We have solved the correspondence problem by applying the matching process in two levels, the first level is Feature based matching, in which we have extracted the features of both images by creating multi-resolution images and applying histogram segmentation. The resulting features are region features; a comparison is done between the regions in the first image with the regions of the second image to get the disparity map.
The second level is Area-based matching in which we applied the Wavelet transform to get an expected window size as a search area for each pixel. We have joined the two levels to obtain more accurate pixel by pixel correspondence. We also obtained an adaptive search range and window size for each pixel to reduce the mismatches. Our procedure introduced high accuracy results and denser depth information.
The depth information is used to get the final 3D model – using only pair of images will create 2.5D model, using more than pair of images will create 3D model, we will refer to 3D model as a general output of stereo reconstruction– After reconstructing the model, in some applications it is needed to be published online. For example suppose the reconstructed model is a model for Sphinx – Famous statue in Egypt – The reconstruction for the model can be done in many days or months; then the model will be published online to let Internet users around the world watch the model. Therefore, techniques should be used to protect the copyright for that model. We have applied new fragile watermarking technique to secure the 3D reconstructed model and protect its copyright.
4. Stereo image definition
Stereo images work as the human eyes, since the pair of eyes take pair of
images of the same view from different angles.
5. The relation between Disparity and Depth
Disparity of Sl
Centroid of Sr
Centroid of Sl
Sl Sl
Sr
Left Image Right Image
Disparity is the shift of the objects in both images, or the difference between the
object positions in both images.
????
Depth is inversely proportional to the disparity, so the objects near the camera
will be shifted more than the objects which is far from the camera
d : Disparity
f : distance from the image plan to the camera
T : The baseline of the two cameras
Z : the third dimension or Depth.
6. The problems of the reconstruction process
Segmentation Matching
Disparity map
3D scene
- Finding the best segmentation method
- The correspondence problem
- The occluded regions
- The density of the 3D scene
7. The problems of the reconstruction process
Segmentation
?
Match
Segmentation
FFoorr eevveerryy ffeeaattuurree iinn
tthhee lleefftt iimmaaggee wwee
sseeaarrcchh ffoorr tthhee
ccoorrrreessppoonnddeenntt ffeeaattuurree
iinn tthhee eennttiirree rriigghhtt
iimmaaggee ffeeaattuurreess
AA ddiissttiinnccttiivvee
ddeessccrriippttoorr ffoorr tthhee
ffeeaattuurreess aarree nneeeeddeedd
ttoo rreedduuccee tthhee
mmiissmmaattcchheess aanndd
oobbttaaiinn hhiigghh aaccccuurraaccyy..
9. The Proposed Stereo Image Watermarking
Approach
1. Stereo Images are reconstructed (make disparity estimation) to obtain disparity map.
2. Disparity map is used as a watermark and is embedded into the left image of the stereo image
pair.
3. The host image is transferred through insecure channel to the receiver.
10. Disparity Estimation
Disparity Estimation is done in two levels
First: Feature based matching, the first level is to extract the feature of
the pair of images and then compare these features together. The
comparison is done according to specific properties for each feature
Second: Wavelet-based matching, in this level we are going to make
dense matching by comparing the pair of images window by window
leading to more accurate matching.
12. Disparity Estimation
(Feature Based Matching)
Histogram segmentation: can be viewed as an operation that involves tests
against a function T of the form
T =T[x, y, p(x, y), f (x, y)] (3)
Left Image Right Image
(4)
if < f x y £
T
0 ( , ) 1
if T < f x y £
T
1 ( , ) 2
if T < f x y £
T
2 ( , ) 3
3 ( , ) 255
g
1
2
3
4
( , )
< £
ì
Î
ï ï
g
í
ï ï
î
if T f x y
g
g
Pixl x y
13. Disparity Estimation
(Feature Based Matching)
For each group a series of (low, medium and high) resolution images will be
created, It is created by reduction in size by a factor of two using Gaussian
convolution .) filter h(x
( ) 2 (5) h x = psAe-2p 2s 2x2
A binary map for each of multi-resolution images will indicate, for each pixel,
). whether it belongs to the group or not, as shown by equation (6
(6)
if f x y T
1 ( , )
>
0 ( , ) .
