Digital Image Processing

Morphological Image Processing

Christophoros Nikou
[email protected]

Images taken from: R. Gonzalez and R. Woods. Digital Image Processing, Prentice Hall, 2008

University of Ioannina - Department of Computer Science and Engineering

2
Morphological Image Processing
and Analysis

In form and feature, face and limb,

I grew so like my brother,
That folks got taking me for him
And each for one another.

Henry Sambrooke Leigh,

Carols of Cockayne, The Twins

C. Nikou – Digital Image Processing

3 Contents
Mathematical morphology provides tools for
the representation and description of image
regions (e.g. boundary extraction, skeleton,
convex hull).
It provides techniques for pre- and post-
processing of an image (morphological
thinning, pruning, filtering).
Its principles are based on set theory.
Applications to both binary and graylevel
images.
C. Nikou – Digital Image Processing
4 Preliminaries
The four horizontal and vertical neighbours of
a pixel p are called 4-neighbours of p and are
denoted by N4(p).
The four diagonal neighbours of a pixel p are
denoted by ND(p).
Together N4(p) and ND(p) are called the 8-
neighbours of pixel p and are denoted by
N8(p).

C. Nikou – Digital Image Processing

5 Preliminaries (cont.)
Adjacency of pixels
Let V be the set of intensity values used to
define the adjacency (e.g. V={1} for binary
images).
4-adjacency. Two pixels p and q with values in
V are 4-adjacent if q is in N4(p).
8-adjacency. Two pixels p and q with values in
V are 8-adjacent if q is in N8(p).

C. Nikou – Digital Image Processing

6 Preliminaries (cont.)
Adjacency of pixels
m-adjacency (mixed adjacency). Two pixels p
and q with values in V are m-adjacent if
• q is in N4(p), or
• q is in ND(p) and the set N4(p) ∩ N4(q) has
no pixels whose values are in V .
Mixed adjacency is a modification of the 8-
adjacency. It is introduced to eliminate
ambiguities of 8-adjacency.
C. Nikou – Digital Image Processing
7 Preliminaries (cont.)
Not m-connected. They
Adjacency of pixels have a common 4-
connected neighbor.

Pixels in a 8-adjacency m-adjacency

binary image m-connected. They do not have any
common 4-connected neighbor.

The role of m-adjacency is to define a single

path between pixels. It is used in many image
analysis and processing algorithms.
C. Nikou – Digital Image Processing
8 Preliminaries (cont.)
Basic set operations.

A  B  {w | w  A OR w  B}
A  B  {w | w  A AND w  B}
A  B  {w | w  A, w  B}  A  Bc
Ac  {w | w  A}
C. Nikou – Digital Image Processing
9 Preliminaries (cont.)
The above operations assume that the
images containing the sets are binary and
involve only the pixel location.
Union and intersection are different when we
define set operations involving intensity
values:


A  B  max(a, b) | a  A, b B
z
 The elements of the sets
are gray values on the

A  B  min(a, b) | a  A, b B
same location z.

C. Nikou – Digital Image Processing

10 Preliminaries (cont.)

Set reflection: Bˆ  {w | w  b, for b  B}

Set translation by z: ( B) z  {c | c  b  z, for b  B}

C. Nikou – Digital Image Processing

11 Preliminaries (cont.)

Set reflection and translation

are employed to structuring
elements (SE).
SE Are small sets or subimages
used to examine the image
under study for properties of
interest. Note: gray colour
represents a value of
The origin must be specified. one and white colour a
Zeros are appended to SE to zero value.
give them a rectangular form.
C. Nikou – Digital Image Processing
12 Preliminaries (cont.)

The origin of the SE B visits every pixel in an image A.

It performs an operation (generally non linear) between
its elements and the pixels under it.
It is then decided if the pixel will belong to the resulting
set or not based on the results of the operation.
Zero padding is necessary (like in convolution) to
ensure that all of the elements of A are processed.

C. Nikou – Digital Image Processing

13 Preliminaries (cont.)

For example, it marks the pixel under its center

as belonging to the result if B is completely
contained in A (A2, B2).

C. Nikou – Digital Image Processing

14 Morphological Operations
• Some basic operations
– Erosion.
– Dilation.
– Opening.
– Closing.
• Applications
– Morphological filtering.
– The hit-or-miss transformation.

C. Nikou – Digital Image Processing

15 Erosion

The erosion of a set A by a SE B is defined as

A  B  {z | ( B) z  A}

The result is the set of all points z such that B

translated by z is contained in A.
Equivalently:

A  B  {z | ( B) z  A  } c

C. Nikou – Digital Image Processing

16 Erosion (cont.)

Erosion is a shrinking operation

C. Nikou – Digital Image Processing

17 Erosion (cont.)

C. Nikou – Digital Image Processing

18 Erosion (cont.)
Erosion by a square SE of varying size

Original image 11x11

15x15 45x45

C. Nikou – Digital Image Processing

19 Erosion (cont.)
Erosion can split apart joined objects

Erosion can strip away extrusions

Watch out: Erosion shrinks objects

C. Nikou – Digital Image Processing
20 Dilation

The dilation of a set A by a SE B is defined as

A  B  {z | ( Bˆ ) z  A  }

The result is the set of all points z such that the

reflected B translated overlap with A at at least
one element.

