Image Segmentation

Dr. Dipti Patra

Department of Electrical Engineering

National Institute of Technology
Rourkela
 Image Segmentation
 Partitioning of an image into different constituent
regions or objects depending on their properties.

 Isolates the object of interest in an application.

Why segmentation is difficult ?

It can be difficult for many reasons:

•Non- uniform illumination
•No control of the environment
•Inadequate model of the object of interest
•Noise
 Applications
 Automated target detection, ex: infrared image
segmentation in military application.

 Automated Industrial Application.

 Remote sensing: Satellite image segmentation for

detecting the crop area, vegetation area, urban area etc.

 Bio-medical: MRI, Ultrasound, CT image

segmentation for diagnosis of diseases and disorders.
 Segmentation

Discontinuity Similarity of pixels

Variation in pixel All neighbouring pixels

intensities. Have similarity in pixel
Appropriate grouping of intensities are grouped
edge pixels. into a region.
Discontinuity based Image Segmentation

Edge detection Techniques are

applied.

A set of edge pixels are obtained

Some edge pixels Some edge

correspond to pixels
boundary of objects. are isolated.
Edges are significant local changes of intensity
in an image.

• Geometric events
– surface orientation (boundary)
discontinuities
– depth discontinuities
– color and texture discontinuities
• Non-geometric events
– illumination changes
– shadows
– inter-reflections
 Edge Descriptors
Edge direction: perpendicular to the direction
of maximum intensity change (i.e., edge normal)
Edge strength: related to the local image
contrast along the normal.
Edge position: the image position at which the
edge is located.
Two edge pixels are linked up based upon two
criteria

 Edge strength gradient

 Edge angle direction of the gradient

Boundary detection is done by examining the

links of edge pixels.
 Modeling Intensity Changes

• Step edge: the image intensity abruptly

changes from one value on one side of the
discontinuity to a different value on the
opposite side.
 Modeling Intensity Changes
(cont’d)
• Ramp edge: a step edge where the
intensity change is not instantaneous but
occur over a finite distance.
 Modeling Intensity Changes
(cont’d)
• Ridge edge: the image intensity abruptly
changes value but then returns to the
starting value within some short distance
(i.e., usually generated by lines).
 Modeling Intensity Changes
(cont’d)
• Roof edge: a ridge edge where the
intensity change is not instantaneous but
occur over a finite distance (i.e., usually
generated by the intersection of two
surfaces).
 Main Steps in Edge Detection

Smoothing: suppress as much noise as

possible, without destroying true edges.

Enhancement: apply differentiation to

enhance the quality of edges (i.e.,
sharpening).
 Main Steps in Edge Detection
Thresholding: determine which edge
pixels should be discarded as noise and
which should be retained (i.e., threshold
edge magnitude).

Localization: determine the exact edge

location.
sub-pixel resolution might be required for
some applications to estimate the location of
an edge to better than the spacing between
pixels.
 Edge Detection Using Derivatives

• Often, points that lie on an edge

are detected by:
(1) Detecting the local maxima or
minima of the first derivative.

(2) Detecting the zero-crossings

1st derivative
of the second derivative.

2nd derivative
The first derivate of an image can be computed using
the gradient:

Gradient of an image f(x,y) at location (x,y)

 f 
Gx   x 
f      f 
G y   
 y 
Magnitude of gradient vector
f  Gx2  Gy 
2 1/ 2

1  G y

Direction of the gradient vector  ( x, y )  tan  
 Gx 
 f f
| || |
x y

18
An edge pixel at (x',y') is similar in magnitude
to the pixel at (x,y) if

f ( x, y)  f ( x, y)  E

where E is a non-negative threshold.

