Академический Документы
Профессиональный Документы
Культура Документы
Isabelle Bloch
http://www.tsi.enst.fr/bloch
ecommunications
A few references
J. Serra (Ed.), Image Analysis and Mathematical Morphology, Part II: Theoretical
Advances, Academic Press, London, 1988.
Introduction
Mathematical bases:
set theory
topology
geometry
algebra (lattice theory)
probabilities, random closed sets
functions
Main characteristics:
non linear
non invertible
strong properties
associated algorithms
Tools for
filtering
segmentation
measures (distances, granulometry, integral geometry, topology, stochastic
processes...)
texture analysis
shape recognition
scene interpretation
...
Example
Example
Structuring element
shape
size
origin (not necessarily in B)
examples:
Continuous:
Digital:
Binary dilation
Minkowski addition:
X Y = {x + y / x X, y Y }
Binary dilation:
D(X, B)
=
=
= {x + y / x X, y B}
(or = X B)
X B
[
x = {x Rn / Bx X 6= }
B
xX
Binary dilation
Minkowski addition:
X Y = {x + y / x X, y Y }
Binary dilation:
D(X, B)
=
=
= {x + y / x X, y B}
(or = X B)
X B
[
x = {x Rn / Bx X 6= }
B
xX
Properties of dilation:
Example of dilation
Binary erosion
E(X, B)
{x Rn / Bx X}
{x / y B, x + y X} = X B.
Binary erosion
E(X, B)
{x Rn / Bx X}
{x / y B, x + y X} = X B.
Properties of erosion:
C)
(or E(X, B) = [D(X C , B)]
anti-extensive (E(X, B) X) if O B;
increasing (X Y E(X, B) E(Y, B));
B B 0 E(X, B 0 ) E(X, B);
commutes with intersection, not with union:
E[(X Y ), B] = E(X, B) E(Y, B),
Example of erosion
Binary opening
XB = D[E(X, B), B]
Binary opening
XB = D[E(X, B), B]
Properties of opening:
anti-extensive (X XB );
increasing (X Y XB YB );
idempotent ((XB )B = XB ).
Morphological filter
B B 0 XB 0 XB ;
(Xn )n0 = (Xn0 )n = Xmax(n,n0 ) .
Example of opening
Binary closing
X B = E[D(X, B), B]
Binary closing
X B = E[D(X, B), B]
Properties of closing:
extensive (X X B );
increasing (X Y X B Y B );
idempotent ((X B )B = X B ).
Morphological filter
0
B B0 X B X B ;
(X n )n = (X n )n = X max(n,n ) ;
X B = [(X C )B ]C .
Example of closing
Digital case
choice of the digital grid (both for the image and the structuring element)
translations on the grid
same properties
sup /
inf /
Dilation of a function
by a flat structuring element
x Rn , D(f, B)(x) = sup{f (y) / y Bx }
Dilation of a function
by a flat structuring element
x Rn , D(f, B)(x) = sup{f (y) / y Bx }
extensivity if O B;
increasingness;
D(f g, B) = D(f, B) D(g, B);
D(f g, B) D(f, B) D(g, B);
iterativity property.
It holds:
D(f , B) = [D(f, B)]
Erosion of a function
Erosion of a function
Functional opening
fB = D[E(f, B), B]
Functional opening
fB = D[E(f, B), B]
anti-extensive;
increasing;
idempotent.
morphological filter
Functional closing
f B = E[D(f, B), B]
Functional closing
f B = E[D(f, B), B]
extensive;
increasing;
idempotent.
morphological filter
Structuring functions
Dilation:
D(f, g)(x) = sup{f (y) + g(y x)}
y
Erosion:
E(f, g)(x) = inf {f (y) g(y x)}
y
g(x) =
on a compact B
elsewhere
DB (x) EB (x)
An application of opening:
top-hat transform
f fB
An application of opening:
top-hat transform
f fB
An application of opening:
top-hat transform
f fB
Granulometry
Granulometry
( ) is a granulometry iff is an opening for each and the class of subsets A which
are invariant under is included in the class of subsets which are invariant under for
Granulometry
( ) is a granulometry iff is an opening for each and the class of subsets A which
are invariant under is included in the class of subsets which are invariant under for
Dilation
componentwise
maximum
Surfacic opening
(f ) =
Surfacic opening
(f ) =
(f ) =
Surfacic opening
Set theory
Lattice theory
structuring element
Topology
relations (, , ...)
adjunctions
algebraic operations
Probability theory
P (A K 6= )
random closed sets
Hit-or-miss topology
G G, G F 6= , N, n N, G Fn 6=
K K, K F = , N 0 , n N 0 , K Fn =
Hit-or-miss topology
G G, G F 6= , N, n N, G Fn 6=
K K, K F = , N 0 , n N 0 , K Fn =
semi-continuity
Semi-continuity
f :F
(F, K) 7 F K u.s.c.
