and half-sizing.
The resistance of the digital watermark to low-pass filtering is
expected as a digital watermark added to relatively high
frequencies is susceptible to filtering and may be removed
relatively easily. The methods of the present invention also
provide enhanced resistance to geometric transformations, such as
translation and rotation.
[0028] Another aspect of a successful digital watermarking method
is ensuring that the digital watermark added is optimal. In other
words, in order for the digital watermark added to be resistant to
different types of attacks, it is necessary to optimize the power
of the digital watermark, while still ensuring that the digital
watermark is not visible to the human eye. In one embodiment, a
method of the present invention defines a neural network-based
algorithm that automatically selects and controls digital watermark
parameters, creating maximum-strength or optimal digital
watermarks. Typically, this process involves generating a digitally
watermarked image, allowing one or more persons to judge the image
to ensure that the digital watermark is invisible, and repeating
the process with increased digital watermark power until just
before the digital watermark becomes visible. The method of the
present invention replaces the human of the conventional process
with an intelligent system, such as a neural network, allowing the
process to be automated. The intelligent system may be, for
example, a neural network, an expert system, a fuzzy model, or any
other suitable system that may learn and mimic the behavior of a
complex non-linear system, such as the human visual system (HVS).
Neural networks and the like may be trained to accurately and
reliably model HVS perception of the quality of digitally
watermarked images.
[0029] Referring to the method 10 of FIG. 2, to train the neural
network, a database of images is created. (Block 12). Preferably,
these images include original images 14, digitally watermarked
images 16, standard images 18, and non-standard images 20. The
images may be digitally watermarked with differing power levels for
the wavelet. The quality of the images is then judged by a
plurality of human subjects. (Block 22). Specifically, each of the
plurality of human subjects assigns a score between 0 and 100 to
each of the images. A score of 0 indicates that the there is no
perceivable difference between a given original image and the
corresponding digitally watermarked image. A score of 100 indicates
that a given digital watermark has highly distorted the
corresponding original image. Optionally, these scores are then
converted to a number between 0 and 1, allowing the use of a
logarithmic-based sigmoid activation function for the output layer.
(Block 24). The information described above is then provided to the
neural network. (Block 26). This information includes the images
28, the power level of the digital watermarks 30, and the scores
32. Using the information, the neural network is trained to
approximate a visibility score that a human subject would assign to
a given digitally watermarked image. In one embodiment of the
present invention, the neural network is a multi-layer sigmoid
neural network including a back-propagation training algorithm.
[0030] FIG. 3 illustrates a training curve 40 for the method
described above. After only 300 epochs, the neural network provides
a negligible scoring error. Table 2 illustrates the scoring of a
different set of digitally watermarked images (a non-training set)
with digital watermarks of varying power using the neural network,
demonstrating how well the trained neural network approximates the
HVS. TABLE-US-00004 TABLE 2 Comparison of Neural network and Human
Digital Watermark Visibility Scores (Digital Watermarking
Algorithm, Decomposition Level 2, db1, Threshold = 40) Digital H
Digital Test Image Watermark NN Digital Watermark Watermark
(Section) Power (.alpha.) Score (scaled by 100) Score Woman (1) 0.4
9.7 10 Woman (2) 0.8 9.8 10 Woman (12) 0.8 27.1 25 Woman (13) 0.4
41.4 40
[0031] Referring to FIG. 4, in another embodiment of the present
invention, after the neural network 54 is trained and tested, an
original image 50 may be presented to a wavelet transform 52, the
neural network 54, and a watermarking algorithm 56 to produce a
digitally watermarked image 58. The neural network 54 is
responsible for deeming that the digital watermark has achieved
maximum power and yet is still invisible to most humans. It should
be noted that the wavelet transform 52 and the neural network 54
may be replaced by any suitable frequency transformation system and
expert system, respectively. Preferably, the original image 50 is
divided into a plurality of blocks and each of the plurality of
blocks is digitally watermarked separately. A comparison of the
original blocks and the digitally watermarked blocks reveals that
the neural network 54 digitally watermarks the blocks such that the
digital watermarks are invisible to the HVS.
[0032] Referring to FIG. 5, in a further embodiment of the present
invention, after the neural network 64 is trained and tested, an
original image 60 may be exposed to a watermarking technique 62,
the neural network 64, and a strength factor adjustment 66
incorporating a strength factor (.alpha.) 68 to produce a
watermarked image 70. Again, the neural network 64 is responsible
for deeming that the watermark has achieved maximum power and yet
is still invisible to most humans. It should be noted that the
neural network 64 may be replaced by any suitable frequency
transformation system and/or expert system. Preferably, the
original image 60 is divided into a plurality of blocks and each of
the plurality of blocks is watermarked separately. A comparison of
the original blocks and the watermarked blocks reveals that the
neural network 64 watermarks the blocks such that the watermarks
are invisible to the HVS.
[0033] It should also be noted that a given type of image (or piece
of speech or shot of video) is capable of accepting a predetermined
level of digital watermarking, this level varying from one type of
image to another. For example, a portrait's capacity to accept a
digital watermark is different from that of a landscape scene.
Thus, for each type of image, a predetermined type of expert system
is trained. During the automated digital watermarking process, the
type of image is first recognized, and then an optimized digital
watermark is added.
[0034] The methods, systems, and computer-readable media of the
present invention may be used in conjunction with conventional
digital and non-digital watermarking applications. In such cases,
the present invention not only automates the watermarking process
and creates a means for the real-time implementation of
watermarking, but also eliminates the need for and costs associated
with human supervision. Likewise, the subjectivity associated with
human scoring is eliminated. Potential applications include, for
example, copyright protection for images, pictures, videos,
multimedia products, and other digital products produced by news
agencies and the like; copyright protection for CDs, DVDs, and
other digital products produced by entertainment companies and the
like; and the protection of biometrics using maximum-strength
digital watermarks.
[0035] The methods, systems, and computer-readable media of the
present invention may also be used where watermarking has not
typically been used. Such applications include: 1) the protection
and distributional control of medical data, such as MRI, FMRI, CT,
and ultrasound images; 2) online and real-time digital watermarking
of pictures and videos in the hardware of digital cameras; 3)
copyright protection of VLSI chips via the watermarking of the body
of the chips; 4) protection against reverse engineering and the
unauthorized duplication of vital parts in manufacturing industries
such as the automotive, aerospace, and defense industries; and 5)
the indexing of negative films of photographs.
[0036] It is apparent that there has been provided, in accordance
with the present invention, automated digital watermarking methods,
systems, and computer-readable media using neural networks. While
the present invention has been particularly shown and described in
conjunction with examples and preferred embodiments thereof, it
will be appreciated that variations in and modifications to the
present invention may be effected by persons of ordinary skill in
the art without departing from the spirit or scope of the
invention. It is to be understood that the principles described
herein apply in a similar manner, where applicable, to all such
examples and embodiments which the following claims are intended to
cover.
* * * * *