An edge pixel at (x',y') has an angle similar

to the pixel at (x,y) if

 ( x, y)   ( x, y)  A
 Approximating Gradient

• We can implement and using the following masks:

(x+1/2,y)

good approximation
(x,y+1/2)
at (x+1/2,y) *
*

good approximation
at (x,y+1/2)
 Approximating Gradient

• A different approximation of the

gradient: good approximation
(x+1/2,y+1/2)

• and can be implemented using

the following masks:
 Common Operators
• Gradient operator

g (m, n)  g12 (m, n)  g 22 (m, n)

Examples: 1. Roberts operator

 0 1 1 0 
  1 0 0 1
   
g1 g2

22
 Common Operators
2. Prewitt operator 3. Sobel operator
 1 0 1   1 0 1
 1 0 1   2 0 2
vertical    
 1 0 1   1 0 1 

 1  1  1  1  2  1

horizontal 0 0 0  0 0 0 
   
 1 1 1   1 2 1 

23
 Examples

original image horizontal edge vertical edge

Prewitt operator
24
Effect of Thresholding Parameters

small threshold large

25
 Compass Operators
1 1 0 1 1 1 0 1 1
1 0  1 0 0 0  1 0 1
   
0 1  1  1  1  1  1 1 0
1 0  1  1 0 1
1 0  1  1 0 1
 
1 0  1  1 0 1
0 1  1  1  1  1  1 1 0
1 0  1 0 0 0  1 0 1
   
1 1 0   1 1 1   0 1 1
g (m, n)  max{| g k (m, n) |}
k 26
 Examples

Compass operator
27
Edge Detection Steps Using Gradient

(i.e., sqrt is costly!)

Prewitt Operator

29
Example
 Magnitude

After thresholding
 Practical Issues
• Noise suppression-localization tradeoff.
– Smoothing depends on mask size (e.g.,
depends on σ for Gaussian filters).
– Larger mask sizes reduce noise, but
worsen localization (i.e., add uncertainty to
the location of the edge) and vice versa.
smaller mask larger mask
Effect of Smoothing on
Derivates

Where is the edge??

Effect of Smoothing on Derivatives
(cont’d)
Combine Smoothing with
Differentiation
(i.e., saves one operation)
 Practical Issues (cont’d)

• Choice of threshold.
gradient magnitude

low threshold high threshold

 Practical Issues (cont’d)

• Edge thinning and linking.

 Criteria for Optimal Edge
Detection
• Good detection
– Minimize the probability of false positives (i.e.,
spurious edges).
– Minimize the probability of false negatives (i.e.,
missing real edges).

• Good localization
– Detected edges must be as close as possible to the
true edges.

• Single response
– Minimize the number of local maxima around the true
edge.
 Canny edge detector
• Canny has shown that the first derivative of the
Gaussian closely approximates the operator that
optimizes the product of signal-to-noise ratio and
localization.
(i.e., analysis based on "step-edges" corrupted by "Gaussian noise“)

J. Canny, A Computational Approach To Edge Detection, IEEE Trans. Pattern Analysis

and Machine Intelligence, 8:679-714, 1986.
 Local edge linking approach: A local
neighbourhood is used. (Canny Edge Detector)

 Global edge linking approach: each edge

pixel is linked with other edge pixels.

Hough transform
Steps of Canny edge detector
Steps of Canny edge detector
(cont’d)

(and direction)
 Canny Edge Detector

• Low error rate of detection

– Well match human perception results
• Good localization of edges
– The distance between actual edges in an
image and the edges found by a
computational algorithm should be minimized
• Single response
– The algorithm should not return multiple
edges pixels when only a single one exists
43
Flow-chart of Canny Edge Detector
Original image

Smoothing by Gaussian convolution

Differential operators along x and y axis

Non-maximum suppression
finds peaks in the image gradient

Hysteresis thresholding locates edge strings

Edge map
44
 Canny Edge Detector Example

original image vertical edges horizontal edges

norm of the gradient after thresholding after thinning

45
 Non-maxima suppression

• Check if gradient magnitude at pixel location (i,j)

is local maximum along gradient direction
Non-maxima suppression
Warning: requires checking
interpolated pixels p and r

(i,j)
 Hysteresis thresholding

• Standard thresholding:

- Can only select “strong” edges.

