Chapter 10

Image Segmentation

國立雲林科技大學 資訊工程所
張傳育 (Chuan-Yu Chang ) 博士
Office: EB 212
TEL: 05-5342601 ext. 4337
E-mail: [email protected]
Website: http://MIPL.yuntech.edu.tw
Introduction
 Image Segmentation
 Subdivides an image into its constituent regions or objects.
 Segmentation should stop when the objects of interest in
an application have been isolated.
 Segmentation accuracy determines the eventual success
or failure of computerized analysis procedures.
 Image segmentation algorithm generally are based on one
of two basic properties of intensity values:
 Discontinuity
 Partitioning an image based on abrupt changes in intensity.
 Similarity
 Partitioning an image into regions that are similar according to a
set of predefined criteria.

2
Detection of discontinuities
 There are three basic types of gray-level discontinuities
 Points, lines, and edges.
 The most common way to look for discontinuities is to run a
mask through the image in the manner described in Section
3.5. (sum of product)
R  w1 z1  w2 z 2  ...  w9 z9
(10.1-1)
9
  wi zi
i 1

3
Detection
Point detection
of discontinuities
 An isolated point will be quit different from its surroundings.
 Measures the weighted differences between the center point and
its neighbors.
 A point has been detected at the location on which the mask is
centered if
R T
where T is a nonnegative threshold.
Mask 計 算
後的結果
飛機渦輪
葉片的 X
光影像

有小破洞
取 (c) 圖 最 大
灰 階 的 90%
作為 threshold
4
後的結果
Detection of discontinuities
 Line detection
 Let R1, R2, R3, and R4 denote the response of the mask in
following.
 Suppose that the four masks are run individually through an
image.
 If, at a certain point in the image, |Ri|>|Rj|, for all j≠i, that point is
said to be more likely associated with a line in the direction of
mask i.
 If we are interested in detecting all the lines in an image in the
direction defined by a given mask, we simply run the mask
through the image and threshold the absolute value of the result.
 The coefficients in each mask sum to zero.

5
Detection of discontinuities
 We are interested in finding all the lines that are one pixel
thick and are oriented at -45.
 Use the last mask shown in Fig. 10.3.

偵測到的 45
度 line

強度最強的
line

6
Detection of discontinuities
 Edge detection
 An edge is a set of connected pixels that lie on
the boundary between two regions.
 The “thickness” of the edge is determined by
the length of the ramp.
 Blurred edges tend to be thick and sharp edges tend
to be thin.

7
 Detection of discontinuities

 The first derivative is positive at the points of transition

into and out of the ramp as we move from left to right
along the profile.
 It is constant for points in the ramp,
 It is zero in areas of constant gray level.
 The second derivative is positive at the transition
associated with the dark side of the edge, negative at
the transition associated wit the light side of the edge,
and zero along the ramp and in areas of constant gray
level.

The magnitude of the first derivative can be used to detect the presence
of an edge.
The sign of the second derivative can be used to determine whether
8
an edge pixel lies on the dark or light side of an edge.
Edge detection (cont.)

 Two additional properties of the second

derivative around an edge:
 It produces two values for every edge in an
image.
 An imaginary straight line joining the extreme
positive and negative values of the second
derivative would cross zero near the midpoint of
the edge.
 The zero-crossing property of the second derivative is
quit useful for locating the centers of thick edges.

9
 Detection of discontinuities
 The entire transition from black to white is a single edge.

=0 Image and gray-level profiles

of a ramp edge

First derivative image and

=0.1 the gray-level profile

=1 Second derivative

image and the
gray-level profile

=10

10
Edge detection (cont.)
 The second derivative is even more sensitive to noise.
 Image smoothing should be a serious consideration prior to the
use of derivatives in applications .
 Summaries of edge detection
 To be classified as a meaningful edge point, the transition in gray
level associated with that point has to be significant stronger than
the background at that point.
 Use a threshold to determine whether a value is “significant” or not.
 We define a point in an image as being an edge point if its two
dimensional first-order derivative is greater than a specified
threshold.
 A set of such points that are connected according to a predefined
criterion of connectedness is by definition an edge.
 Edge segmentation is used if the edge is short in relation to the
dimensions of the image.
 A key problem in segmentation is to assemble edge
segmentations into linger edges.
 If we elect to use the second-derivative to define the edge points
in an image as the zero crossing of its second derivative.
11
Detection of discontinuities
 Gradient operator
 The gradient of an image f(x,y) at location (x,y) is defined as
the vector f  
G x   
f      fx 
G y   
 y 

the gradient vector points is the direction of maximum rate of

change of f at coordinates (x,y).
 The magnitude of the vector denoted ∇f, where
 
1
f  mag (f )  G  G 2
x
2 2
y

 The direction of the gradient vector denoted by

 Gy 
 ( x, y )  tan 1  
 Gx 
The angle is measured with respect to the x-axis.

12
Detection of discontinuities
 Roberts cross-gradient operator
 Gx=(z9-z5)
 Gy=(z8-z6)
 Masks of size 2x2 are awkward to
implement because they do not have
a clear center.
 Prewitt operator
 Gx=(z7+z8+z9)-(z1+z2+z3)
 Gy=(z3+z6+z9)-(z1+z4+z7)
 Sobel operator
 Uses a weight of 2 in the center
coefficient.
 Gx=(z7+2z8+z9)-(z1+2z2+z3)
 Gy=(z3+2z6+z9)-(z1+2z4+z7)
13
Detection of discontinuities
 Computation of the gradient requires Gx and Gy be combined in Eq.
(10.1-4), however, this implementations is not always desirable
because of the computational burden required by squares and
square roots.
 An approach used frequently is to approximate the gradient by
absolute values:
f  G x  G y
 The two additional Prewitt and Sobel masks for detecting
discontinuities in the diagonal directions are shown in Fig. 10.9

用來偵測對角邊界
的 Prewitt 及 Sobel
mask 。

14
Detection of discontinuities
 Fig. 10.10 shows the response of the two components of
the gradient, |Gx| and |Gy|.
 The gradient image formed the sum of these two
components.

