|| Shri Hari ||

BCSE0101: DIGITAL IMAGE PROCESSING

MODULE II

Prof. Charul Bhatnagar

Room no. – 325, AB – I
WA: 7055510537
MODULE II
Morphological Image Processing: Introduction, Logical Operations involving Binary
Images, Dilation and Erosion, Opening and Closing, The Hit-or-Miss Transformation,
Morphological Algorithms – Boundary Extraction, Region Filling, Extraction of
Connected Components, Convex Hull, Thinning, Thickening.
Image Segmentation: Point, Line & Edge detection, Thresholding, Region-based
Segmentation, Region Extraction - Pixel Based Approach & Region Based Approach, Edge
and Line Detection - Basic Edge Detection, Canny Edge Detection, Edge Linking - Hough
Transform.
Representation & Description: Representation - Boundary Following, Chain Codes;
Boundary Descriptors – Shape Numbers.
 BASIC IDEA OF SE

 for each pixel in binary image:

 check if SE is “satisfied”
 output pixel is set to 0 or 1 depending on the operation used

Output is 1 if the SE fits in

Oct 14, 2013
 BASIC COMPONENTS IN MORPHOLOGY
Every Operation has two elements.
1) Input Image

2) Structuring element Origin is not a

member of the
The results of the operation structuring element
mainly depends upon the
structuring element chosen.

Typical Structuring elements

Oct 14, 2013 X denotes the origin of the structuring element
EROSION
 Shrinks or thins objects in a binary image
 Translate structuring element, check if it entirely fits
within foreground of image

 The mathematical defination is

 Let,
A = {(1,0), (1,1), (1,2), (0,3), (1,3), (2,3), (3,3), (1,4)}
B = {(0,0), (1,0)}

Image A Image B
Structuring
Oct 14, 2013 element Erosion
DILATION
 Dilation is an operation that grows or thickens
objects in a binary image.
 The specific manner of this thickening is controlled
by the shape of structuring element.
Dilation combines two sets using vector addition
(a1, a2) + (b1, b2) = (a1 + b1, a2 + b2)
 Let,
A = {(1,0), (1,1), (1,2), (2,2), (0,3), (0,4)}
B = {(0,0), (1,0)}

Image B
Image A
Oct 14, 2013
Structuring element Dilation
 Dilation
(Another way of defining it)

Theoretical way of generation:

 Obtain the reflection of B about its origin
 Shift this reflection by z
 The dilation of A by B then is set of all displacements z, such
that B^ and A overlap by at least one element
Oct 14, 2013
  The reflected structuring element is translated
throughout the domain of the image to see where it
overlaps with 1-valued pixels. The output image is 1
at each location of the origin s.t. the reflected
structuring element overlaps at least one 1-valued
pixel in the input image.

SE

Reflected SE
Oct 14, 2013 Image A Dilation
 DILATION & EROSION ARE DUALS OF EACH OTHER WRT
SET COMPLEMENTATION & REFLECTION

Proof:

If (B)z is contained in set A, then

Oct 14, 2013

 But the complement of the set of z’s that satisfy

is the set of z’s that satisfy

Thus,

Remember???

Oct 14, 2013

 Dilation & Erosion are not inverse transformations
 If an image is eroded & then dilated (or vice-versa), the original
image is not re-obtained
 In practical applications, dilation and erosion are used most
often in various combinations.
 Three of the most common combinations of dilation and
erosion are
 Opening
 Closing
 Hit-or-miss transformation
Oct 14, 2013
 In practical applications, dilation and erosion
are used most often in various combinations.
Three of the most common combinations of
dilation and erosion are
 Opening
 Closing
 Hit-or-miss transformation
Oct 14, 2013
OPENING - EROSION FOLLOWED BY DILATION

Opening
 Generally smoothes the contours of an object
 Breaks narrow isthmuses
 Eliminates thin protrusions
 It is less destructive than the Erosion
GEOMETRICAL INTERPRETATION OF OPENING
The opening of A by B is the union of all the translations of B so
that B fits entirely in A.

a) Image I, composed of set (object) A

and background.
b) Structuring element, B.
c) Translations of B while being
contained in A. (A is shown dark for
clarity.)
d) Opening of A by B.
CLOSING - DILATION FOLLOWED BY EROSION

Closing
 Also tends to smooth sections of the contours
 Generally fuses narrow breaks & long thin gulfs
 Eliminates small holes & fills gaps in the contour
GEOMETRIC INTERPRETATION OF CLOSING

a) Image I, composed of set (object)