- Does not guarantee “continuity”.
gradient magnitude low threshold high threshold
 Hysteresis thresholding (cont’d)

• Hysteresis thresholding uses two thresholds:

- low threshold tl
- high threshold th (usually, th = 2tl)

th
tl th
tl

• For “maybe” edges, decide on the edge if

neighboring pixel is a strong edge.
 Hysteresis thresholding/Edge Linking

Idea: use a high threshold to start edge

curves and a low threshold to continue
them.

Use edge
“direction” for
linking edges
 Hysteresis Thresholding/Edge Linking
(cont’d)

(using tl and th)

Note: large gaps are still difficult to bridge.

(i.e., more sophisticated algorithms are required)
Second Derivative in 2D: Laplacian
(cont’d)
Variations of Laplacian
Laplacian - Example

detect zero-crossings
 Properties of Laplacian

• It is an isotropic operator.
• It is cheaper to implement than the
gradient (i.e., one mask only).
• It does not provide information about edge
direction.
• It is more sensitive to noise (i.e.,
differentiates twice).
 Laplacian of Gaussian (LoG)
(Marr-Hildreth operator)
• To reduce the noise effect, the image is first smoothed.

• When the filter chosen is a Gaussian, we call it the LoG

edge detector.

σ controls smoothing

• It can be shown that: (inverted

LoG)
2σ2
Laplacian of Gaussian (LoG)
(inverted LoG) (inverted LoG)

filtering zero-crossings
 Decomposition of LoG
• It can be shown than LoG can be
written as follows:

• 2D LoG convolution can be

implemented using 4, 1D convolutions.
Decomposition of LoG (cont’d)

Steps
Difference of Gaussians (DoG)
• The Laplacian of Gaussian can be approximated by
the difference between two Gaussian functions:

approximation
actual LoG
Difference of Gaussians
(DoG) (cont’d)

(a) (b) (b)-(a)

 Gradient vs LoG
• Gradient works well when the image contains sharp
intensity transitions and low noise.
• Zero-crossings of LOG offer better localization,
especially when the edges are not very sharp.

step edge

ramp edge
 Gradient vs LoG (cont’d)

LoG behaves poorly at corners

 Marr and Hildreth’s Method
Edge is scale-dependent
A different edge map can be generated at different scale
• Scale space representation

f ( x , y ; s )  f ( x , y ;0 )  g ( x , y ; s )
coarse-scale fine-scale Gaussian kernel
image image with width of s
1 x2  y2
g ( x, y; s )  exp(  )
2s 2s
64
 Scale-Space Edge Detection Examples

fine coarse

65
 Image Segmentation based on
property of similarity

f1 f3
f4

f2

Extract the regions or object of interest based on their

properties of similarity.
 Segmentation based on Thresholding
• Based on similarity of intensity value.
• Above threshold  Binary 1
• Below Threshold  Binary 0
• Threshold selection:
• Ex. With dark object and bright background
• Image intensity:
f x. y   T : object point

f x. y   T : background
No. of Pixels Object Background

0 255
T Gray Values
Dark Bright
 Object Background

No. of Pixels

0 Gray Values 255

Threshold operator
• Factors on which T depends:
T  f1x, y, px, y , f x, y 
 Depends on local neighbourhood : f(x,y)
 Vary from region to region : (x,y)
 Depends on local property: P(x,y)
• A thresholded image is defined as

1 if f  x, y   T
g  x, y    f(x,y) = intensity at (x,y)
0 if f  x, y   T
P(x,y) = some property at (x,y)
 Threshold
Types

Global Threshold Local Threshold

T only depends T is a function of
on f(x,y). both P(x,y) and f(x,y).

Dynamic threshold
T is a function
of [x,y, P(x,y) f(x,y)].
 In a controlled environment ( proper lighting condition)
Global thresholding is successful.