15
Detection of discontinuities
 Figure 10.11 shows the same sequence of images
as in Fig. 10.10, but with the original image being
smoothed first using a 5x5 averaging filter.
 The response of each mask no shows almost no
contribution due to the bricks, with the result being
dominated mostly by the principal edges.

16
Detection of discontinuities
 The horizontal and vertical Sobel masks respond about equally
well to edges oriented in the minus and plus 45° direction.
 If we emphasize edges along the diagonal directions, the one of
the mask pairs in Fig. 10.9 should be used.
 The absolute responses of the diagonal Sobel masks are shown
in Fig. 10.12.
 The stronger diagonal response of these masks is evident in
these figures.

17
Detection of discontinuities
 The Laplacian of a 2-D function f(x,y) is a second-order
derivative defined as
2 f 2 f (10.1-13)
 f  2  2
2

x y
 For a 3x3 region, one of the two forms encountered most
frequently in practice is ( 水平與垂直邊 )
 2 f  4 z 5  z 2  z 4  z 6  z 8  (10.1-14)
 A digital approximation including the diagonal neighbors is
given by ( 含水平、垂直、及對角邊 )
 2 f  8 z5  z1  z 2  z3  z 4  z6  z7  z8  z9  (10.1-15)

18
Detection of discontinuities
 The Laplacian generally is not used in its original
form for edge detection for several reasons:
 As a second-order derivative, the Laplacian typically is
unacceptably sensitive noise.
 The magnitude of the Laplacian produces double edges.
 The Laplacian is unable to detect edge direction.
 The role of the Laplacian in segmentation consists of
 Using its zero-crossing property for edge location.
 Using it for complementary purpose of establishing whether
a pixel is on the dark or light side of an image.

19
Detection of discontinuities
 Laplacian of a Gaussian (LoG)
 The Laplacian is combined with smoothing as a precursor to
finding edges via zero-crossing, consider the function
r2

h( r )  e 2 2 (10.1-16)
where r2=x2+y2 and  is the standard deviation.
 Convolving this function with an image blurs the image, with
the degree of blurring being determined by the value of .
 The Laplacian of h (the second derivative of h with respect to
r) is
r2
 r 2   2   2 2 (10.1-17)
 h( r )   
2
e
 
2


 The function is commonly referred to as the “Laplacian of a

Gaussian” (LoG), sometimes is called the ”Mexican hat”
function.
20
Detection of discontinuities

 The purpose of the Gaussian function in the LoG formulation is

to smooth the image, and the purpose of the Laplacian operator
is to provide an image with zero crossings used to establish the
location of edges.

21
Detection of discontinuities
Fig. 10.15(c) is a spatial Gaussian

function (with a standard deviation
of five pixels) used to obtain a
27x27 spatial smoothing mask.
The mask was obtained by
sampling this Gaussian function at
equal interval.
 ∇2h can be computed by
application of (c) followed by (d).
 The LoG result shown in Fig.
10.15(e) is the image from which
zero crossings are computed to
find edges.
 One straightforward approach for
approximating zero-crossings is to
threshold the LoG image by setting
all its positive values to white, and
all negative values to black.
 Zero-crossing occur between
positive and negative values of the
Laplacian.
 Estimated zero-crossing, obtained
by scanning the threshold image
and noting the transitions between
22
black and white.
Detection of discontinuities

 Comparing Figs/ 10.15(b) and (g)

 The edges in the zero-crossing image are thinner
than the gradient edges.
 The edges determined by zero-crossings form
numbers closed loops.
 “spaghetti effect” is one of the most serious drawbacks
of this method.
 Another major drawback is the computation of
zero crossing.

23
Edge Linking and Boundary
Detection
Ideally, edge detection should yield pixels lying only on
edges.
 In practice, this set of pixels seldom characterizes an
edge completely because of noise, breaks in the edge
from nonuniform illumination, and other effects that
introduce spurious intensity discontinuities.
 Thus, edge detection algorithms are followed by linking
procedures to assemble edge pixels into meaningful
edges.
 Local Processing
 To analyze the characteristics of pixels in a small neighborhood
about every point (x,y) in an image that has been labeled an
edge point.
 All points that are similar according to a set of predefined criteria
are linked.

24
Edge Linking and Boundary
Detection
 The two principal properties used for
establishing similarity of edge pixels:
 The strength of the response of the gradient
operator used to produce the edge pixel. As
defined in Eq.(10.1-4)
 The direction of the gradient vector. As defined in
Eq. (10.1-5)

25
Edge Linking and Boundary
Detection
An edge pixel with coordinates (x ,y ) in a predefined neighborhood of
0 0
(x,y), is similar in magnitude to the pixel at (x,y) if

f ( x , y )  f ( x 0 , y 0 )  E
 where E is a nonnegative threshold
 An edge pixel at (x0,y0) is the predefined neighborhood of (x,y) has an
angle similar to the pixel at (x,y) if
 ( x, y )   ( x0 , y 0 )  A
 where A is a nonnegative angle threshold
 A point in the predefined neighborhood of (x,y) is linked to the pixel at
(x,y) if both magnitude and direction criteria are satisfied 。
 This process is repeated at every location in the image.
 A record must be kept of linked points as the center of the
neighborhood is moved from pixel to pixel.