( , )
î í ì
£
=
if f x y T
g x y
Thus the pixels labeled 1 correspond to the region segment, whereas pixels
labeled 0 correspond to the background
14. Disparity Estimation
(Wavelet Based Matching)
For each pixel in , the stereo images
- Execute a convolution by an adaptive window size.
- The correspondence search range will be limited by using the disparity map
gained from feature based matching.
- The correspondence will be estimated by using SSD,
å
AI x y AI x d y
+ + - + + +
x h x h
[ ( , ) ( , )]
x m
( , )
l r
å å
2 2
AI x y AI x y
+ + + +
=
x h x h
( , ) ( , )
x m x m
( , ) ( , )
C
l r
:The disparity estimate for a pixel x,y is the one which minimizes SSD error
)d(x,y) = arg min C(x,y,d
16. Experimental Results
(Image acquisition techniques)
the image acquisition techniques, (a) The Object Registration device. Source: Borg Al Arab
City for Scientific Research and Technology Applications, Informatics Research Institute,
Alexandria, Egypt (b) Aircraft Stereo images
17. Experimental Results
(Disparity Estimation)
AAllggoorriitthhmm NNuummbbeerr ooff mmaattcchhiinngg
lleevveellss
MMiissmmaattcchheess
QQuuaannttiizzaattiioonn aanndd MMoodduullaattiioonn 11 lleevveell 2233 %%
DDCCTT bbaasseedd 11 lleevveell 2211 %%
DDWWTT bbaasseedd 11 lleevveell 2299 %%
FFrraaccttiioonnaall FFoouurriieerr ttrraannssffoorrmm ((FFrrFFTT)) 11 lleevveell 1199 %%
GGeenneettiicc AAllggoorriitthhmm ((GGAA)) 11 lleevveell 1166 %%
OOuurr AAllggoorriitthhmm 22 lleevveellss 44 %%
Due to the two levels approach we have reduced the correspondence mismatches
18. Experimental Results
(Random LSB Stereo Image Watermarking)
The PSNR value is used to measure the quality of watermarking in our proposed stereo image
watermarking. The PSNR value between host image I and the host image I’ is calculated as
PSNR 20log 255 10
ö çè
÷ø
= æ
RMSE
Where
åå[ m
]
= =
1
= -
i
n
j
I i j I i j
m n
RMSE
1 1
( , )2 '( , )2
*
Lower values of PSNR mean less invisibility of the watermark, while higher PSNR represents
better invisibility of the watermark in the host image.
19. Experimental Results
(Disparity Estimation)
The proposed disparity estimation algorithm results applied on Tsukuba image, (a) The
original image (b) the disparity map obtained in [7] (c) disparity map obtained in [9] (d) The
disparity map obtained in [6] (e) The disparity map obtained in [10] (f) The disparity map
obtained by our proposed approach
21. 3D Watermarking ( Main idea )
Localized embedding
Using ratios of 2-D and 3-D geometrical measures
3D Triangular mesh model
22. 3D Watermarking ( Attacks 1/2 )
Rotation, translation, and uniform scaling
This kind of operations disables the watermark detector by changing the
orientation, location and/or scale of the model to make the detector not able
to locate the watermark.
Polygon simplification
This kind of operations changes the coordinates of vertex and the topology of
the model by reducing the number of faces in a dense mesh while minimally
perturbing the shape.
Mesh smoothing
This operation attacks the watermark by deleting some high-frequency
components of the model and smoothing the model surface.
23. 3D Watermarking ( Attacks 2/2 )
Re-meshing
This kind of operations attacks the watermark through using another
sampling mesh to representing the same model.
Reordering of points
This operation scrambles the vertices in the 3D model to destroy the
watermark information embedded in the store sequence of the vertices.