 In an uncontrolled environment, Local and Dynamic

thresholding are successful.
 Picking the threshold is the hard part

• Human operator decided the threshold

• Use mean gray level of the image
• A fixed proportion of pixels are detected ( set to 1)
by the thresholding operation
•Analyzing the histogram of an image
 Dynamic (Adaptive) Thresholding

A more complex thresholding algorithm would be to use a

spatially varying threshold. This approach is very useful to
compensate for the effects of non-uniform illumination. If T
depends on coordinates x and y, this referred to as Dynamic
Thresholding or Adaptive Thresholding.
Preprocessing the image to remove noise of other non-
uniformities can improve the performance of the
thresholding.
A technique which often provides better results is to only use
edge points when creating the grey level histogram .
 Adaptive thresholding - how it works?
There are two main approaches to finding the threshold:
(i) the Chow and Kaneko approach and
(ii) local thresholding.
The assumption behind both methods is that smaller image
regions are more likely to have approximately uniform
illumination, thus being more suitable for thresholding.
Chow and Kaneko divide an image into an array of overlapping
subimages and then find the optimum threshold for each
subimage by investigating its histogram. The threshold for each
single pixel is found by interpolating the results of the
subimages. The drawback of this method is that it is
computational expensive and, therefore, is not appropriate for
real-time applications.
ADAPTIVE THRESHOLDING
T12 T13
T11

Image

T22 T23
T21
 Adaptive thresholding - Local thresholding
An alternative approach to find the local threshold is to statistically
examine the intensity values of the local neighborhood of each
pixel. The statistic which is most appropriate depends largely on
the input image.

The size of the neighborhood has to be large enough to cover

sufficient foreground and background pixels, otherwise a poor
threshold is chosen. On the other hand, choosing regions which are
too large can violate the assumption of approximately uniform
illumination. This method is less computationally intensive than
the Chow and Kaneko approach and produces good results for
some applications.
 Optimal thresholding
• We will consider in detail minimisation method
for determining the threshold

• Idealized object/background image histogram

h(i)
2500.00

2000.00

1500.00

1000.00

500.00

0.00 i
0.00 50.00 100.00 150.00 200.00 250.00
85
T
 Optimal thresholding

• Any threshold separates the histogram into 2

groups with each group having its own statistics
(mean, variance)
• The homogeneity of each group is measured by
the within group variance
• The optimum threshold is that threshold which
minimizes the within group variance thus
maximizing the homogeneity of each group

86
 Optimal Thresholding
• p1(z): pdf of dark region.
• p2(z): pdf of bright region.

Overall probability density condition,

h(i)
2500.00

pz   P1 p1 z   P2 p2 z 
2000.00
Background

1500.00

P2(z)
• P1 + P2 = 1 Object
1000.00

500.00
P1(z)
0.00 i
0.00 50.00 100.00 150.00 200.00 250.00

T
 Optimal Thresholding

Dark region : object

Bright region : background

mean of object < mean of background

μ1< μ2

Aim: T should be selected such that

all pixels with a level below T are
object pixels.
Optimal Threshold:
• What is the probability of erroneous classification of
object point as a background point?
• Solution is:
• Classifying background as object
E1 T    p2 z dz
T


• Classifying object as background

E2 T    p1 z dz
• Overall probability of error is T

ET   P2 E1 T   P1 E2 T 
• Threshold value for minimization of error is:
E T 
0
T
• so, P1 p1 T   P2 p2 T 

• p1(T), p2(T): Gaussian PDF

P1   z   1 2  P2   z   2 2 
Pz   exp   exp  
2 1 2 1  2 2 2 2 
2 2
 
• General solution is: AT 2  BT  C  0
A  1   2
2 2

where

B  2 1 2   2 1
2 2

  2 P1 
C   1  2   2 1  2 1  2 ln  
2 2 2 2 2 2

  1 P2 
• If variances are same
1   2 2  P2 
1   2   2
2 2
T  ln  
2 1   2  P1 
• If P1=P2 (probabilities of two levels are equal)
1   2
T
2
Histogram : Approximation of actual pdf
p(z) depends upon p1(z), p2(z), μ1, μ2, σ1, σ2.