26
 Edge Linking and Boundary
Detection
Example 10-6: the objective is to find rectangles whose sizes

makes them suitable candidates for license plates.
 The formation of these rectangles can be accomplished by
detecting strong horizontal and vertical edges.
 Linking all points, that had a gradient value greater than 25 and
whose gradient directions did not differ by more than 15°.
使用垂直的
Sobel operator

使用水平的 分別對圖 (b) 及

Sobel operator (c) 進行 edge
linking 的動作，
將梯度大於 25 ，
且角度小於 15 度
的點連起來。
27
Edge Linking and Boundary
Global Processing via the Hough Transform
Detection
Points are linked by determining first if they lie on a curve of specified

shape.
 Given n points in an image, suppose that, we want to find subsets of
these points that lie on straight lines.
 Consider a point (xi,yi) and the general equation of a straight line in slope-
intercept form, yi=axi+b.
 Infinitely many lines pass through (xi,yi), but they all satisfy the equation
yi=axi+b for varying values of a and b.
 However, writing this equation as b=-xia+yi and considering the ab-plane
yields the equation of a single line for a fixed pair (xi,yi) .
 A second point (xj,yj) also has a line in parameter space associated with
it, and this line intersects the lines associated with (xi,yi) at (a’, b’).
 All points contained on this line have lines in parameter space that
intersect at (a’, b’).

28
Edge Linking and Boundary
 Hough Transform

Detection



Subdividing the parameter space into so-called accumulator cell
Initially, these cells are set to 0.
For every point (xk, yk) in the image plane, we let the parameter a equal each of the allowed
subdivisions values on the a-axis and solve for the corresponding b using the equation b=-
xka+yk.
 The resulting b’s are then rounded off to the nearest allowed value in the b-axis.
 If a choice of ap results in solution bq ， we let A(p,q)=A(p,q)+1 .
 At the end of this procedure, a value of Q in A(i,j) corresponds to Q points in the xy-plane lying
on the line y=aix+bj.
 The number of subdivisions in the ab-plane determines the accuracy of the co-linearity of
these points.

29
Edge Linking and Boundary
A problem with using the equation y=ax+b to represent a line
Detection
is that the slope approaches infinity as the line approaches
the vertical.
 To use normal representation of a line
x cos   y sin    (10.2-3)
 Figure 10.19(a) shows the geometrical interpretation od the
parameters used in Eq. (10.2-3).
 Q collinear points lying on a line xcosj +ysinj=i yield Q sinusoidal
curves that intersect at (i,j) in the parameter space.
 Incrementing  and solving for the corresponding  gives Q entries in
accumulator A(i,j) associated with the cell determined by (i,j) .

30
Edge Linking and Boundary
Detection
X, Y 平面
上五個點
(1, 2, 3, 4,
5) ，在
平面的曲
X, Y 平面上有 線
五個點 (1, 2,
3, 4, 5)

從交點 A 知
道， 點 1,
3, 5) 共線。
交點 B 表示
點 2,3,4 共
線
31
Edge Linking and Boundary
Detection
 Edge-linking based on Hough transform
 Compute the gradient of an image and threshold it to obtain
a binary image.
 Specify subdivisions in the -plane.
 Examine the counts of the accumulator cells for high pixel
concentrations.
 Examine the relationship between pixels in a chosen cell.
( 依其對應的找出直線 ) 。
 Based on computing the distance between disconnected
pixels identified during traversal of the set of pixels
corresponding to a given accumulator cell.
 A gap at any point is significant if the distance between that
point and its closest neighbor exceeds a certain threshold.

32
 Edge Linking and Boundary
Detection
Fig. (a) is an aerial infrared image containing two hangars and a runway.

 Fig. (b) is a thresholded gradient image obtained using the Sobel operator.
 Fig. (c) shows the Hough transform of the gradient image.
 Fig. (d) shows the set of pixels linked according to the criteria
 They belonged to one of the three accumulator cells with the highest count.
 No gaps were longer than five pixels.

33
 Edge Linking and Boundary
 Global Processing via Graph-Theoretic Techniques
Detection
 A global approach for edge detection and linking based on representing
edge segments in the form of a graph and searching the graph for low-cost
paths that correspond to significant edges.
 This representation provides a rugged approach that performs well in the
presence of noise.
 Graph G=(N,U)
 N: set of node
 U: unordered pairs of distinct elements of N
 Each pair (ni,nj) of U is called arc ， ni, is said to be a parent ， nj is said to be a
successor 。
 The process of identifying the successor of a node is called expansion 。
 In each graph we define levels, such that level 0 consists of a single node,
called the start or root, and the nodes in the last level are called goal nodes.
 Cost (ni,nj) can be associated with every arc (ni,nj).
 A sequence of nodes n1, n2,…,nk, with each node ni being a successor of node ni-
1, is called a path from n1 to nk.
 The cost of the k
entire path is
p q
c   cni 1 , ni 
i 2
34
Edge Linking and Boundary
Detection Each edge element defined by pixels p and q,
has an associated cost, defined as
c ( p, q )  H   f ( p )  f ( q ) 

座標 成本 where H is the highest gray-level value

灰階值 in the image, and f(x) is the gray-level
value of x.

35
Edge Linking and Boundary
Detection
By convention, the point p is on the right-hand side of the direction


of travel along edge elements.

 To simplify, we assume that edges start in the top row and
terminate in the last row.
 p and q are 4-neighbors.
 An arc exists between two nodes if the two corresponding edge
elements taken in succession can be part of an edge.
 The minimum cost path is shown dashed.
 Let r(n) be an estimate of the cost of a minimum-cost path from s
to n plus an estimate of the cost of that path from n to a goal node;

r n   g n   hn  (10.2-7)

 Here, g(n) can be chosen as the lowest-cost path from s to n found

so far, and h(n) is obtained by using any variable heuristic
information.

