Statistical Gaussian Model of Image Regions in Stochastic Watershed Segmentation

Statistical Gaussian Model of Image Regions in Stochastic Watershed Segmentation
Statistical Gaussian Model of Image Regions
in Stochastic Watershed Segmentation
Jess Angulo
jesus.angulo@mines-paristech.fr ; http://cmm.ensmp.fr/angulo
MINES ParisTech, PSL-Research University,

CMM-Centre de Morphologie Mathmatique
GSI'2015 - 2nd Conference on Geometric Science of Information

Ecole Polytechnique, Paris-Saclay (France) - October 28th-30th 2015
1 / 34
Motivation: Unsupervised segmentation of generic
images
Custard: Color image
Large homogenous areas, well contrasted objects as well as textured zones

and fuzzy boundaries
2 / 34
images
Custard: its color gradient image
Large homogenous areas, well contrasted objects as well as textured zones

and fuzzy boundaries
3 / 34
images
Custard: pdf of contours using stochastic watershed
Using watershed based techniques large homogeneous areas are

oversegmented and textured zones are not always well contoured
4 / 34
images
Custard: h-dynamics watershed cut from SW pdf, h = 0.1

5 / 34
images
Custard: h-dynamics watershed cut from SW pdf, h = 0.3

6 / 34
Context and goal
7 / 34
Context and goal
Context:
Segmentation approaches based on statistical modeling of pixels and
regions, e.g, mean shift and statistical region merging
Hierarchical contour detection and segmentation, e.g., machine
learned edge detection, watershed transform
Stochastic watershed (SW): to estimate a probability density
function of contours
7 / 34
Context and goal
Context:
Segmentation approaches based on statistical modeling of pixels and
regions, e.g, mean shift and statistical region merging
Hierarchical contour detection and segmentation, e.g., machine
learned edge detection, watershed transform
Stochastic watershed (SW): to estimate a probability density
function of contours
Goal: Take into account regional information in the probability

estimation by SW by means of a statistical gaussian model More
perceptual strength function of contours
7 / 34
Plan
1 Stochastic Watershed using MonteCarlo Simulations
2 Multivariate Gaussian Model of Regions in SW
3 Perspectives
8 / 34
Stochastic Watershed using MonteCarlo Simulations
3 Perspectives
9 / 34
Regionalized Poisson points
Uniform random germs
Generate realizations of a Poisson point process with a constant

intensity (i.e., average number of points per unit area)
Random number of points N(D) falling in a domain D (bounded

Borel set), with area |D|, follows a Poisson distribution with
parameter |D|, i.e.,
n
(|D|)
Pr{N(D) = n} = e |D|
n!
Conditionally to the fact that N(D) = n, the n points are
independently and uniformly distributed over D , and the average
number of points in D is |D| (i.e., the mean and variance of a
Poisson distribution is its parameter
10 / 34
Regionalized random germs
Let us suppose that the density is not constant; but considered as

measurable positive-valued function, dened in Rd . For simplicity,
(x)d x
R
let us write (D) =
Number of points falling in a Borel set B according to a regionalized
density function follows a Poisson distribution of parameter (D),
i.e.,
n
((D))
Pr{N(D) = n} = e (D)
n!
If N(D) = n, the n are independently distributed over D with the
probability density function:
b x) = (x)/(D)
(
11 / 34
Generate N random germs in an image m : E {0, 1} according to
density (x)
b using inverse transform sampling
1 Initialization: m(xi ) = 0 xi E ; P = Card(E )

P
(x
b k)
2 Compute cumulative distribution function: cdf (xi ) = ki
PP
k=1
(x
b k)
3 for j =1 to N
4 rj U (1, P)
5 Find the value sj such that rj cdf (xsj ) .
6 m(xsj ) = 1
12 / 34
Stochastic watershed paradigm
Spreading random germs as markers on the watershed segmentation.

This arbitrary choice is stochastically balanced by the use of a given
number M of realizations, in order to lter out non signicant
uctuations
Each piece of contour may then be assigned the number of times it

appears during the various simulations in order to estimate a
probability density function (pdf ) of contours
In the case of uniformly distributed random germs, large regions will

be sampled more frequently than smaller regions and will be selected
more often
Image gradient as density for regionalization of random germs

involves sampling high contrasted image areas: probability of
selecting a contour will oer a trade-o between strength of the
contours and size of the adjacent regions
13 / 34
Probability density of contours using MonteCarlo
simulations of watershed
Let {mrkn (x)}M

n=1 be a series of M realizations of N spatially
distributed random markers according to its gradient image g
Each realization of random germs considered as the marker image
for a watershed segmentation of gradient image g in order to obtain
the binary image:

