WATERMARKING USING FEATURE POINTS
Madhukar V Kakaraparthi
Dr. Donald Adjeroh
Cs591 Multimedia Systems
What is WaterMarking?
• Is the practice of imperceptibly altering a image to
embed a message about that image
Two Important properties of WM:
Imperceptibility– The modifications caused
by watermark embedding should be below the
Robustness– The ability of the watermark to
resist distortion introduced by standard or
malicious data processing.
A watermark may be
Fragile: distorted or broken under slight changes
Semi-fragile: designed to break under all changes
that exceed a user- specified threshold
Robust: withstand moderate to severe signal
processing attacks (compression, rescaling, filtering)
Applications Of WaterMarking
1. Copyright Protection
Most prominent application
Embed information about the owner to prevent others from claiming
Require very high level of robustness
2. Database Retrieval
For the identification of the content.
Robustness not decisive
Embed a watermark to detect modifications to the cover
The watermark in this case has low robustness, “fragile”
The detection of the watermark requires a
synchronization step to locate the embedded
mark in the content.
We can illustrate this idea in:-
2. Digital Images.
Classical Self Synchronization
Using Periodical Sequences
Using Templates insertions
Using Invariant transforms
Using original image.
Comparison of the Sync
Local Global Notes
Periodic No Yes -
Template No Yes Can be
Invariant No Yes -
Non-blind Yes Yes Computational
Content Based WaterMarking
This represents the new class of the water
marking techniques which link the water mark
with image semantics and not the image
The problem of geometrical synchronization is
solved because the image content represents an
invariant reference to geometrical
Principle of a content based
water Marking Scheme
Alghoniemy and tewfik
In the embedder side:
Detect robust feature points in an image.
Generate a triangular tessellation of the image based on
the set of feature points.
Map a triangular spread sequence (watermark) into each
triangle of the tessellation via affine transform.
Add the mapped sequence on each triangle.
In the detector side
Reconstruct the tessellation.
Map each triangle to the shape of the original triangular
Compute the correlation of each mapped triangle with the
Accumulate the correlations to detect the watermark over
the whole image.
Feature Point Detectors
Achard Rouquet Detector
Detector Bench Mark
Score= Nb pre Nbcre Nbdes
Nb pre Nbdes
Basing on the score Harris Detector is choosen.
• A random sequence Tw is generated of the shape of a right angle isosceles
traingle(64 X 64).
• Improved Harris detector is used for the detection of feature points
P pi R 2 , i 1,...., n
• A Delaunay tessellation of the set P is the unique triangulaton of the convex cover of
P such as the interior of the containing circle C pi , p j , pk of each traingle of P 3 does
not contain another vertex of P
• An affine transformation is performed on triangle Tw to map the shape of Tk .The
affine function A is defined by six real parameters a,b,c,d,e,f.
xm a b xw e xw
Axw , yw
y c d y f L y T
m w w
• Tmis multiplied by a visual mask is
based on image luminance.
Tp i, j i, j Tm (i, j )
• The marked triangle Ts T p Tk
• Ts is substituted for Tk
A random sequence T is generated of the shape of a right angle isosceles traingle(64 X 64).
wis used for the detection of feature points
Improved Harris detector
• A Delaunay tessellation T i0 i N
T , is performed using image
• feature points. The mark is inserted in each triangle T T
• The traingle k is warped into right angled isosceles triangle (64 X 64).
• TL is given to a wiener filter and using ˆ
Twstatistics to obtain Tw
Vi Tw i, j
Tw i, j
ˆ TL i, j M l TL i, j
Vi Tw i, j Vl TL i, j
• Weiner prediction is a denoising operation as it allows separation of
the image components from the marked components
• The mark is detected inside the traingle only if:
corr (Tw ; Tw ) Pfa
• A global decision is obtained using the global sum of ˆ
corr (Tw ; Tw )
• The mean of correlation is Z i w ; Tw
i 1,...., N
• The mark can be also detected if: Z
• The final decision is obtained from the different detection results
So to summarize the paper:
The detection of the mark doesnot require the original
Using content-based techniques we obtain automatic
resynchronization of the mark after both local and global
The orientation of the signature is carried by the content
of the image and consequently cannot be erased,
Content Based techniques donot depend on the
template insertion which is easily removed.
• The robustness of the scheme is
The capacity of the feature point detector to
preserve the feature points after geometrical
The content of the image. Like highly textured
images are more error prone.