Equivalently: A  B  {z | [( Bˆ ) z  A]  A}

C. Nikou – Digital Image Processing

21 Dilation (cont.)

Dilation is a thickening operation

C. Nikou – Digital Image Processing
22 Dilation (cont.)

Dilation bridges gaps.

Contrary to low pass filtering it produces a binary image.
C. Nikou – Digital Image Processing
23 Dilation (cont.)
Dilation can repair breaks

Dilation can repair intrusions

Watch out: Dilation enlarges objects

C. Nikou – Digital Image Processing

24 Duality

Erosion and dilation are dual operations with

respect to set complementation and reflection:
( A  B)c  Ac  Bˆ
Also,
( A  B)c  Ac  Bˆ
The duality is useful when the SE is symmetric:
The erosion of an image is the dilation of its
background.

C. Nikou – Digital Image Processing

25 Compound Operations
More interesting morphological operations
can be performed by combining erosions
and dilations in order to reduce shrinking or
thickening.
The most widely used of these compound
operations are:
– Opening
– Closing

C. Nikou – Digital Image Processing

26 Opening
The opening of set A by structuring element B
is defined as
A  B  ( A  B)  B

which is an erosion of A by B followed by a

dilation of the result by B.

C. Nikou – Digital Image Processing

27 Opening (cont.)
Geometric interpretation: The boundary of
the opening is defined by points of the SE
that reach the farthest into the boundary of A
as B is “rolled” inside of this boundary.

A  B  ( B) z | ( B) z  A
C. Nikou – Digital Image Processing
28 Opening (cont.)
Notice the difference with the simple erosion:

A  B  {z | ( B) z  A} A  B  ( B) z | ( B) z  A

If B translated by z lies inside A, then the result

contains the whole set of points covered by the SE
and not only its center as it is done in the erosion.

C. Nikou – Digital Image Processing

29 Closing
The closing of set A by structuring element B
is defined as
A  B  ( A  B)  B

which is an dilation of A by B followed by an

erosion of the result by B.

C. Nikou – Digital Image Processing

30 Closing (cont.)
It has a similar geometric interpretation
except that B is rolled on the outside of the
boundary:

A  B  {w |( B) z  A  ,for all translates of ( B) z containing w}

C. Nikou – Digital Image Processing
31 Opening and Closing

Erosion: elements where the

disk can not fit are eliminated.

Opening: outward corners are

rounded.

Dilation: inward intrusions are

reduced in depth.

Closing: inward corners are

rounded.

C. Nikou – Digital Image Processing

32 Duality

Opening and closing are dual operations.

Erosion-Dilation duality Opening-Closing duality

( A  B)c  Ac  Bˆ ( A  B)c  Ac  Bˆ

( A  B)  A  Bˆ
c c
( A  B)  A  Bˆ
c c

C. Nikou – Digital Image Processing

33 Properties of Opening and Closing

Opening: A B  A
C  D C  B  D B
( A  B)  B  A  B

A  A B
Closing:
C  D C  B  D B
( A  B)  B  A  B
The last properties, in each case, indicate that
multiple openings or closings have no effect after
the first application of the operator
C. Nikou – Digital Image Processing
34 Morphological Filtering Example

The image contains noise:

•Light elements on dark
background.
•Dark elements on the light
components of the
fingerprint.

Objective: Eliminate noise while distorting the

image as little as possible.
We will apply an opening followed by closing.

C. Nikou – Digital Image Processing

35
Morphological Filtering Example
(cont.)

A A B
Background noise completely removed (noise
components smaller than the SE).
The size of the dark noise elements in the fingerprint
structure increased (inner dark structures).
C. Nikou – Digital Image Processing
36
Morphological Filtering Example
(cont.)

A B ( A  B)  B  A  B

The dilation reduced the size of the inner noise or

eliminated it completely.
However, new gaps were created by the opening
between the fingerprint ridges.
C. Nikou – Digital Image Processing
37
Morphological Filtering Example
(cont.)

A B A B  B

The dilation reduces the new gaps between the

ridges but it also thickens the ridges.

C. Nikou – Digital Image Processing

38
Morphological Filtering Example
(cont.)

A B  B [ A  B  B]  B  ( A  B)  B

The final erosion (resulting to a closing of the

opened image) makes the ridges thinner.

C. Nikou – Digital Image Processing

39
Morphological Filtering Example
(cont.)

A ( A  B)  B

The final result is clean of noise but some ridges were

not fully repaired.
We should impose conditions for maintaining the
connectivity (we will see a more advanced algorithm).
C. Nikou – Digital Image Processing
40 The Hit-or-Miss Transformation
Basic tool for shape detection.
Erosion of A by B: the set of all locations of the
origin of B that B is completely contained in A.
Alternatively, it is the set of all locations that B
found a match (hit) in A.