1 if x Watershed lines
WS(g , mrkn )(x) =
0 if x/ Watershed lines
Probability density function of contours is computed by the kernel

density estimation method:
M
1 X
pdf (x) = WS(g , mrkn )(x) K (x).
M n=1
where the smoothing kernel K (x) is a spatial Gaussian function of

width
14 / 34
Color image f (x) Color gradient g (x)
15 / 34
{mrkn (x)}M
n=1 : M realizations of N regionalized Poisson points of
(x) = g (x)
{WS(g , mrkn )}1nM : Watershed segmentations
16 / 34
Density of contours pdf (x) Segmented with h = 0.1
17 / 34
18 / 34
Multivariate Gaussian Model of Regions in SW
3 Perspectives
19 / 34
Watershed transform Tessellation
Tessellation of E from watershed WS(x): (Finite) family of disjoint

open sets (or classes, or regions)
= {Rr }1r N , with i 6= j Ri Rj =

such that
[
E = r Rr WS(x) WS(x) = E \ r Rr = li,j
Boundary between regions Ri and Rj (1 i, j N , i 6= j): Irregular arc

segment
li,j = Ri Rj
20 / 34
Color regions as multivariate normal distributions
The color image values restricted to each region of the partition,

Pi = f (Ri ), can be modeled by dierent statistical distributions
Here we focuss on a multivariate normal model
Pi N (i , i ),
of mean i and covariance matrix i

Dierent (statistical) distances are dened in the space of N (i , i )
Boundary li,j will be weighted with a function depending on the
distance between N (i , i ) and N (j , j )
21 / 34
Distances for multivariate normal distributions
Bhattacharyya distance DB (P1 , P2 )

It measures the similarity of two discrete or continuous probability
distributions P1 and P2 by computing the amount of overlap
between the two statistical populations:
Z p
DB (P1 , P2 ) = log P1 (x)P2 (x)dx
For multivariate normal distributions

1 1 det
DB (P1 , P2 ) = (1 2 )T 1 (1 2 )+ log ,
8 2 det 1 det 2
where
1 + 2
=
2
Note that the rst term in the Bhattacharyya distance is related to

the Mahalanobis distance, both are the same when the covariance of
both distributions is the same
22 / 34
Hellinger metric distance DH (P1 , P2 )

0 DB and it is symmetric DB (P1 , P2 ), but DB does not obey
the triangle inequality and therefore it is not a metric
Bhattacharyya distance can be metrized by transforming it into to

the following Hellinger metric distance
p
DH (P1 , P2 ) = 1 exp (DB (P1 , P2 )),
For multivariate normal distributions

s
1/2
det 1
DH (P1 , P2 ) = 1 e ( 4 (1 2 ) 1 +2 )1 (1 2 ))
T (
det 1 det 2
Hellinger distance is an -divergence, which corresponds to the case

=0 and it is the solely being a metric distance. Hellinger distance
can be related to measure theory and asymptotic statistics
23 / 34
Wasserstein metric distance DW (P1 , P2 )

Wasserstein metric is a distance function dened between probability
measures and on Rn is based on the notion optimal transport:
W2 (, ) = inf E(kX Y k2 )1/2 ,

where the inmum runs over all random vectors (X , Y ) Rn Rn
with X and Y
For the case of discrete distributions, it corresponds to the
well-known earth mover's distance
For two multivariate normal distributions:
q
DW (P1 , P2 ) = k1 2 k2 + Tr (1 + 2 21,2 ),
where
1/2
1/2 1/2
1,2 = 1 2 1 .
In particular, in the commutative case 1 2 = 2 1 one has
2 2 1/2 1 /2 2
DW (P1 , P2 ) = k1 2 k + k 1 2 k .
F 24 / 34
Probability density function MonteCarlo estimation
25 / 34
To assign to each piece of contour li,j between regions Ri and Rj the

normalized statistical distance between the color gaussian
distributions Pi and Pj :
D(Pi , Pj )
i,j = P ,
lk,l WS D(Pk , Pl )
where D(Pi , Pj ) is any of the distances discussed above
25 / 34

D(Pi , Pj )
i,j = P ,
lk,l WS D(Pk , Pl )

For any realization n of SW, denoted WS(x, n), one can compute an
image of weighted contours:
n

i,j if x li,j
Pr (x, n) = n
0 if x/ li,j
25 / 34

D(Pi , Pj )
i,j = P ,
lk,l WS D(Pk , Pl )

For any realization n of SW, denoted WS(x, n), one can compute an
image of weighted contours:
n

i,j if x li,j
Pr (x, n) = n
0 if x/ li,j
Integrating across the M realizations, the MonteCarlo estimate of
the probability density function of contours:
M
1 X
pdf (x) = Pr (x, n) K (x)
M n=1
25 / 34
{WS(g , mrkn )}1nM : Watershed segmentations
{Pr (x, n)}1nM : Weighted contours (Bhattacharyya distance)
26 / 34
Bhattacharyya distance
27 / 34
1
Distance of means Bhattacharyya distance
Hellinger distance Wasserstein distance
1 F. Lpez-Mir, V. Naranjo, S. Morales, J. Angulo. Probability Density Function of