A, and background.
b) Structuring element B.
c) Translations of B such that B does
not overlap any part of A. (A is
shown dark for clarity.)
d) Closing of A by B.
Oct 14, 2013
PROPERTIES

a) A ∘ B is a subset of A. a) A is a subset of A B.
b) If C is a subset of D, then b) If C is a subset of D, then
C ∘ B is a subset of C  B is a subset of
D ∘ B. D  B.
c) (A ∘ B) ∘ B = A ∘ B. c) (A  B) B = A  B.
What will be the result of
i. Erosion
ii. Dilation
iii. Opening
iv. Closing
 THE HIT-OR-MISS TRANSFORM (HMT)
♦ HMT is a basic tool for shape detection.
♦ Let I be a binary image composed of
foreground (A) and background pixels,
respectively.
♦ HMT utilizes two structuring elements:
 B1 for detecting shapes in the
foreground
 B2 for detecting shapes in the
background
 B1 = [1 1 1]
1 1 1 1
Eroding the image with a 1 × 3 1 1 1 1 1
mask that corresponds to the shape SE B1 1 1 1
of the object
 all objects that are smaller A
than the target object are
removed
1 1
 But all objects that are larger
than the mask, i. e., where the
shifted mask is a subset of the 1
object are retained
A ⊖ B1
We need a second operation to 1 1 1 1 1 1 1 1 1
1 1 1 1 1
remove all objects larger than the
1 0 0 0 1 1 0 0 0 0 1 1 1 1
target object.
 Can be done by analyzing the 1 1 1 1 1 1 1 1 1 1 1 0 0 1
background of the original SE B2 1 1 0 0 0 1 1 1 1
binary image. 1 1 1 1 1 1 1 1 1
 Use as a second step an
AC
erosion of the background with 0 0 0 0 0 0 0 0 0
a 3 × 5 mask B2 in which all
coefficients are zero except for 0 0 0 0 0 0 0 0 0
the pixels in the background, 0 0 0 0 0 0 0 0 0
surrounding the object. 0 0 0 1 0 0 0 0 0
 This is a negative mask for the 0 0 0 0 0 0 0 0 0
object
AC ⊖ B2
 0 0 0 0 0 0 0 0 0
1 1 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
1  0 0 0 1 0 0 0 0 0
0 0 0 0 0 0 0 0 0
A ⊖ B1 AC ⊖ B2

1 1 1 1
1 1
1 A B 1 1 1
As the hit and miss masks of the hit-miss 0 0 0 0 0
operator are disjunct, they can be combined into 0 1 1 1 0
one mask using a hit (1), miss (0), and don’t care 0 0 0 0 0
(x) notation.

What will be the mask for detecting isolated

points?

What will be the mask for detecting objects with

horizontal rows of 3 to 5 consecutive pixels?
 || Shri Hari ||

Some practical uses

of morphology
BOUNDARY EXTRACTION
1. Erode A by B
2. Perform set difference
between A & its erosion
 EXAMPLE

Input image Output Image

Structure element used is a 3X3 all 1s matrix

Oct 14, 2013
HOLE FILLING /
Hole - a background region surrounded by a connected
border of foreground pixels.

 Let A denote a set whose elements are 8-connected

boundaries, with each boundary enclosing a
background region (i.e., a hole).

 Given a point in each hole, the objective is to fill

all the holes with foreground elements (1’s).

Set A
 All non-boundary points are labelled 0.
Arbitrarily select a point p, inside the p
boundary and assign it value 1
The objective is to fill the entire region
with 1s.
The region is filled with 1s by the
following procedure.
Xk = (Xk-1  B) ∩ Ac for k = 1, 2, …
where X0 = p
Set A
p origin

Structuring element B

Set A Set Ac
 origin

Structuring element B

X0

∩
Oct 14, 2013 Set Ac
 The algorithm terminates at iteration step k if Xk = Xk-1.
The set union of Xk and A contains the filled set & its
boundary.

The dilation process would fill the entire area if left unchecked.
The intersection with Ac at each step limits the result to inside the region
of interest.
EXTRACTION OF CONNECTED
COMPONENTS
Extraction of connected components
Structuring
in a binary image is central to many X element B
automated image analysis applications
Let A be a set and Y be a connected
component in it.
Point p of Y is known.
The following iterative expression p
yields all the elements of Y.
Y

X0 = p
Set A
a) Structuring element. p

b) Image containing a set with

one connected component.
c) Initial array containing a 1
in the region of the
connected component.
d) – (g) Various steps in the
iteration
 CONVEX HULL
♦ Set A is said to be 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 is
the smallest convex set containing S
♦ The set difference H – S is called the
convex deficiency of S
♦ Convex hull & Convex deficiency
are useful for object description
 There are four SEs (convex hull masks)
masks).