36
Edge Linking and Boundary
Detection

37
 Edge Linking and Boundary

Detection
Graph search algorithm
Step1: Mark the start node OPEN and set g(s)=0.
 Step 2: If no node is OPEN exit with failure; otherwise, continue.
 Step 3: Mark CLOSE the OPEN node n whose estimate r(n) computed from Eq.
(10.2-7) is smallest.
 Step 4: If n is a goal node, exit with the solution path obtained by tracing back
through the pointets; otherwise, continue.
 Step 5: Expand node n, generating all of its successors (If there are no
successors go to step 2)
 Step 6: If a successor ni is not marked, set

r ni   g n   cn, ni 
 Step 7: if a successor ni is marked CLOSED or OPEN, update its value by
letting
g ' ni   ming ni , g n   cn, ni 

Mark OPEN those CLOSED successors whose g’ value were thus lowered and
redirect to n the pointers from all nodes whose g’ values were lowered. Go to
Step 2. 38
 Edge Linking and Boundary
Detection
 Example 10-9: noisy chromosome silhouette and an

edge found using a heuristic graph search.

 The edge is shown in white, superimposed on the
original image.

39
Thresholding
 Thresholding
 To select a threshold T, that separates the objects from

the background.
 Then any point (x,y) for which f(x,y)>T is called an object point;
otherwise, the point is called a background point.
 Multilevel thresholding
 Classifies a point (x,y) as belonging to one object class if T1
<f(x,y) ≤T2, and to the other object class if f(x,y) >T2
 And to the background if f(x,y) ≤T2

40
Thresholding
 In general, segmentation problems requiring multiple thresholds are
best solved using region growing methods.
 The thresholding may 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 

where f(x,y) is the gray-level of point (x,y) and p(x,y) denotes some
local property of this point.
 A threshold image g(x,y) is defined as
1 if f ( x, y )  T
g x, y   
0 if f ( x, y )  T
 Thus, pixels labeled 1 correspond to objects, whereas pixels labeled
0 correspond to the background.
 When T depends only on f(x,y) the threshold is called global. If T
depends on both f(x,y) and p(x,y), the threshold is called local.
 If T depends on the spatial coordinates x and y, the threshold is
called dynamic or adaptive.

41
The role of illumination
 An image f(x,y) is formed as the product of a
reflectance component r(x,y) and an illumination
component i(x,y).
(10.3-4)
f x, y   i x, y r x, y 
 In ideal illumination, the reflective nature of objects
and background could be easily separable.
 However, the image resulting from poor illumination
could be quit difficult to segment.
 Taking the natural logarithm of Eq.(10.3-3)
z x, y   ln f x, y 
 ln i x, y   ln r x, y  (10.3-5)

 i ' x, y   r ' x, y 

42
The role of illumination

 From probability theory,

 If i’(x,y) and r’(x,y) are independent random
variables, the histogram of z(x,y) is given by the
convolution of the histograms of i’(x,y) and r’(x,y).

43
 The role of illumination
電腦產生的反射函數

f ( x, y )  i ( x, y ) r ( x, y )

影像 f(x,y) 可看作是反 電腦產生的照度函數

射量 r(x,y) 和照度 i(x,y)
的乘積。

Fig. (a)*Fig(c)
物體和背景的反射特性，
使她們容易被分割，但
差的照明，會使產生的
影像難以分割。
44
Thresholding
 Basic global thresholding
 Select an initial estimate for T
 Segment the image using T
 G1 : consisting of all pixels with gray level values > T

 G2 : consisting of all pixels with gray level values <= T

 Compute the average gray level values 1 and 2 for pixels

in regions G1 and G2.
 Compute a new threshold value
 T=0.5*( +  )
1 2
 Repeat step 2 through 4 until the difference in T in
successive iterations is smaller than a predefined
parameter T0.
 The parameter T0 is used to stop the algorithm after
changes become small.
45
Basic Global Thresholding
 To partition the image histogram by using a single global
threshold T.
 Segmentation is then accomplished by scanning the image
pixel by pixel and labeling each pixel as object or
background, depending on whether the gray level of that
pixel is great or less than the value T.

46
 Basic Global Thresholding
 Fig. (a) is the original image, (b) is the image histogram.
 The clear valley of the histogram.

 Application of the iterative algorithm resulted in a value of 125.4

after three iterations starting with the average gray level and T0=0.
 The result obtained using T=125 to segment the original image is
shown in Fig. (c).

47
 Thresholding
 Basic adaptive thresholding
 Imaging factors such as uneven illumination can transform a
perfectly segmentable histogram into a histogram that cannot be
partitioned effectively by a single global threshold.
 To divide the original image into subimages and then utilize a
different threshold to segment each subimage.
 The key issues are
 How to subdivide the image?
 How to estimate the threshold for each resulting subimages?

Global threshold
手動將 T 設在山谷
處。
將原始影像根據亮度的
變化分成 16 區塊。

48
Thresholding 灰階值得分佈極不
均勻，全域門限法
注定失敗

圖 10.30(c) 的 (1,2)
及 (2,2) 子影像

分成更多的子影像

49
Thresholding (cont.)
 Optimal Global and Adaptive Thresholding
 Estimating threshold that produce the minimum average
segmentation error.
 Suppose that an image contains only two principal gray-
level regions.
 Let z denote gray-level values.
 We can view these values as random quantities, and their
histogram may be considered an estimate of their
probability density function, p(z).
 The overall density function is the sum or mixture of two
densities in the image.
 If the form of the densities is known or assumed, it is
possible to determine an optimal threshold for segmenting
the image into two distinct regions.