These factors should be such that MSE is minimized

between p(z) and observed pdf h(z).

Mean square error between mixture density

and image histogram is:
2
   p  zi   h zi 
1 n
ems
n i 1
n = no of points in the histogram.
 Optimal thresholding
– Minimisation of the within group variance
– Robot Vision, Haralick & Shapiro, volume 1,
page 20
• Let group o (object) be those pixels with
greylevel <=T
• Let group b (background) be those pixels
with greylevel >T
• The prior probability of group o is po(T)
• The prior probability of group b is pb(T)
93
• The following expressions can easily be derived
for prior probabilities of object and background
T
po ( T )   P( i )
i 0
255
pb ( T )   P( i )
i  T 1

P(i )  h( i ) / N
• where h(i) is the histogram of an N pixel image

94
• The mean and variance of each group are
as follows :
T
 o ( T )   iP(i ) / po ( T )
i0
255
 b ( T )   iP( i ) / pb ( T )
i  T 1

 ( T )   i   o ( T ) P(i ) / po ( T )
T 2
2
o
i0

 ( T )   i   b ( T ) P( i ) / pb ( T )
255 2
2
b
i  T 1
95
• The within group variance is defined as :
W2 ( T )   2o ( T ) po ( T )   b2 ( T ) pb ( T )

• We determine the optimum T by

minimizing this expression with respect to
T
– Only requires 256 comparisons for and 8-bit
greylevel image

96
Greylevel thresholding

h(i)
Topt
2500.00

2000.00

Histogram
1500.00
Within group variance

1000.00

500.00

0.00 i
0.00 50.00 100.00 150.00 200.00 250.00

97
 Greylevel thresholding

Low noise image Thresholded at T=124

98
Greylevel thresholding

Low noise image Thresholded at T=124

99
• High level of pixel miss-classification
noticeable

• This is typical performance for

thresholding

– The extent of pixel miss-classification is

determined by the overlap between object
and background histograms.

100
Greylevel thresholding

p(x)

0.02

Background

0.01
Object

x
0.00

o b

T
101
Greylevel thresholding

p(x)

0.02

Background
0.01

Object

x
0.00
o b

T
102
• Easy to see that, in both cases, for any
value of the threshold, object pixels will
be miss-classified as background and
vice versa

• For greater histogram overlap, the pixel

miss-classification is obviously greater
– We could even quantify the probability of
error in terms of the mean and standard
deviations of the object and background
histograms

103
 Greylevel clustering
• Consider an idealized
object/background histogram

Background

Object

c1 c2 104
 Greylevel clustering

• Clustering tries to separate the histogram

into 2 groups

• Defined by two cluster centres c1 and c2

– Greylevels classified according to the nearest
cluster centre

105
 Greylevel clustering

• A nearest neighbour clustering algorithm

allows us perform a greylevel
segmentation using clustering

– A simple case of a more general and widely

used K-means clustering

– A simple iterative algorithm which has known

convergence properties
106
 Greylevel clustering

• Given a set of greylevels

g(1), g( 2 )...... g( N )
• We can partition this set into two groups
g (1), g ( 2 )...... g ( N )
1 1 1 1

g (1), g ( 2 )...... g ( N )
2 2 2 2

107
 Greylevel clustering

• Compute the local means of each group

1 N1
c1   g1 ( i )
N 1 i 1
N2
1
c2 
N2
g
i 1
2 (i )

108
 Greylevel clustering
• Re-define the new groupings

g1 ( k )  c1  g1 ( k )  c2 k  1.. N 1

g2 ( k )  c2  g 2 ( k )  c1 k  1.. N 2

• In other words all grey levels in set 1 are nearer

to cluster centre c1 and all grey levels in set 2
are nearer to cluster centre c2