1 x x 1 1 1 x x 1 x x x
1 0 x x 0 x x 0 1 x 0 x
1 x x x x x x x 1 1 1 1

B1 B2 B3 B4
  Following equations are iteratively applied

+1 +1

The procedure converges using the ith structuring element when

X ik  X ik 1 we let Di  X ik
Then the convex hull of A is the union of the four results:

4
CA    Di
i 1
Oct 14, 2013
Oct 14, 2013
 1 x x

1 0 x

1 x x

B1

Oct 14, 2013

 1 1 1

x 0 x

x x x

B2

Oct 14, 2013

 x x 1

x 0 1

x x 1

B3

Oct 14, 2013

 x x x

x 0 x

1 1 1

B4

Oct 14, 2013

Oct 14, 2013
 Shortcoming of the procedure -
The convex hull can grow
beyond the minimum dimensions
required to guarantee convexity.

To reduce this effect limit growth

so that it does not extend past
the vertical & horizontal
dimensions of the original set of
points

Oct 14, 2013

 Boundaries of greater complexity can be
used to limit growth even further in images
with more detail.
e.g. the maximum dimensions of the original set of
points along vertical, horizontal & diagonal
directions can be used.
Price paid for such refinements ??
Additional complexity
Oct 14, 2013
THINNING
Thinning is used to remove selected foreground pixels
from binary images
Erase foreground pixels such that
An object without holes erodes to a minimally Original image
connected stroke located equidistant from its
nearest outer boundaries
An object with holes erodes to a minimally
connected ring midway between each hole & its
nearest outer boundary. Thinned image
 The thinning of a set A by a structuring element B can be
defined in terms of the hit-or-miss transform

 A more useful expression for thinning A symmetrically is based

on a sequence of structuring elements:
{B} = {B1, B2, ..., Bn}
where Bi is the rotated version of Bi-1
(a) Structuring elements.
(b) Set A.
(c) Result of thinning
A with B1 (shaded).
(d) Result of thinning
A1 with B2.
(e)–(i) Results of
thinning with the next
six SEs. (There was no
change between A7
and A8 .)
(j)–(k) Result of using
the first four elements
again.
(l) Result after
convergence.
(m) Result converted
to m-connectivity.
 THICKENING
Thickening is the morphological dual of thinning

As in thinning, thickening can be defined as a sequential operation:

The structuring elements used for thickening have the same form as those for
thinning, but with all 1’s and 0’s interchanged.
 A separate algorithm is rarely used for thickening

The usual procedure??

Thin the background of the set in question
& then take the complement of the result

To thicken A
 Let C = Ac
 Thin C
 Take Cc
Oct 14, 2013
 54
IMAGE SEGMENTATION
 Segmentation divides an image into its
constituent regions or objects.
 Segmentation of non trivial images is
one of the difficult task in image
processing. Still under research.
 Segmentation accuracy determines the
eventual success or failure of
computerized analysis procedure.
 Example Application: Automated
inspection of electronic assemblies.
(mother boards)
Nov 11, 2013
REGION
Let v be the spatial domain on which the image is defined. Image
segmentation divides v into n regions, Ri (i = 1 to n) such that
indicates that the segmentation must be complete, in the sense
that every pixel must be in a region.
requires that points in a region be
connected in some predefined sense
deals with the(e.g.,
properties
thethe thatmust
points must
regions be8-connected).
be
must satisfied
be disjointby the
pixels in a segmented region - for example, Q (Ri ) =
TRUE if all pixels in Riindicates
have thethat
sametwo
intensity.
adjacent
regions Ri and Rj must be different
in the sense of predicate Q
POINT, LINE AND EDGE DETECTION

♦ Point
a foreground (background) pixel surrounded by background (foreground)
pixels
♦ Line
a (typically) thin edge segment in which the intensity of the background on
either side of the line is either much higher or much lower than the
intensity of the line pixels.
♦ Edge
a set of connected edge pixels where edge pixels are pixels at which the
intensity of an image changes abruptly
DETECTION OF ISOLATED POINTS
o Point detection is based on the second derivative, the Laplacian

o A point is detected at a location (x, y) on which the kernel is centered if the