50
Thresholding (cont.)
 Optimal Global and Adaptive Thresholding
 The mixture probability density function describing the overall
gray-level variation in the image is
p ( z )  P1 p1 ( z )  P2 p 2 ( z ) (10.3-5)

Here, P1 and P2 are the probabilities of occurrence of the two

classes of pixels.
 Assuming that any given pixel belongs either to an object or to the
background,
P1  P2  1 (10.3-6)

background

object

51
Thresholding (cont.)
 An image is segmented by classifying as background all pixels with
gray levels greater than a threshold T. All other pixels are called
object pixels.
 The main objective is to select the value of T that minimizes the
average error in making the decisions that a given pixel belongs to
an object or to the background.
 The probability of erroneously classifying a background point as an
object point is
T
E1 (T )  

p 2 ( z )dz 將背景誤認為物體的機率。 (10.3-7)
 The probability of erroneously classifying a object point as an
background point is

E 2 (T )   p1 ( z )dz 將物體誤認為背景的機率。 (10.3-8)
T

 The overall probability of error is

E (T )  P2 E1 (T )  P1 E 2 (T ) 總誤差 (10.3-9)

52
Thresholding (cont.)
 To find the threshold value for which this error is minimal requires
differentiating E(T) with respect to T and equating the result to 0.
The result is為求得最小誤差的門限值 T ，須求 E(T) 對 T 的偏微分。
P1 p1 (T )  P2 p 2 (T ) (10.3-10)
This equation is solved for T to find the optimum threshold. If
P1=P2, the optimum threshold is where the curves for p1(z) and
p2(z) intersect.
 Obtaining an analytical expression for T requires that we know the
equations for the two PDFs.
 Estimating these densities in practice is not always feasible. One of
the principal densities used is the Gaussian density, which is
completely characterized by two parameters, the mean and the
variance ( 以 Gaussian 密度函數來近似 p(z))

z  1 2 
z   2 2
pz  
P1
e 2 12

P2
e 2 22 (10.3-11)
2  1 2  2
where 1 and  are the mean and variance of the Gaussian density of
one class of pixels.
53
Thresholding (cont.)
 Using this equation in the general solution of Eq.(10.3-10) results
in the following solution for the threshold T: (10.3-10) 式的通解 )
AT 2  BT  C  0 (10.3-12)

where
A   12   22

B  2 1 22   2 12  (10.3-13)

C   12  22   22 12  2 12 22 ln  2 P1 /  1 P2 
 Since a quadratic equation has two possible solution, two
threshold values may be required to obtain the optimal solution.
 If the variances are equal, a single threshold is sufficient:
1   2 2 P  (10.3-14)
T  ln 2 
2 1   2  P1 

 If P1=P2, the optimal threshold is the average of the means

54
Thresholding (cont.)
 Instead of assuming a functional form for p(z), a minimum
mean-square-error approach may be used to estimate a
composite gray-level PDF of an image from the image
histogram.
 The mean square error between the (continuous) mixture
density p(z) and the (discrete) image histogram h(zi) is
1 n
ems    pz i   hz i 2 (10.3-15)
n i 1
where an n-point histogram is assumed.
 The principal reason for estimating the complete density is to
determine the presence or absence dominant modes in the
PDF.
 Two dominant modes typically indicate the presence of edges in
the image (or region) over which the PDF is computed.

55
 Thresholding
Example 10.13
(cont.)
 The general problem is to outline automatically the boundaries of
heart ventricles in cardioangiograms.
 Pre -processings:
 Each pixel was mapped with a log function to counter exponential
effects caused by radioactive absorption.
 An image obtained before application of the contrast medium was
subtracted from each image captured after the medium was injected
in order to remove the spinal column present in both images.
 Several angiograms were summed to reduce random noise.

原始的心臟影像，打
有顯影劑，目的在於
前處理 :
描繪出左心室的輪廓。 (1) 每個像素點經 log function 轉換
(2) 將打藥前與打藥後的影像相減，以消除脊椎部份。
(3) 將多張心臟影像平均以消除雜訊。
56
Thresholding (cont.)
 Each preprocessed image was subdivided into 49 regions by placing a 7x7
grid with 50% overlap over each image.
 Each of the 49 resulting overlapped regions contained 64x64 pixels.
 The histogram for region A is bimodal, indicating the presence of a boundary.
 The histogram for region B is unimodal, indicating the absence of two
markedly distinct regions.
 After all 49 histograms were computed, a test of bimodality was performed to
reject the unimodal histograms.
 The remaining histograms were then fitted by bimodal Gaussian density
curves [see Eq. (10.3-11)] using a conjugate gradient hill-climbing method to
minimize the error function given in Eq(10.3-15).

Block A 和 Block B
的影像 histogram 57
Thresholding (cont.)
 The x’s and o’s in Fig. 10.34 (a) are two fits to the histogram shown in
black dots.
 The optimal thresholds were then obtained by using Eqs.(10.3-12) and .
(10.3-13)
 For non-bimodal regions, the thresholds were obtained by interpolating
these thresholds.
 Then a second interpolation was carried out point by point using
neighboring threshold values so that, at the end of the procedure, every
point in the image had been assigned a threshold.
 Finally, a binary decision was carried out for each pixel using the rule
1 if f(x,y)  Txy
f x, y   
0 otherwise

58
 Thresholding
 Use of boundary characteristics for histogram improvement and
local thresholding
 The chances of selecting a “good” threshold are enhanced
considerably if the histogram peaks are tall, narrow, symmetric,
and separated by deep valleys.
 One approach for improving the shape of histogram is to consider
only those pixels that lie on or near the edges between objects and
the background. 。
 If only the pixels on or near the edge between object and the
background were used, the resulting histogram would have peaks
of approximately the same height.
 In practice the valleys of histograms formed from the pixels
selected by a gradient/Laplacian criterion can be expected to be
sparsely populated.
0 f  T 非邊界點標為 0

sx, y    f  T and  2 f  0 邊界點的深色邊標為 +
 f  T and  2 f  0
 邊界點的亮色邊標為 -

59
Thresholding (cont.)
 Figure 10.36 shows the labeling produced by Eq. (10.3-16) for an
image of a dark, underlined stroke written on a light background.
 The information obtained with this procedure can be used to
generate a segmented, binary image in which 1’s correspond to
object of interest and 0’s correspond to the background.
 The transition from a light background to a dark object must be
characterized by the occurrence of a – followed by a + in s(x,y).
 The interior of the object is composed of pixels that are labeled
either 0 or +.
 A horizontal or vertical scan line containing a section of an object
has the following structure:
(…)(-,+)(o or +)(+, -)(…)

60
 Thresholding (cont.) An ordinary scenic bank
check

Example 10.14
The histogram as a function of
gradient values for pixels with
gradients greater than 5.
This histogram has two
dominant modes that are
symmetric, nearly of the same
height, and are separated by a
distinct valley.