109
 Greylevel clustering

• But, we have a chicken and egg situation

– The problem with the above definition is that

each group mean is defined in terms of the
partitions and vice versa

– The solution is to define an iterative algorithm

and worry about the convergence of the
algorithm later
110
 Greylevel clustering

• The iterative algorithm is as follows

Initialize the label of each pixel randomly

Repeat
c1= mean of pixels assigned to object label
c2= mean of pixels assigned to background label

Compute partition g (1), g ( 2 )...... g ( N )

1 1 1 1

Compute partition g (1), g ( 2 )...... g ( N )

2 2 2 2

Until none pixel labelling changes

111
 Greylevel clustering

• Outline proof of algorithm convergence

– Define a ‘cost function’ at iteration r

  g1 ( i )  c1     g2 ( i )  c2 
1 N1 ( r ) ( r 1 ) 2 1 N2
( r 1 ) 2
E(r )  (r )

N1 i 1 N 2 i 1
– E(r) >0

112
 Greylevel clustering

• Now update the cluster centres

1 N1 ( r )
c1( r )   g1 ( i )
N1 i 1
N1
1
c2( r ) 
N2
 2 (i)
g (r )

i 1

• Finally update the cost function

  g1 ( i )  c1     g2 ( i )  c2 
1 N1 ( r ) 2 1 N2
(r ) 2
E1( r )  (r ) (r )

N1 i 1 N 2 i 1
113
 Greylevel clustering

• Easy to show that

( r 1 )
E E (r )
1 E (r )

• Since E(r) >0, we conclude that the

algorithm must converge
– but
• What does the algorithm converge to?

114
 Greylevel clustering

• E1 is simply the sum of the variances

within each cluster which is minimised at
convergence
– Gives sensible results for well separated
clusters
– Similar performance to thresholding

115
Greylevel clustering

g2

g1

c1 c2
116
 Let there is a small object against a vast background.

Edge detection Estimate the

boundary pixels Compute
the histogram of the edge
pixels.
No. of pixels

Histogram of the image

Gray labels
 Three regions:

1. High gradient = non-uniform intensity

boundary

2. Low gradient = uniform intensity

1. object
2. Background

Change in intensity f from +ve to –ve

Change in intensity f from –ve to +ve
Change in intensity = 0
 0iff  T

S ( x, y )   iff  Tand 2 f  0
 iff  Tand 2 f  0


Three distinct gray levels:

0 = All pixels that are not on an edge.

+ = All pixels on the dark side of an edge

- = All pixels on the light side of an edge

These boundary pixels having value +
and – are estimated and the histogram is
computed.

The threshold of this histogram is used

to classify the edge pixels into object and
background pixels.

This information is used to generate a

segmented binary image.
 Applications
 Oneof the most powerful and important tool for image
segmentation.

 It has the advantages of smaller storage space, fast

processing speed.

 The simplicity and speed of the thresholding algorithm

make it one of the most widely used algorithms in
automated systems ranging from medical applications to
industrial manufacturing.
Region Orientated Image Segmentation:

Let R be the region describing the complete image.

R1, R2, R3,….,Rn are n different Partitions.

R1 R2

R3

R4
R5
 Properties

Completeness: Image is fully

n
1. U Ri  R
i 1
divided into different regions.
2. Ri is a connected region. All pixels in a region are
connected.
3. Ri  R j   for all i, j where i ≠ j

4. P( Ri )  True for i = 1, 2, 3, ….,n

5. P( Ri  R j )  False for i ≠ j
 Region
segmentation

Region growing Region split and merge

by pixel aggregation
 Region Growing
1. Define the set of seed points.
2. Define the similarity function = P (Ri)
3. Append the neighbouring pixels based upon
the similarity function.
1 0 0 5 6 7 P(Ri) = pixels within Ri
1 2 1 7 6 5 must not have difference
0 1 2 1 4 6 in intensity more than 3
2 1 0 6 7 7
2 2 1 4 5 6
Difficulties:

1. How to start the seed point

2. What is the similarity function

How do we choose the seed(s) in practice ?