C. Nikou – Digital Image Processing

41
The Hit-or-Miss Transformation
(cont.)
There are many possible locations for the shape
we search (the SE!). If we are looking for disjoint
(disconnected) shapes it is natural to assume a
background for it.
Therefore, we seek to match B in A and
simultaneously we seek to match the background
of B in Ac.
Mathematically, the hit-or-miss transformation is:

A * B  ( A  B)  ( A  Bb ) c

C. Nikou – Digital Image Processing

42
The Hit-or-Miss Transformation
(cont.)
We seek to locate the shape
D in the image A.

We define a thin background

W for the shape.

We take the intersection of

the two results

A * D  ( A  D)   Ac  (W  D) 

C. Nikou – Digital Image Processing

43 Morphological Algorithms
Using these morphological operations we may
extract image components for shape representation:
• Shape boundaries.
• Region filling.
• Connected components
• Convex hull.
• Shape thinning and thickening.
• Skeletons.
We may also accomplish a morphological image
reconstruction.
C. Nikou – Digital Image Processing
44 Boundary Extraction
The boundary of a set A, denoted by β(A),
may be obtained by:
 ( A)  A  ( A  B)

C. Nikou – Digital Image Processing

45 Boundary Extraction (cont.)
The boundary is one pixel thick due to the
3x3 SE. Other SE would result in thicker
boundaries.

Original Image Extracted Boundary

C. Nikou – Digital Image Processing
46 Region Filling
Given a pixel inside a boundary, region
filling attempts to fill the area surrounded by
that boundary with 1s.

Given a point inside

here, can we fill the
whole circle?

C. Nikou – Digital Image Processing

47 Region Filling (cont.)
Form a set X0 with zeros everywhere except
at the seed point of the region.
Then,
X k  ( X k 1  B)  Ac , k  1, 2,3,...

Where B is a 3x3 cross-shaped SE.

The algorithm terminates when Xk=Xk-1.
The set union of Xk and A contains all the
filled holes and their boundaries.
C. Nikou – Digital Image Processing
48 Region Filling (cont.)
This is a first example where the morphological
operation (dilation) is conditioned.

The intersection of the result with Ac limits the

result inside the region of interest.

X k  ( X k 1  B)  Ac , k  1, 2,3,...

C. Nikou – Digital Image Processing

49 Region Filling (cont.)

C. Nikou – Digital Image Processing

50 Region Filling (cont.)

Original Image One Region All Regions

Filled Filled

C. Nikou – Digital Image Processing

51 Extraction of connected components
• Given a pixel on a connected component, find the rest of
the pixels of that component.
• The algorithm may be applied to many connected
components provided we know a pixel on each one of
them.
• Disadvantage:
• we have to provide a pixel on the connected
component.
• There are more sophisticated algorithms that detect the
number of components without manual interaction. The
purpose here is to demonstrate the flexibility of
mathematical morphology.

C. Nikou – Digital Image Processing

52
Extraction of connected components
(cont.)
Form a set X0 with zeros everywhere except at
the seed point of the connected components.
Then,
X k  ( X k 1  B)  A, k  1, 2,3,...

Where B is a 3x3 square-shaped SE.

The algorithm terminates when Xk=Xk-1.
Xk contains all the connected components.

C. Nikou – Digital Image Processing

53
Extraction of connected components
(cont.)
Note the similarity with region filling. The only
difference is the use of A instead of Ac.
This is not surprising as we search for foreground
objects.

C. Nikou – Digital Image Processing

54
Extraction of connected components
(cont.)
Image of chicken filet
containing bone fragments

Result of simple thresholding

Image erosion to retain only

objects of significant size.

15 connected components detected with four of them being

significant in size. This is an indication to remove the chicken
filet from packaging.
C. Nikou – Digital Image Processing
55 Convex Hull
A set A is convex if the straight line segment joining
any two points in A lies entirely within A.
The convex hull H of an arbitrary set S the smallest
convex set containing S.
The difference H-S is called convex deficiency.
The convex hull and the convex deficiency are useful
quantities to characterize shapes.

We present here a morphological algorithm to obtain

the convex hull C(A) of a shape A.

C. Nikou – Digital Image Processing

56 Convex Hull (cont.)
The procedure requires four SE Bi, i=1, 2, 3, 4, and
implements the following equation:
X ki  ( X k 1 * Bi )  A, i  1, 2,3, 4 and k  1, 2,3,...
with X 0i  A

with i referring to the SE and k to iteration.