Object Contours Using Regional Regularized Stochastic Watershed. In IEEE ICIP'14. 28 / 34
2
2 F. Lpez-Mir, V. Naranjo, S. Morales, J. Angulo. Probability Density Function of

Object Contours Using Regional Regularized Stochastic Watershed. In IEEE ICIP'14. 29 / 34
Probability density function MonteCarlo Robust
estimation
30 / 34
Probability density function MonteCarlo Robust
estimation
31 / 34
Comparison with SRM
3
Statistical Region Merging (SRM) depending on scale parameter Q
Segmentation for Q = 128
Sum of contours from nine Q

256, 128, 64, 32, 16, 8, 4, 2, 1
Segmentation for Q = 32
3 R.
Nock, F. Nielsen. Statistical Region Merging. IEEE Trans. on PAMI,
26(11):14521458, 2004. 32 / 34
Perspectives
3 Perspectives
33 / 34
Perspectives
Perspectives
In addition to the color, each pixel x described also by its structure

tensor T (x) SPD(2):
Ri : N (0, i ), i = |Ri |1 xRi T (x)
P
Each region
Each region Ri : The histogram of structure tensors {T (x)}xRi
From to color to multi/hyper-spectral images: High-dimensional
covariance matrix estimated locally in regions
Supervised segmentation: Distance learning from training images of

annotated contours
MATLAB code available.
34 / 34

Statistical Gaussian Model of Image Regions in Stochastic Watershed Segmentation

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Statistical Gaussian Model of Image Regions in Stochastic Watershed Segmentation

Uploaded by

Copyright:

Available Formats

Statistical Gaussian Model of Image Regions in Stochastic Watershed Segmentation

Statistical Gaussian Model of Image Regions

in Stochastic Watershed Segmentation

MINES ParisTech, PSL-Research University,

GSI'2015 - 2nd Conference on Geometric Science of Information

Motivation: Unsupervised segmentation of generic

Custard: Color image

Large homogenous areas, well contrasted objects as well as textured zones

Motivation: Unsupervised segmentation of generic

Custard: its color gradient image

Large homogenous areas, well contrasted objects as well as textured zones

Motivation: Unsupervised segmentation of generic

Custard: pdf of contours using stochastic watershed

Using watershed based techniques large homogeneous areas are

Motivation: Unsupervised segmentation of generic

Custard: h-dynamics watershed cut from SW pdf, h = 0.1

Using watershed based techniques large homogeneous areas are

Motivation: Unsupervised segmentation of generic

Custard: h-dynamics watershed cut from SW pdf, h = 0.3

Using watershed based techniques large homogeneous areas are

Context and goal

Context and goal

Context and goal

Goal: Take into account regional information in the probability

1 Stochastic Watershed using MonteCarlo Simulations

2 Multivariate Gaussian Model of Regions in SW

1 Stochastic Watershed using MonteCarlo Simulations

2 Multivariate Gaussian Model of Regions in SW

Regionalized Poisson points

Uniform random germs

Generate realizations of a Poisson point process with a constant

Random number of points N(D) falling in a domain D (bounded

Regionalized Poisson points

Regionalized random germs

Let us suppose that the density is not constant; but considered as

Regionalized Poisson points

Generate N random germs in an image m : E {0, 1} according to

1 Initialization: m(xi ) = 0 xi E ; P = Card(E )

Stochastic watershed paradigm

Spreading random germs as markers on the watershed segmentation.

Each piece of contour may then be assigned the number of times it

In the case of uniformly distributed random germs, large regions will

Image gradient as density for regionalization of random germs

Probability density of contours using MonteCarlo

Let {mrkn (x)}M

Probability density function of contours is computed by the kernel

where the smoothing kernel K (x) is a spatial Gaussian function of

Probability density of contours using MonteCarlo

Color image f (x) Color gradient g (x)

Probability density of contours using MonteCarlo

{WS(g , mrkn )}1nM : Watershed segmentations

Probability density of contours using MonteCarlo

Color image f (x) Color gradient g (x)

Density of contours pdf (x) Segmented with h = 0.1

Probability density of contours using MonteCarlo

Color image f (x) Color gradient g (x)

Density of contours pdf (x) Segmented with h = 0.3

1 Stochastic Watershed using MonteCarlo Simulations

2 Multivariate Gaussian Model of Regions in SW

Watershed transform Tessellation

Tessellation of E from watershed WS(x): (Finite) family of disjoint

= {Rr }1r N , with i 6= j Ri Rj =

Boundary between regions Ri and Rj (1 i, j N , i 6= j): Irregular arc

Color regions as multivariate normal distributions

Custard: Color image

Custard: its color gradient image

Custard: pdf of contours using stochastic watershed

Custard: h-dynamics watershed cut from SW pdf, h = 0.1

Custard: h-dynamics watershed cut from SW pdf, h = 0.3

Note that the rst term in the Bhattacharyya distance is related to

Bhattacharyya distance can be metrized by transforming it into to