absolute value of the response of the filter at that point exceeds a specified
threshold. Such points are labeled 1 and all others are labeled 0 in the output
image, thus producing a binary image.
a) Laplacian kernel used for point
detection.
b) X-ray image of a turbine blade
with a porosity manifested by
a single black pixel.
c) Result of convolving the
kernel with the image.
d) Result of thresholding was a
single point
 LINE DETECTION
o For line detection the second
derivatives result in a stronger
filter response and produce
thinner lines than first derivatives.
o The Laplacian kernel can be used.
The Laplacian image contains
negative values, so absolute value (a) Original image. (b) Laplacian image; the
can be taken for display Magnified section shows the positive/negative
o The double-line effect of the double-line effect characteristic of the
second derivative must be handled Laplacian. (c) Absolute value of the
properly. Laplacian. (d) Positive values of the
Laplacian.
 The Laplacian is isotropic, i.e. independent of direction.
 If we would like to detect lines on a certain direction only we might want
to use masks that would emphasize a certain direction and be less sensitive
to other directions.
Y

Line detection kernels. Detection angles are with respect to the axis system in which
positive angles are measured counterclockwise with respect to the (vertical) x-axis.
EDGE MODELS
Frequently used for segmenting images based on abrupt (local) changes in
intensity.
Edge models are classified according to their intensity profiles.
Step edge
Characterized by a transition between two
intensity levels occurring ideally over the
distance of one pixel.
Step edges occur, for example, in images
generated by a computer for use in areas Ideal representation of a step edge
such as solid modeling and animation. & it’s corresponding intensity profile.
In practice, digital images have edges that are blurred and noisy,
 blurring due to limitations in the focusing mechanism (e.g., lenses in the case
of optical images)
 noise level determined by the electronic components of the imaging system.

Ramp
The slope of the ramp is inversely
proportional to the degree to which the
edge is blurred.
No single “edge point” along the profile.
An edge point now is any point contained in
the ramp Ideal representation of a ramp
and it’s corresponding intensity
An edge segment a set of such points that profile.
are connected.
Roof Edge
Roof edges are models of lines through a
region, with the width of the edge being
determined by the thickness and sharpness of
the line.
Arise, for example, in range imaging, when thin
objects (such as pipes) are closer to the
sensor than the background (such as walls). Ideal representation of a roof
The pipes appear brighter . edge and it’s corresponding
Also appear in the digitization of line drawings intensity profile.
& in satellite images, where thin features, such
as roads, can be modeled by this type of edge.
 The First derivative
– is positive at the onset of the ramp
and at points on the ramp,
– it is zero in areas of constant
intensity.
– the magnitude of the first derivative
can be used to detect the presence
of an edge at a point in an image
a) Two regions of constant intensity separated
by an ideal ramp edge.
b) Detail near the edge, showing a horizontal
intensity profile & its first & second
derivatives.
 The Second derivative
– is positive at the beginning &
negative at the end of the ramp
– zero at points on the ramp & at
points of constant intensity.
– The intersection between the zero
intensity axis and a line extending
between the extrema of the second
a) Two regions of constant intensity separated derivative marks a point called the
by an ideal ramp edge. zero crossing of the second
b) Detail near the edge, showing a horizontal derivative.
intensity profile & its first & second
derivatives.
Properties of Second Derivative
1. The sign of the second derivative can be used to determine whether an
edge pixel lies on the dark or light side of an edge.
2. it produces two values for every edge in an image
3. its zero crossings can be used for locating the centers of thick edges
4. It is very sensitive to noise
 Three steps for edge detection are:
1. Image smoothing
for noise reduction.
2. Detection of edge points
a local operation that extracts from an image all points that are potential
First column: 8-bit images with
edge-point candidates.
values in the range [0,255], and
3. Edge
intensity localization
profiles of a ramp edge
corrupted by Gaussian noise.
to select from the candidate points only the points that are members of
Second column: First-derivative
imagesthe
andset of points
intensity comprising an edge.
profiles.
Third column: Second-derivative
images and intensity profiles.
HOW CAN YOU TELL THAT A PIXEL IS ON AN
EDGE?

We want an Edge Operator that

produces:
 Edge Magnitude

 Edge Orientation

 High Detection Rate and Good

Localization

Real Edges are Noisy and

Discrete!
 THE IMAGE GRADIENT AND ITS PROPERTIES
Edge strength and direction at an
arbitrary location (x, y) of an image, f, is
the gradient, denoted by f and defined as
the vector

The magnitude, M(x, y), of this gradient

vector at a point (x, y) is given by its
Euclidean vector norm

The direction of the gradient vector

at a point (x, y) is given by
♦ An image containing a straight edge segment.
♦ The shaded pixels have value 0, and the white pixels have value 1.
♦ Angles are measured in the counterclockwise direction with respect to
the x-axis
♦ The edge direction is perpendicular to the direction of the gradient vector
at the point where the gradient is computed.
 Find the strength & the direction of the edge at the
highlighted pixel
y
Pixels in gray have value 0 & in white have
value 1.