Then, letting
Di  X ki
The convex hull of A is 4
C ( A)   Di
i 1
C. Nikou – Digital Image Processing
57 Convex Hull (cont.)
X ki  ( X k 1 * Bi )  A i  1, 2,3, 4 and k  1, 2,3,... with X 0i  A
4
Di  X ki , C ( A)   Di
i 1

The method consists of iteratively applying the hit-or miss

transform to A with B1.
When no changes occur we perform the union with A and
save the result to D1.
The procedure is then continued with B2 (applied to A) and so
on.
The union of the results is the convex hull of A.
Note that a simple implementation of the hit or miss is applied
(no background match is required).
C. Nikou – Digital Image Processing
58 Convex Hull (cont.)

The hit-or-miss transform tries to find (“hit”) these structures

in the image.
The SE has points with “don’t care” condition. For all the SE,
a match is found in the image when these conditions hold:
• the central pixel in the 3x3 region in the image is 0.
• the three shaded pixels under the mask are 1s.
The remaining pixels do not matter.

C. Nikou – Digital Image Processing

59 Convex Hull (cont.)

- Problem.
The result is convex
but greater than the
true convex hull.

C. Nikou – Digital Image Processing

60 Convex Hull (cont.)
Solution: limit the growth so that it does not extend past the
horizontal and vertical limits of the original set of points.

Original image Initial convex hull Refined convex hull

More complex boundaries have been imposed to images with

finer details in their structure (e.g. the maximum of the
horizontal vertical and diagonal dimensions could be used).
C. Nikou – Digital Image Processing
61 Thinning
The thinning of a set A, by a SE B may be
defined in terms of the hit-or-miss transform:
A  B  A  ( A * B)  A  ( A * B) c

No background match is required and the hit-

or-miss part is reduced to simple erosion.
A more advanced expression is based on a
sequence of SE {B }  {B1
, B 2
, B 3
,..., B n
},
where each Bi is a rotated version of Bi-1.
C. Nikou – Digital Image Processing
62 Thinning (cont.)
The thinning by a sequence of SE is defined by:
A {B}  ((...(( A  B )  B )...)  B )
1 2 n

The process is to thin A by one pass by B1,

then thin the result with one pass of B2, and
so on, until we employ Bn.
The entire process is repeated until no further
changes occur. Each individual thinning is
performed by:
A  B  A  ( A * B)
C. Nikou – Digital Image Processing
63 Thinning (cont.)

• No change between the

result of B7 and B8 at the
first pass.
• No change between the
results of B1, B2, B3, B4 at
the second pass.
• No change occurs after
the second pass by B6 .

•The final result is

converted to
m-connectivity to have a
one pixel thick structure.
C. Nikou – Digital Image Processing
64 Thickening
Thickening is a morphological dual of thinning:
A  B  A  ( A * B)
The SE have the same form as the ones used
for thinning with the 1s and 0s interchanged.
It may also be defined by a sequence of
operations:

A  {B}  ((...(( A  B1 )  B 2 )...)  Bn )

C. Nikou – Digital Image Processing

65 Thickening (cont.)

In practice, a separate algorithm is seldom used for

thickening.
The usual process is to thin the background of the
set in question and then take the complement of the
result.
The advantage is that the thinned background forms
a boundary for the thickening process. Direct
implementation of thickening has no stopping
criterion.
A disadvantage is that there may be isolated points
needing post-processing.
C. Nikou – Digital Image Processing
66 Thickening (cont.)

Original set A Ac

Thickened set
Thinning of Ac obtained by
complementing the
result of thinning.

Elimination of disconnected points.

C. Nikou – Digital Image Processing

67 Skeletons
The notion of a skeleton S(A) of a
set A, intuitively, has the following
properties:
• If z is a point belonging to S(A)
and (D)z is the largest disk
centered at z and contained in A:
one cannot find a larger disk (not
necessarily centered at z)
containing (D)z and included in A.
• (D)z is then called maximum disk.
• The maximum disk touches the
boundary of A at two or more
different points.
C. Nikou – Digital Image Processing
68 Skeletons (cont.)

It may be shown that a definition of the skeleton

may be given in terms of erosions and openings:
K
S ( A)   Sk ( A), with Sk ( A)  ( A  kB)  ( A  kB)  B
k 0

with ( A  kB)  ((...( A  B)  B)  ...)  B)


k successive erosions

K is the last iterative step before A erodes to an

empty set:
K  max{k | A  kb  }
C. Nikou – Digital Image Processing
69 Skeletons (cont.)

The previous formulation allows the iterative

reconstruction of A from the sets forming its
skeleton by:
K
A    Sk ( A)  kB  ,
k 0

with Sk ( A)  B  ((...( Sk ( A)  B)  B)  ...)  B)


k successive dilations of the set Sk ( A )

C. Nikou – Digital Image Processing

70 Skeletons (cont.)
The skeleton is
• thicker than essential.
• disconnected.

The morphological
formulation does not
guarantee connectivity.

More assumptions are

needed to obtain a
maximally thin and
connected skeleton.
C. Nikou – Digital Image Processing
71 Morphological Reconstruction
• The morphological algorithms discussed so
far involve an image and a SE.
• Morphological reconstruction involves two
images and a SE.
– The marker image containing the starting
point of the transformation.
– The mask image, which constraints the
transformation.
– The SE is used to define connectivity.