The segmented result obtained by

using Eq.(10.3-16) with T at or near
of the midpoint of the valley.
61
 Thresholding (cont.)
 Thresholds based on several variables
 Multispectral thresholding
 Constructing a 3-D “histogram” 。
 The basic procedure is analogous to the method used for one variable.
 The concept of the threshold now becomes one of finding clusters of points in
3-D space.
 The principal difficultly is that cluster seeking becomes an increasingly
complex task as the number of variables increases.

Obtained by thresholding about one of the histogram

clusters corresponding to facial tones.

A monochrome
picture of a color
photograph.

Obtained by thresholding about a

cluster close to the red axis. 62
Region-Based Segmentation
 Basic Formulation
 Let R represent the entire image region.
 We may view segmentation as a process that partitions R into
n subregions, R1, R2,…,Rn
n
(a) R i 1
i R Every pixel must be in a region

Points in a region must be connected

(b) Ri is a connected region, i  1, 2 ,..., n in some predefined sense.
(c) Ri  R j   for all i and j , i  j The regions must be disjoint.

(d) P ( Ri )  TRUE for i  1,2,..., n The properties must be satisfied by the pixels

(e) P Ri  R j   FALSE for i  j

in a segmented region
Regions Ri and Rj are different in the sense of
predicate P.

63
Region-Based Segmentation
 Region Growing
 Group pixels or subregions into regions based on predefined
criteria.
 Start with a set of “seed” points and from these grow regions
by appending to each seed those neighboring pixels that have
properties similar to the seed.
 Problems about region growing:
 Selecting a set of seed points

 Selecting of similarity criteria depends not only on the

problem under consideration, but also on the type of image
data available.
 The formulation of a stooping rule.
 Growing a region should stop when no more pixels satisfy the criteria
for inclusion in that region.

64
Region-Based Segmentation
 Example 10-16 焊接點的檢測
 To use region growing to segment the regions of the weld
failures.
An X-ray image of a We selected as
weld, the pixels of starting points
defective welds tend all pixels having
to have values of values of 255.
255.

Criteria for region growing:

(1) The absolute gray-level difference between any pixel and the seed had to be less than 65.
(2)To be included in one of the regions, the pixel had to be 8-connected to at least one pixel in 65
that region
Region-Based Segmentation
 Problems having multimodal histograms generally are best
solved using region-based approach.
 Histogram of Fig.10.40 .( 無法清楚分離物體與背景 )
 This problem cannot be solved effectively by any reasonably
general method of automatic threshold selection based on
gray levels alone.
 The use of connectivity was fundamental in solving the
problem.

66
 Region-Based Segmentation
 Region Splitting and Merging
 To subdivide an image initially into a set of arbitrary, disjointed

regions and then merge and/or split the regions in an attempt to

satisfy the conditions stated in Section 10.4.1. ( 一開始將影像分割
成任意子影像，然後再合併或分割成滿足條件的 segmentation 。 )
 Let R represent the entire image region and select a predicate P.

 Subdivide R successively into smaller and smaller quadrant regions

so that, for any region Ri, P(Ri)=TRUE

 We start with the entire region. If P(R)=FALSE, we divide the image into
quadrants. ( 表示區域 R 內的 pixel 有不同的灰階值 ) ，則將 R 分成四
的子影像。
 If P is FALSE for any quadrant, we subdivide that quadrant into
subquadrants, and so on. ( 如果子影像的 P 為 FALSE 則繼續分割成四個
子影像。 )
 This splitting technique is called “quadtree”
 The root of the tree corresponds to the entire image and that node
corresponds to a subdividion.

67
Region-Based Segmentation

Region Splitting and Merging

68
Region-Based Segmentation

 Region Splitting and Merging

 Split into four disjoint quadrants any region Ri for
which P(Ri)=FALSE.
 Merge any adjacent regions Rj and Rk for which
P(Rj ∪Rk)=TRUE.
 Stop when no further merging or splitting is
possible.

69
Region-Based Segmentation

We define P(Ri)=TRUE if at least 80% of the pixels in Ri have the

property |zj-mi|<=2i, where zi is the gray level of the j-th pixel in Ri, mi
is the mean gray level of that region, and i is the standard deviation
of the gray levels in Ri.

Original maple leaf.