Cropping
This operation removes parts of the model to destroy the watermark
24. Problems of 3D watermarking
The Causality Problem - 1
The location of a former processed vertex will be changed by perturbing the latter
processed vertex
V1
V2 V3
V1
V2 V3
V1
V2
V3
We have solved this problem by proposing The Contagious Diffusion Technique for
)traversing the model (details later
25. Problems of 3D watermarking
The Convergence Problem - 2
In embedding stage, it required to perturb all the pixels in the model to every pixel value
index equal to the pixel’s location index. In practice, some vertices should be perturbed
more and more until it satisfy the requirements
)We have solved this problem by proposing our NNM technique (details later
26. The Proposed Algorithm (1/5)
Our proposed algorithm steps
The original 3d model ( 3ds file)
Extract VBT (Vertex Body Table)
Extract PNT from VBT
Navigate by using Contagious
Diffusion Technique.
Embed the watermark by
changing topology
The watermarked 3d model (3ds
file)
1. Extract VBT from the saved model
file.
2. Extract PNT from the saved model
file.
3. Navigate through the model by
using Contagious Diffusion
Technique.
4. Embed the watermark by using
NNM (Nearest Neighbor Move)
technique.
27. The Proposed Algorithm (2/5)
.Extracting VBT from 3DS file- 1
The 3ds file is saved on the hard disk as chunks; each chunk has a unique address.
Our algorithm extracts the chunks that represent the model vertices. The vertices are
then saved into a database table
28. The Proposed Algorithm (3/5)
Extracting PNT from VBT- 2
For each polygon we search for a polygon sharing the same edge with the current
polygon. We notice this way is a time consuming since the search will be done by
comparing each edge (two vertices) with each other two vertices in the table. When a
matching edge is found, this means the edge (two vertices) belongs to the same
polygon and the tow polygons are neighbors
29. The Proposed Algorithm (4/5)
Contagious Diffusion Technique- 3
In The Contagious Diffusion Technique the polygon has three statuses,
Suspected, means it has no embedded data, and its entire neighborhood has no
embedded data.
Infected, means some of its neighbor are infected others are not.
Immune means all the polygon neighbors are entirely infected and have embedded
data.
30. The Proposed Algorithm (5/5)
(The embedding Procedure (Nearest Neighbor Move NNM- 4
x = x + h -
( ) (2) 1 2 1 x x
L
z = z + h -
( ) (4) 1 2 1 z z
L
y = y + h -
( ) (3) 1 2 1 y y
L
where L is the Euclidean distance between vertices Va and Vb, which can be calculated with
( ) ( ) ( )2 (5)
2 1
2
2 1
2
2 1 L = x - x + y - y + z - z
31. Experimental Results
Tools
3D Max Software to build up our models – 1
C++ as a programming language – 2
OpenGL Library to Embed and Extract the secret data into the 3DS files – 3
MMooddeell
NNaammee SSppaacceesshhiipp CCuubbeess HHeeaarrtt CChheessss ppaawwnn
VVeerrttiicceess
664499 11772288 886611 11003366
PPoollyyggoonnss
550000 22559922 11771177 22006688
33. Experimental Results
The effect of changing on the visual quality of the watermarked model. The
experiment is done on three models, Helmet, Chesspawn and Spaceship.
34. Experimental Results
(a),(b) and (c) are the cover models for Chesspawn, Spaceship and Helmet respectively. (d),
(e) and (f) are the watermarked 3D models.
37. Conclusions & Future Work
The process of stereo image reconstruction is enhanced by applying
two levels of disparity estimation
A highly secure stereo image watermarking system has been
implemented. The Random LSB is applied together with a key to
increase the security of the watermarked images.
High visual quality of 3D models is obtained after Watermarking the 3D
model. The two common problems of 3D watermarking has been solved,
namely the causality problem and the convergence problem.
In Future Work, We are planning to perform the watermarking for 3D
models in the frequency domain and evaluate the results.