C. Nikou – Digital Image Processing

72 Morphological Reconstruction (cont.)
The geodesic dilation of size 1 of a marker image
F by a SE B, with respect to a mask image G is
defined by: (1)
DG ( F )  ( F  B)  G
Similarly, the geodesic dilation of size n is defined by:

DG( n ) ( F )  DG(1)  DG( n1) ( F )  with DG(0) ( F )  F

The intersection operator at each step guarantees

that the growth (dilation) of marker F is limited by
the mask G.
C. Nikou – Digital Image Processing
73 Morphological Reconstruction (cont.)

Geodesic dilation of size 1.

The result will not contain elements not belonging to
the mask G.
C. Nikou – Digital Image Processing
74 Morphological Reconstruction (cont.)
The geodesic erosion of size 1 of a marker image
F by a SE B, with respect to a mask image G is
defined by: (1)
EG ( F )  ( F  B)  G
Similarly, the geodesic erosion of size n is defined by:

EG( n ) ( F )  EG(1)  EG( n1) ( F )  with EG(0) ( F )  F

The union operator guarantees that the geodesic

erosion of marker F remains greater than or equal
to the mask G.
C. Nikou – Digital Image Processing
75 Morphological Reconstruction (cont.)

Geodesic erosion of size 1.

The result will at least contain the mask G.
C. Nikou – Digital Image Processing
76 Morphological Reconstruction (cont.)
• The geodesic dilation and erosion are duals
with respect to set complementation.
• They always converge after a finite number
of steps:
– Geodesic dilation: propagation of the marker is
constrained by the mask image.
– Geodesic erosion: shrinking of the marker is
constrained by the mask.

C. Nikou – Digital Image Processing

77 Morphological Reconstruction (cont.)
The morphological reconstruction by dilation of
mask image G from a marker image F is defined as
the geodesic dilation of F with respect to G, iterated
until stability s achieved:

RGD ( F )  DG( k ) ( F )
with k such that:

DG( k ) ( F )  DG( k 1) ( F )

C. Nikou – Digital Image Processing

78 Morphological Reconstruction (cont.)
Example of morphological
reconstruction by dilation.
The mask, marker, SE and the
first step of the algorithm are
from the example of geodesic
dilation.

C. Nikou – Digital Image Processing

79 Morphological Reconstruction (cont.)
The morphological reconstruction by erosion of
mask image G from a marker image F is defined as
the geodesic erosion of F with respect to G, iterated
until stability is achieved:

RGE ( F )  EG( k ) ( F )
with k such that:

( k 1)
E (F )  E
(k )
G G (F )
The example is left as an exercise!
C. Nikou – Digital Image Processing
80
Applications
Opening by Reconstruction
In morphological opening, erosion removes small
objects and dilation attempts to restore the shape of
the objects that remain without the small objects.

This is not accurate as it depends on the similarity

between the shapes to be removed and the SE.

Opening by reconstruction restores exactly the

shapes of the objects that remain after erosion.

C. Nikou – Digital Image Processing

81
Applications
Opening by Reconstruction (cont.)
The opening by reconstruction of size n of an image
F is defined as the reconstruction by dilation of F
from the erosion of size n of F:

OR( n ) ( F )  RFD ( F  nB)

The image F is used as the mask and the n

erosions of F by B are used as the initial marker
image.

C. Nikou – Digital Image Processing

82
Applications
Opening by Reconstruction (cont.)
We are interested in extracting characters with long vertical
strokes (~50 pixels high).
Original image One erosion by a 51x1 SE

Opening Opening by reconstruction

C. Nikou – Digital Image Processing
83
Applications
Region filling
No starting point is needed to be provided.
The original image I(x,y) is used as a mask.
The marker image is
1  I ( x, y ) if ( x, y) is on the border of I
F ( x, y )  
 0 otherwise
Only dark pixels of I(x,y) touching the border
have a value of 1 in F(x,y).

The binary image with all regions (holes) filled is

given by: H   RI c ( F ) 
D c

C. Nikou – Digital Image Processing

84
Applications
Region filling (cont.)

We wish to fill the hole of the image I.

The complement builds a wall around the hole.
The marker image F is one at the border except
from border pixels of the original image.

C. Nikou – Digital Image Processing

85
Applications
Region filling (cont.)

The dilation of the marker F starts from the

border and grows inward.
The complement is used as an AND mask: it
protects all foreground pixels (including the wall)
from changing during the iterations.
The last operation provides only the hole points.
C. Nikou – Digital Image Processing
86
Applications
Region filling (cont.)
Original image Complement of original image

Marker image (1s almost everywhere Result of hole filling

on the border apart of some points
on the right border) C. Nikou – Digital Image Processing
87
Applications
Border Clearing
The extraction of objects from an image is a
fundamental task in automated image analysis.
An algorithm for removing objects that touch
(are connected) to the image border is useful
because
• only complete objects remain for further
processing.
• it is a signal that partial objects remain in the
field of view.