(K 0 , K) 7 K 0K u.s.c.
Myopic topology
generated by:
F
= {K K, K F = , K G 6= }
KG
(F F , G G)
x0 K 0
T =
et I =
T =
(F (Rd ), )
functions of Rn in R for the ordering :
f g x Rn , f (x) g(x)
partitions
Semi-continuity of functions
u.s.c. :
t > f (x), V (x), y V (x), t > f (y)
(V (x) neighborhood of x in Rn )
l.s.c. :
t < f (x), V (x), y V (x), t < f (y)
complete lattice (T , )
Algebraic dilation:
(xi ) T , (i xi ) = i (xi )
Algebraic erosion:
(xi ) T , (i xi ) = i (xi )
Properties:
Adjunctions
(, ) adjunction on (T , ):
(x, y), (x) y x (y)
Properties:
W
openings and i i (anti-extensive filters)
V
closings and i i (extensive filters)
, , and are filters
Inv() = Inv() and Inv() = Inv()
is the smallest filter which is largest than
Mi
mi mi1 ...m2 m1
Ni
ni ni1 ...n2 n1
Ri
ri ri1 ...r2 r1
Si
si si1 ...s2 s1
Property: i j Mj Mi = Mj , Nj Ni = Nj , ...
Auto-dual filters
Operators which are independent of the local contrast, acting similarly on bright
and dark areas.
Auto-dual filters
Operators which are independent of the local contrast, acting similarly on bright
and dark areas.
O
F
centre
O
F
Closing
Opening
morphological center
Hit-or-Miss Transformation
Structuring element: T = (T1 , T2 ), with T1 T2 =
HMT:
X T = E(X, T1 ) E(X C , T2 )
Hit-or-Miss Transformation
Structuring element: T = (T1 , T2 ), with T1 T2 =
HMT:
X T = E(X, T1 ) E(X C , T2 )
Thinning (if O T1 ):
X T =X \X T
Thickening (if O T2 ):
X T =X X T
For T 0 = (T2 , T1 ):
X T = (X C T 0 )C
HMT: examples
HMT: examples
Skeleton: requirements
s (A)
>0
Characterization:
s (A) =
>0
r(A) =
[ \
>0 >0
Reconstruction:
A=
D(s , B )
>0
[E(X, Bn ) \ E(X, Bn )B ]
nN
Properties:
[E(X, Bn ) \ E(X, Bn )B ]
nN
Properties:
1
1
.
0
Properties:
perfect topology
no reconstruction
Geodesic operators
d (x,y)
X
X
d(x,y)
X convex dX = d
Properties:
r
B)
n
X]n
Digital case:
DX (Y, Br ) = [D(Y, B1 ) X]r
Reconstruction:
[D(Y, B1 ) X] = DX
(Y )
Regional maxima
X regional maximum of f if
x X, f (x) = et X = CC(f )
Computation of regional maxima:
f Df (f 1)
h-maxima (grey level dynamics):
f Df (f h)
robust maxima
X=
Xi
Influence zone of Xi in X C :
ZI(Xi ) = {x X C / d(x, Xi ) < d(x, X \ Xi )}
Skeleton by influence zones:
Skiz(X) = (
ZI(Xi ))C
Skiz(X) Skel(X C )
Skiz is not necessarily connected (even if X C is)
Skiz(X) r(X C )
Skiz(X) r(X C )
Y = i Yi
Geodesic influence zone of Yi conditionally to X :
ZIX (Yi ) = {x X, dX (x, Yi ) < dX (x, Y \ Yi )}
Geodesic skeleton by influence zones:
SKIZX (Y ) = X \
ZIX (Yi )
SKIZ(Y)
SKIZ X (Y)
Y1
Y1
Y2
Y2
X
Cortex segmentation
(PhD of Arnaud Cachia)
Cortex segmentation
(PhD of Arnaud Cachia)
Watersheds
catchment basins
watersheds
minimum
minimum
Watersheds: definition
Steepest descent:
Desc(x) = max{
f (x) f (y)
, y V (x)}
d(x, y)
Tf () =
n
X
d(xi1 , xi )Cost(xi1 , xi )
i=1
with
Desc(x)
Cost(x, y) =
Desc(y)
(Desc(x) + Desc(y))/2
Watersheds: definition
Topographic distance
Tf (x, y) = inf{Tf (), = (x0 = x, x1 , ..., xn = y)}
(equals 0 on a plateau)
Catchment basin associated to the regional minimum Mi :
BV (Mi ) = {x, j 6= i, Tf (x, Mi ) < Tf (x, Mj )}
Watersheds:
LP E(f ) = [i BV (Mi )]C
Approach by immersion
Xhmin
f hmin
Xh+1
BV
Xhmax
LP E(f )
XhCmax
Geodesic erosion
in order to impose markers