The shading and the stem of the leaf were erroneously

eliminated by the thresholding procedure. 70
Segmentation by Morphological
 Basic Concepts
Watersheds
 The concept of watersheds is based on visualizing an
image in three dimensions: two spatial coordinates versus
gray levels.
 In such a “topographic” interpretation, we consider three
types of points:
 Points belonging to a regional minimum. ( 區域最小值的點。 )
 Points at which a drop of water, if placed at the location of any
of those points, would fall with certainty to a single minimum.
[ 一滴水的點 ( 單一最小值 ) 。形成集水盆 (catchment basin)
或分水嶺 (watershed)]
 The set of points satisfying condition (b) is called the catchment
basin or watershed of that minimum.
 Points at which water would be equally likely to fall to more
than one such minimum. 水可能會掉入的多個最小值的點。形
成峰線 (crest line)
 The points satisfying condition (c) form crest lines on the
topographic surface and are termed divide lines or watershed
lines.
 The principal objective of segmentation algorithms based
on these concepts is to find the watershed line 。 71
Segmentation by Morphological
Watersheds
The basic idea is
 Suppose that a hole is punched in each regional minimum.
 The entire topography is flooded from below by letting
water rise through the holes at a uniform rate.
 When the rising water in distinct catchment basins is about
to merge, a dam is built to prevent the merging.
 The flooding will eventually reach a stage when only the
tops of the dams are visible above the water line.
 These dam boundaries correspond to the divide lines of the
watersheds.
 They are the continuous boundaries extracted by a watershed
segmentation algorithm.

72
 Segmentation by Morphological
Watersheds
A simple gray- A topographic view.
scale image

A hole is punched in
each regional
minimum and that the
entire topography is As the water
flooded from below by continues to rise,
letting water rise it will eventually
through the holes at a overflow from one
uniform rate. catchment basin
into another.

The water has risen

into the first and
second catchment 73
 Segmentation by Morphological
Watersheds
水持續溢入第
二個集水區

Water from the left

basin actually
overflowed into the
basin on the right
and a short dam
was built to prevent
water from merging
at that level of
flooding. The final dams
corresponds to the
watershed lines, which
are the desired
segmentation result.

74
Segmentation by Morphological
Watersheds
 Watershed 法的優點：
 Watershed line form a connected path, thus giving
continuous boundaries between regions
 Watershed segmentation is in the extraction of
nearly uniform (bloblike) objects from the
background.

75
Dam Construction
 Dam construction is based on binary images.
 The simple way to construct dams separating sets of binary point
is to use morphological dilation.
 Suppose that each of the connected components is dilated by the
structuring element, subject to two conditions:
 The dilation has to be constrained to q. (the center of the structuring
element can be located only at points in q during dilation)
 The dilation cannot be performed on points that would cause the sets
being dilated to merge (become a single connected component).
 Only points in q that satisfy the two conditions under
consideration describe the one-pixel-thick connected path.
 This path constitutes the desired separating dam at stage n of
flooding.
 Construction of the dam at this level of flooding is completed by
setting all the points in the path just determined to a value greater
than the maximum gray-level value of the image.
 The height of all dams is generally set at 1 plus the maximum allowed
value in the image.
76
 在第 n-1 次注入水後
Dam Construction 的集水區 2 ， Cn-
使用二元化影像的 dilation 1 (M2)

Two partially flooded catchment

basins at stage n-1 of flooding. 在第
C[n-1]
n-1 次注入水後 的集水區 1 ， Cn-
1 (M1)

Flooding at stage
n, showing that
water has spilled
第一次 between basins.
dilation 第 n 次注入水後
的集水區 q

第二次
dilation

將二次 dilation 重疊的

區域建成水壩 77
Watershed Segmentation Algorithm
 Let M1, M2,…,MR be sets denoting the coordinates of the points in
the regional minima of an image g(x,y).
 Let C(Mi) be a set denoting the coordinates of the points in the
catchment basin associated with regional minimum Mi.
 Let T[n] represent the set of coordinates (s,t) for which g(s,t)<n,
 
That is T n  s, t  g s, t   n (10.5-1)
 T[n] is the set of coordinates of points in g(x,y) lying below the plane
g(x,y)=n.
 The topography will be flooded in integer flood increments, from
n = min + 1 to n = max + 1.
 At any step n of the flooding process, the algorithm needs to
know the number of points below the flood depth.
 Suppose that the coordinates in T[n] that are below the plane
g(x,y) = n are “marked” black, and all other coordinates are
marked white.
 When we look “down” on the xy-plane at any increment n of
flooding, we will see a binary image in which black points
correspond to points in the function that are below the plane
g(x,y)=n.
78
Watershed Segmentation Algorithm
 Let Cn(Mi) denote the set of coordinates of points in the
catchment basin associated with minimum Mi that are
flooded at stage n.
 Cn(Mi) may be viewed as a binary image given by
C n M i   C M i   T n (10.5-2)

Cn(mi)=1 at location (x,y) if (x,y) C(Mi) AND (x,y)  T[n];

otherwise Cn(Mi)=0.
 Next, we let C[n] denote the union of the flooded catchment
basins portions
R at stage n:
C n  C n M i  (10.5-3)
i 1

 Then C[max+1] isR the union of all catchment basins:

C max  1  C M i  (10.5-4)
i 1
79
Watershed Segmentation Algorithm
 Procedure for obtaining C[n] from C[n-1]
 Let Q denote the set of connected components in T[n].

 For each connected component qQ[n], there are three

possibilities:
 q∩C[n-1] is empty.
 q∩C[n-1] contains one connected component of C[n-1] .
 q∩C[n-1] contains more than one connected component of C[n-1] .
 Condition (a) occurs when a new minimum is encountered, in which
case connected component q is incorporated into C[n-1] to form C[n].
 Condition (b) occurs when q lies within the catchment basin of some
regional minimum, in which case q is incorporated into C[n-1] to form
C[n].
 Condition (c) occurs when all, or part, of a ridge separating two or
more catchment basins is encountered. Further flooding would cause
the water level in these catchment basins to merge. Thus a dam
must be built within q to prevent overflow between the catchment
basins.