C. Nikou – Digital Image Processing

88
Applications
Border Clearing (cont.)
The original image is used as a mask.
The marker image is
 I ( x, y ) if ( x, y) is on the border of I
F ( x, y )  
 0 otherwise
The border clearing algorithm first computes the
D
morphological reconstruction I ( F ),
R
which simply extracts the objects touching the
border and then obtains the new image with no
objects touching the borders I  R D
I ( F ).

C. Nikou – Digital Image Processing

89
Applications
Border Clearing (cont.)

Original image I

Reconstruction by dilation of Reconstructed image I  RID ( F )

the 1s touching the border
C. Nikou – Digital Image Processing
90 Gray-Scale Morphology
• The image f (x,y) and the SE
b(x,y) take real or integer
values.
• SE may be flat or nonflat.
• Due to a number of difficulties
(result interpretation, erosion is
not bounded by the image,
etc.) symmetrical flat SE with
origin at the center are
employed.
• Set reflection: bˆ( x, y)  b( x, y)

C. Nikou – Digital Image Processing

91 Gray-Scale Erosion
The erosion of image f by a SE b at any
location (x,y) is defined as the minimum value
of the image in the region coincident with b
when the origin of b is at (x,y):

 f  b ( x, y)  (min{
s ,t )b
f ( x  s, y  t )}

In practice, we place the center of the SE at

every pixel and select the minimum value of
the image under the window of the SE.
C. Nikou – Digital Image Processing
92 Gray-Scale Dilation
The dilation of image f by a SE b at any
location (x,y) is defined as the maximum value
of the image in the window outlined by b:

 f  b ( x, y)  max{
( s ,t )b
f ( x  s, y  t )}

The SE is reflected as in the binary case.

C. Nikou – Digital Image Processing

93 Gray-Scale Erosion and Dilation

Original image Erosion by a flat disk SE of Dilation by a flat disk SE of

radius 2: radius 2:
Darker background, Lighter background,
small bright dots reduced, small dark dots reduced,
dark features grew. light features grew.

C. Nikou – Digital Image Processing

94 Gray-Scale Morphology (nonflat SE)
The erosion of image f by a nonflat SE bN is
defined as:
 f  bN  ( x, y)  min { f ( x  s, y  t )  bN (s, t )}
( s ,t )bN

The dilation of image f by a nonflat SE bN is

defined as:
 f  bN  ( x, y)  max { f ( x  s, y  t )  bN (s, t )}
( s ,t )bN

When the SE is flat the equations reduce to

the previous formulas up to a constant.
C. Nikou – Digital Image Processing
95 Duality

As in the binary case, erosion and dilation are

dual operations with respect to function
complementation and reflection:
( f  b) ( x, y)  ( f  bˆ)( x, y)
c c

Similarly,
( f  b)c ( x, y)  ( f c  bˆ)( x, y)

In what follows, we omit the coordinates for

simplicity.
C. Nikou – Digital Image Processing
96 Gray-Scale Opening and Closing
The opening of image f by SE b is:
f  b  ( f  b)  b

The closing of image f by SE b is:

f  b  ( f  b)  b

They are also duals with respect to function

complementation and reflection:
c c ˆ
( f  b)  f  b ( f  b)  f  bˆ
c c

C. Nikou – Digital Image Processing

97
Gray-Scale Opening and Closing
(cont.)

Geometric interpretation of opening:

It is the highest value reached by any part of the SE as it
pushes up against the under-surface of the image (up to the
point it fits completely).
It removes small bright details.
Notice the similarity with binary opening (smooths outward
corners from the inside).
C. Nikou – Digital Image Processing
98
Gray-Scale Opening and Closing
(cont.)

Geometric interpretation of closing:

It is the lowest value reached by any part of the SE as it
pushes down against the upper side of the image intensity
curve.
It highlights small dark regions of the image.
Notice the similarity with binary closing (smooths inward
corners from the outside).
C. Nikou – Digital Image Processing
99
Gray-Scale Opening and Closing
(cont.)
Properties of opening:
(1) f  b  f
(2) If f1  f 2 , then f1  b  f 2  b
(3) ( f  b)  b  f  b
The first property indicates that:
• the domain of the opening is a subset of the
domain of f and
• [ f  b ]( x, y)  f ( x, y)
C. Nikou – Digital Image Processing
100
Gray-Scale Opening and Closing
(cont.)
Properties of closing:
(1) f  f  b
(2) If f1  f 2 , then f1  b  f 2  b
(3) ( f  b)  b  f  b

The first property indicates that:

• the domain of f is a subset of the domain of the
closing and
• f ( x, y)  [ f  b ]( x, y)
C. Nikou – Digital Image Processing
101
Gray-Scale Opening and Closing
(cont.)