•It depends on the nature of the problem.

•If targets need to be detected using infrared

images for example, choose the brightest pixel(s).

•Without a-priori knowledge, compute the

histogram and choose the gray-level values
corresponding to the strongest peaks
How do we choose the similarity criteria (predicate)?

 The homogeneity predicate can be based on any

characteristic of the regions in the image such as:

• average intensity
• variance
• color
• texture
 Region Merging
• Region merging operations eliminate false
boundaries and spurious regions by merging
adjacent regions that belong to the same
object.
• Merging schemes begin with a partition
satisfying condition (4) (e.g., regions
produced using thresholding).

• Then, they proceed to fulfill condition (5) by

gradually merging adjacent image regions.
Region Merging (cont’d)
 How to determine region similarity?
(1) Based on the gray values of the regions –
examples:
– Compare their mean intensities.
– Use surface fitting to determine whether the
regions may be approximated by one surface.
– Use hypothesis testing to judge the similarity
of adjacent region

(2) Based on the weakness of boundaries between

the regions.
Region merging using hypothesis testing
• This approach considers whether or not to merge
adjacent regions based on the probability that they
will have the same statistical distribution of intensity
values.

• Assume that the gray-level values in an image region

are drawn from a Gaussian distribution
– Parameters can be estimated using sample
mean/variance:
R2
R1
 Region merging using hypothesis testing

• Given two regions R1 and R2 with m1 and m2 pixels

respectively, there are two possible hypotheses:
R2
R1
H0: Both regions belong to the same object.
The intensities are all drawn from a single Gaussian
distribution N(μ0, σ0)

H1: The regions belong to different objects.

The intensities of each region are drawn from separate
Gaussian distributions N(μ1, σ1) and N(μ2, σ2)
 Region merging using hypothesis testing

• The joint probability density under H0,

assuming all pixels are independently
drawn, is given by:

• The joint probability density under H1 is

given by
 Region merging using hypothesis testing

• The likelihood ratio is defined as the ratio

of the probability densities under the two
hypotheses:
R2
R1

• If the likelihood ratio is below a threshold

value, there is strong evidence that there
is only one region and the two regions
may be merged.
 Region merging by removing weak edges

• The idea is to combine two regions if the

boundary between them is weak.

• A weak boundary is one for which the

intensities on either side differ by less than
some threshold.

• The relative lengths between the weak

boundary and the region boundaries must
be also considered.
Region merging by removing weak edges

• Approach 1: merge adjacent regions R1 and

R2 if

where:
W is the length of the weak part of the
boundary
S = min(S1, S2) is the minimum of the
perimeter of the two regions.
 Region merging by removing weak edges

• Approach 2: Merge adjacent regions R1 and R2

if

where:
W is the length of the weak part of the boundary
S is the common boundary between R1 and R2.
 Region Splitting
• Region splitting operations add missing
boundaries by splitting regions that contain
parts of different objects.
• Splitting schemes begin with a partition
satisfying condition (5), for example, the
whole image.

• Then, they proceed to satisfy condition (4) by

gradually splitting image regions.
 Region Splitting
• Two main difficulties in implementing
this approach:
– Deciding when to split a region (e.g., use
variance, surface fitting).
– Deciding how to split a region.
 Region Splitting and Merging

• Splitting or merging might not produce

good results when applied separately.
• Better results can be obtained by
interleaving merge and split operations.
• This strategy takes a partition that possibly
satisfies neither condition (4) or (5) with
the goal of producing a segmentation that
satisfies both conditions.
Region Splitting and Merging
Region Splitting and Merging

thresholding split and merge