80
 Segmentation by Morphological
Watersheds
Original blobs
image Image grident

Watershed lines
superimposed on
Watershed lines original image
of the gradient
image

81
Segmentation by Morphological
Watersheds
 直接應用 watershed segmentation 會導致 over segmentation
 可使用 marker 來控制 over segmentation 。
 Marker 是影像中相連接的元件。
 Internal marker 結合有興趣的物件。
 External marker 結合背景。

82
 Segmentation by Morphological
 Watersheds
The use of marker
 Direct application of the watershed segmentation algorithm
generally leads to oversegmentation due to noise and other local
irregularities of the gradient.
 A practical solution is to limit the number of allowable regions by
incorporating a preprocessing stage designed to bring additional
knowledge into the segmentation procedure.
 An approach used to control oversegmentation is based on the concept
of markers.
 A marker is a connected component belonging to an image.
 Internal markers associated with objects of interest.
 External markers associated with the background.
 A procedure for marker selection will consist of two steps ：
 preprocessing ： for minimizing the effect of small spatial detail is to
filter the image with a smoothing filter.
 Definition of a set of criteria that markers must satisfy: (internal marker)
 A region is surrounded by points of higher “altitude”
 The points in the region form a connected.
 All the points in the connected component have the same gray-level value.
83
Segmentation by Morphological
Watersheds

1. 先找出 internal marker

2. 再以 watershed 找出 watershed line
84
The Use of Motion in Segmentation
 Spatial Techniques
 The simplest approaches for detecting changes between two
image frames f(x, y, ti) and f(x,y, tj) taken at times ti and tj,
respectively, is to compare the two images pixel by pixel.
 To form a difference image.
 Suppose that we have a reference image containing only stationary
components.
 Comparing this image against a subsequent image of the same
scene, but including a moving object, result in the difference of the
two images canceling the stationary elements, leaving only nonzero
entries that correspond to the nonstationary image components.
 A difference image between two images taken at time ti and tj may be
defined as
1 f ( x, y, ti )  f ( x, y, t j )  T
d ij ( x, y )  
0 otherwise
 where T is a specified threshold. dij(x,y) has a value of 1 at spatial
coordinates (x,y) only if the gray-level difference between the two
images is appreciably different at those coordinates, as determined
by the specified threshold T.

85
The Use of Motion in Segmentation
(cont.)
In dynamic image processing, all pixels in d (x,y) with value

ij
1 are considered the result of object motion.
 This approach is applicable only if the two images are
registered spatially and if the illumination is relatively
constant within the bounds established by T.
 In practice, 1-valued entries in dij(x,y) often arise as a result
of noise.
 These entries are isolated points in the difference image,
and a simple approach to their removal is to form 4- or 8-
connected regions of 1’s in dij(x,y) and then ignore any
region that has less than a predetermined number of
entries.

86
The Use of Motion in Segmentation
(cont.)
 Accumulative differences
 Consider a sequence of image frames f(x, y, t1), f(x, y, t2), …, f(x,
y, tn).
 Let f(x, y, t1) be the reference images.
 An accumulative difference image (ADI) is formed by comparing
this reference image with every subsequent image in the
sequence.
 A counter for each pixel location in the accumulative image is
incremented every time a difference occurs at that pixel location
between the reference and an image in the sequence.
 When the kth frame is being compared with the reference, the
entry in a given pixel of the accumulative image gives the
number of times the gray level at that position was different from
the corresponding pixel value in the reference image.
 Three types of accumulative difference images (ADI)
 Absolute, Positive, and Negative ADI
 Assuming that the gray-level values of the moving objects are large
than the background, these three types of ADIs of the moving
objects are defined as follows.
87
The Use of Motion in Segmentation
(cont.)
 Let R(x,y) denote the reference image.

 Let k denote tk, so that f(x,y, k)= f(x, y, tk)

 Assume that R(x,y)=f(x,y,1)
 For any k>1, and the values of the ADIs are
counts,  A ( x, y )  1 R ( x, y )  f ( x, y , k )  T
k 1
A ( x, y )  
k
 A ( x, y ) otherwise
 Absolute k 1

 P ( x, y )  1 R ( x, y )  f ( x, y, k )  T
k 1
P ( x, y )  
 Positive k
 P ( x, y )
k 1 otherwise
 N ( x, y )  1 R ( x, y )  f ( x, y, k )  T
 Negative N ( x , y )  
k
k 1

 N ( x, y )
k 1 otherwise

88
 The Use of Motion in Segmentation
(cont.)
Example 10.19
Image size: 256x256, object szie:75x50, moving in a southeasterly direction
at a speed of 5 2 pixels per frame.
1. Positive ADI 中非零的區域代表移動物體的大小，及移動物體在參考影像中的位置
2. Absolute ADI 包含 positive 和 negative ADI 的區域
3. 移動物體的方向和速度可由 absolute 及 negative ADI 中決定。

Negative
ADI
Absolute
ADI

Positive
ADI 89
The Use of Motion in Segmentation
 Establishing a reference image
 The difference between two images in a dynamic imaging problem has
the tendency to cancel all stationary components, leaving only image
elements that correspond to noise and to the moving objects.
 In practice, obtaining a reference image with only stationary elements is
not always possible, and building a reference from a set of images
containing one or more moving objects becomes necessary.
 One procedure for generating a reference image is as follows
 Consider the first image in a sequence to be the reference image.
 When a nonstationary component has moved completely out of its position in
the reference frame, the corresponding background in the present frame can
be duplicated in the location originally occupied by the object in the reference
frame.
 When all moving objects have moved completely out of their original positions,
a reference image containing only stationary components will have been
created.
 Object displacement can be established by monitoring the changes in the
positive ADI.

90
The Use of Motion in Segmentation (cont.)

行駛中的汽車 行駛中的汽車 汽車已被移除

Frame 1 Frame 2 Result

91
The Use of Motion in Segmentation (cont.)

92
The Use of Motion in Segmentation (cont.)

93
The Use of Motion in Segmentation (cont.)

94
The Use of Motion in Segmentation (cont.)

95