Original image Opening by a flat disk SE Closing by a flat disk SE of

of radius 3: radius 5:
Intensities of bright Intensities of dark features
features decreased, increased,
Effects on background are Effects on background are
negligible (as opposed to negligible (as opposed to
erosion). dilation).
C. Nikou – Digital Image Processing
102
Gray-Scale
Morphological Algorithms
• Morphological smoothing
• Morphological gradient
• Top-hat transformation
• Bottom-hat transformation
• Granulometry
• Textural segmentation

C. Nikou – Digital Image Processing

103 Morphological Smoothing
Opening suppresses light details smaller than the SE
and closing suppresses (makes lighter) dark details
smaller than the SE.
They are used in combination as morphological filters
to eliminate undesired structures.

Cygnus Loop supernova.

We wish to extract the
central light region.
C. Nikou – Digital Image Processing
104 Morphological Smoothing (cont.)
Opening followed by closing with disk SE of varying size

Radius 1
Original image

Radius 3 Radius 5

C. Nikou – Digital Image Processing

105 Morphological Gradient
The difference of the dilation and the erosion of an
image emphasizes the boundaries between regions:

g  ( f  b)  ( f  b)

Homogeneous areas are not affected and the

subtraction provides a derivative-like effect.
The net result is an image with flat regions
suppressed and edges enhanced.

C. Nikou – Digital Image Processing

106 Morphological Gradient (cont.)

Original Dilation
image

Erosion Difference

C. Nikou – Digital Image Processing

107
Top-hat and Bottom-hat
Transformations
• Opening suppresses light details smaller than
the SE.
• Closing suppresses dark details smaller than
the SE.
• Choosing an appropriate SE eliminates
image details where the SE does not fit.
• Subtracting the outputs of opening or closing
from the original image provides the removed
components.
C. Nikou – Digital Image Processing
108
Top-hat and Bottom-hat
Transformations (cont.)
Because the results look like the top or bottom
of a hat these algorithms are called top-hat
and bottom-hat transformations:
That ( f )  f  ( f  b) Light details remain

Bhat ( f )  ( f  b)  f Dark details remain

An important application is the correction of

nonuniform illumination which is a pre-
segmentation step.
C. Nikou – Digital Image Processing
109
Top-hat and Bottom-hat
Transformations (cont.)

Original image Thresholded image

(Otsu’s method)

Opened image Top-hat reduced Thresholded top-hat

(disk SE r=40) nonuniformity
Does not fit to grains
and eliminates them
C. Nikou – Digital Image Processing
110 Granulometry
• Determination of the size distribution of particles in
an image. Particles are seldom separated.
• The method described here measures their
distribution indirectly.
• It applies openings with SE of increasing size.
• Each opening suppresses bright features where the
SE does not fit.
• For each opening the sum of pixel values is
computed and a histogram of the size of the SE vs
the remaining pixel intensities is drawn.

C. Nikou – Digital Image Processing

111 Granulometry (cont.)
Image of Smoothed Opening by SE
wooden plugs image of radius 10

Opening by SE Opening by SE Opening by SE

of radius 20. of radius 25 of radius 30
Small dowels Large dowels
disappeared. disappeared.
C. Nikou – Digital Image Processing
112 Granulometry (cont.)
• Histogram of the differences of the total image
intensities between successive openings as a
function of the radius of the SE.
• There are two peaks indicating two dominant
particle sizes (of radii 19 and 27 ).

C. Nikou – Digital Image Processing

113 Textural segmentation

The objective is to find a

boundary between the
large and the small blobs
(texture segmentation).
The objects of interest
are darker than the
background.
A closing with a SE
larger than the blobs
would eliminate them.
C. Nikou – Digital Image Processing
114 Textural segmentation (cont.)

• Closing with a SE of radius 30.

• The small blobs disappeared as they have a
radius of approximately 25 pixels.

C. Nikou – Digital Image Processing

115 Textural segmentation (cont.)
The background is lighter than the large blobs.
If we open the image with a SE larger than the
distance between the large blobs then the
blobs would disappear and the background
would be dominant.

C. Nikou – Digital Image Processing

116 Textural segmentation (cont.)
Opening with a SE of radius 60.
The lighter background was suppressed to the
level of the blobs.

C. Nikou – Digital Image Processing

117 Textural segmentation (cont.)
A morphological gradient with a 3x3 SE gives
the boundary between the two regions which is
superimposed on the initial image.

C. Nikou – Digital Image Processing

118 Textural segmentation (cont.)

Closing with a SE
Original image of radius 30 (small
blobs are removed)

Opening with a SE Morphological

of radius 60 (large gradient
blobs flooded the superimposed onto
background) the original image

C. Nikou – Digital Image Processing

119
Gray-Scale Morphological
Reconstruction
The geodesic dilation of size 1 of a marker image f
by a SE b, with respect to a mask image g is defined
by:
Dg ( f )  ( f  b)  g
(1)

where  is the point-wise minimum operator.

This equation indicates that the geodesic dilation of

size 1 is obtained by first computing the dilation of f
by b and then selecting the minimum between the
result and g at every point (x,y).

