Transcript Introduction

Maximizing Strength of Digital
Watermarks Using Neural
Network
Kenneth J.Davis; Kayvan Najarian
International Conference on Neural Networks,
2001. Proceedings.
Presented by Bin-Cheng Tzeng
5/21 2002
Outlines




Introduction
A Watermarking Technique in the
DWT Domain
Neural Technique for Maximum
Watermark
Conclusions
Introduction


For watermarking to be successful
1.Unobtrusive
2.robust
In other words, one would like to
insert the watermark with maximum
strength before it becomes visible to
the human visual system(HVS)
Introduction(Cont.)


The way the strength of the added
watermark is chosen is of highest
importance.
This paper attempts to define a neural
network based algorithm to
automatically control and select the
watermarking parameters to create
maximum-strength watermarks.
A Watermarking Technique in
the DWT Domain


The paper use a wavelet-based scheme
for digital watermarking.
(reference “A New Wavelet-Based
Scheme for Watermarking Images”)
The technique was tested by cropping,
JPEG compression, Gaussian noise,
halfsizing, and median filtering.
A Watermarking Technique in
the DWT Domain
A Watermarking Technique in
the DWT Domain


A threshold was used to determine the
significant coefficients.
The watermark is added to the
significant coefficients of all the bands
other than the low pass subband.
A Watermarking Technique in
the DWT Domain
 : The scaling parameter
ci : The coefficient of the original image
mi: The watermark to be added
ci’ : the watermarked coefficient
Neural Technique for
Maximum Watermark

To achieve maximal watermarking while
remaining invisible to the human eye.
1.Generating a watermarked image
using a given power
2.allowing one or more persons to
judge the image,repeat while
increasing the power until the
humans deem the watermark visible
Neural Technique for
Maximum Watermark


Replacing the humans in the process
with a neural network allowing the
process to be automated.
To train the neural network, a database
of original and watermarked images
whose qualities are judged by several
human subjects is being created.
Neural Technique for
Maximum Watermark


When judging the images, a score is
given between 0 and 100
0 means no perceivable difference
between the original image and
watermarked image and 100 means the
watermark has highly distorted the
image.
Neural Technique for
Maximum Watermark




Feed forward back-propagation network
Being able to properly approximate
non-linear functions and if properly
trained will perform reasonably well
when presented with inputs it has not
seen before
HVS is non-linear
To be useful.
Neural Technique for
Maximum Watermark
Neural Technique for
Maximum Watermark



Each image is subdivided into blocks of
64x64 pixels to be treated as a
complete image.
4096 inputs and 1 final input ()
The hidden layer with 256 or 512
neurons
Neural Technique for
Maximum Watermark


The network is trained using the scaled
conjugate gradient algorithm(SCG)
Trained for 300-600 iterations or until
the mean square error is less than
0.00001
Comparison of Neural Network and
Human watermark visibility scores
Conclusions



The watermark is added to both low
and high scales of DWT.
To aid in maximizing the watermark a
neural network that mimics the HVS
was proposed.
When properly trained, the neural
network can allow it to be used in place
of several human reviewers.