C. Nikou – Digital Image Processing

120
Gray-Scale Morphological
Reconstruction (cont.)
The geodesic dilation of size n of a marker image f
by a SE b, with respect to a mask image g is defined
by:

Dg( n ) ( f )  Dg(1)  Dg(n1) ( f )

with Dg(0) ( f )  f

C. Nikou – Digital Image Processing

121
Gray-Scale Morphological
Reconstruction (cont.)
The geodesic erosion of size 1 of a marker image
f by a SE b, with respect to a mask image g is
defined by:
Eg ( f )  ( f  b)  g
(1)

where  is the point-wise maximum operator.

This equation indicates that the geodesic erosion of

size 1 is obtained by first computing the erosion of f
by b and then selecting the maximum between the
result and g at every point (x,y).

C. Nikou – Digital Image Processing

122
Gray-Scale Morphological
Reconstruction (cont.)
The geodesic erosion of size n of a marker image f
by a SE b, with respect to a mask image g is defined
by:

Eg( n ) ( f )  Eg(1)  Eg(n1) ( f )

with Eg(0) ( f )  f

C. Nikou – Digital Image Processing

123
Gray-Scale Morphological
Reconstruction (cont.)
The morphological reconstruction by dilation of
gray scale image g from a marker image f is defined
as the geodesic dilation of f with respect to g,
iterated until stability is achieved:

R (F )  D (F )
D
g
(k )
g

with k such that:

( k 1)
D (F )  D
(k )
g g (F )

C. Nikou – Digital Image Processing

124
Gray-Scale Morphological
Reconstruction (cont.)
The morphological reconstruction by erosion of
gray scale image g from a marker image f is defined
as the geodesic erosion of f with respect to g,
iterated until stability is achieved:

R (F )  E (F )
D
g
(k )
g

with k such that:

( k 1)
E (F )  E
(k )
g g (F )

C. Nikou – Digital Image Processing

125
Gray-Scale Morphological
Reconstruction (cont.)
The opening by reconstruction of size n of an
image f is defined as the reconstruction by dilation
of f from the erosion of size n of f:

OR( n ) ( f )  R Df ( f  nB)

The image f is used as the mask and the n erosions

of f by b are used as the initial marker image.
Recall that the objective is to preserve the shape of
the image components that remain after erosion.
C. Nikou – Digital Image Processing
126
Gray-Scale Morphological
Reconstruction (cont.)
• The image has a size of
1134x1360.
• The target is to leave only
the text on a flat
background of constant
intensity
• In other words, we want to
remove the relief effect of
the keys.

C. Nikou – Digital Image Processing

127
Gray-Scale Morphological
Reconstruction (cont.)
At first we suppress the
horizontal reflections on the
top of the keys.
The reflections are wider than
any single character.
An opening by reconstruction
using a long horizontal line
SE (1x71) in the erosion
operation provides the keys
and their reflections.

C. Nikou – Digital Image Processing

128
Gray-Scale Morphological
Reconstruction (cont.)
A standard opening would not be sufficient as the
background would not have been as uniform ( e.g.
look at the regions between the keys horizontally).

Opening by reconstruction Standard opening

C. Nikou – Digital Image Processing

129
Gray-Scale Morphological
Reconstruction (cont.)
Then, subtracting this result from the original image
(top-hat by reconstruction) eliminates the reflections.

 

Original Image Opening by Top-hat by

reconstruction reconstruction

C. Nikou – Digital Image Processing

130
Gray-Scale Morphological
Reconstruction (cont.)
A standard top-hat transformation would not be
sufficient as the background is not as uniform as in
the top-hat by reconstruction operation.

Top-hat by reconstruction Standard top-hat

C. Nikou – Digital Image Processing

131
Gray-Scale Morphological
Reconstruction (cont.)
We now try to suppress the vertical
reflections on the sides of the keys.

An opening by reconstruction using a

horizontal line SE (1x11) in the
erosion operation provides the keys
and their reflections (after subtracting
the result from the previous image).

Notice however that vertically

oriented characters are eliminated
(The “I” in the ”SIN” key)
C. Nikou – Digital Image Processing
132
Gray-Scale Morphological
Reconstruction (cont.)
How can we restore the suppressed character?
A dilation is not sufficient as the area of the
suppressed character is now occupied by the
expansion of its neighbors.

Dilation
(SE 1x21)

C. Nikou – Digital Image Processing

133
Gray-Scale Morphological
Reconstruction (cont.)
We form an image by taking the point-wise minimum
between the top-hat by reconstruction image and the
dilated image:

 

Top-hat by Dilated image The result is close to

reconstruction our objective but the
“I” is still missing
C. Nikou – Digital Image Processing
134
Gray-Scale Morphological
Reconstruction (cont.)
Using the last image
as a marker and the
dilated image as a
mask we perform a
gray-scale
Marker Mask
reconstruction by
dilation and we obtain
the desired result.
Result

C. Nikou – Digital Image Processing