Derivative is computed by using a 3x3

x
neighbourhood – subtract the pixels in the
top row from the pixels in the bottom
row to get the derivative in the x
direction. Similarly obtain the derivative in
the y direction.
Nov 11, 2013 71
∂f / ∂x = - 2

∂f / ∂y = 2

M(x, y) = 22

= tan-1(gy / gx)
-45°
which is the same as
135° measured in
the positive
(counterclockwise)
direction with respect to
the x-axis
 Roberts Operator (Roberts [1965])
 Used when diagonal edge direction is of interest.
The Roberts cross-gradient operators use 2-D
kernels with a diagonal preference.
 Kernels of size 2 × 2 are simple conceptually, but
they are not as useful for computing edge
direction as compared with kernels that are
symmetric about their centers.
 The smallest such kernel is of size 3 × 3. These
kernels take into account the nature of the data
on opposite sides of the center point, and thus
carry more information regarding the direction
of an edge.
Prewitt Operators (Prewitt [1970])

Sobel Operators (Sobel [1970])

 Examples 9 9 9 9 9 9 9 2 2
d1
9 8 9 9 9 9 2 2 2
9 9 9 9 9 9 3 2 2 0 1
9 9 9 9 9 2 2 2 2  1 0
A 9x9 original  
image is given by 7 9 9 9 9 2 2 2 2
9 9 9 9 2 2 2 2 2 1 0
0 1
1) Use Robert gradient 9 9 9 9 2 2 2 4 2  
operator to find its 9 9 9 2 2 2 2 2 2
edges
9 9 2 2 2 2 1 2 2 d2

d ( x, y )  d12 ( x, y )  d 22 ( x, y )
Use to estimate the
gradient magnitude, and use T = 5 as the threshold for edge detection
Nov 11, 2013 75
 0 1 0 0 0 0 0 0
-1 0 0 0 0 -7 -1 0
0 0 0 0 0 1 0 0 edge map
2 0 0 0 -7 0 0 0

d1 0 0 0 0 0 0 0 0
0 1 0 0 0 -7 0 0 0 -2
 1 1 1 0 0 0 7 7 0 0 0 0 0 0 1 1 0
 0
 0 0 0 0 0 0 2 0
1 1 0 0 0 9.2 1 0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
0 0 0 0 7 7.1 1 0 0 0 0 0 1 1 0 0
original 2 0 0 0 9.9 0 0 0 0 0 0 0 1 0 0 0
image 1 0 0 0 0 7 7 0
2 0 0 7 7 0 0 0 0 0 0 1 1 0 0 0
0 0 0 9.9 0 0 2 2 0 0 0 1 0 0 0 0
0 -1 0 0 0 6 0 0
0 0 7 7 0 0 2 2 0 0 1 1 0 0 0 0
0 0 0 0 7 7 1 0
0 0 0 0 7 0 0 0 0 7 7 0 0 1 1 0 0 1 1 0 0 0 0 0
1 0
d
0 21
 
-2 0 0 7 7 0 0 0
0 0 0 7 0 0 -2 0
thresholding
0 0 7 7 0 0 0 2
0 7 7 0 0 1 0 0
T=5
Nov 11, 2013
d 2 ( x, y ) Note: 8x8 images (because of boundary problems)
SIMPLE EDGE
DETECTION
USING
GRADIENTS
…
THE CANNY EDGE DETECTOR
Algorithm more Performance
complex superior

Canny’s approach based on 3 basic objectives:

Low error rate

 All edges should be found
 There should be no spurious responses
Edge points should be well localized
 The edges detected should be as close as possible to the true
edges.
 The distance between a point marked as an edge by the detector &
the centre of the true edge should be minimum.

Single Edge point response

 The detector should return only one point for each true edge
point.
 The number of local maxima around the true edge point should be
minimum.
 The detector should not identify multiple edge pixels where only a
single edge point exists.
OPTIMAL EDGE DETECTION: CANNY …
image
Canny
smoothing with Gaussian
Edge
Detector
differentiating along x and y axis

finds peaks in gradient image

thresholding

thinning (locating edge strings)

edge map
 ALGORITHM

Step 1
Let f (x, y) denote the input image and G(x, y) denote
the Gaussian function:

A smoothed image, f s(x, y), is formed by convolving f The image used as

and G: example of Canny edge
detection.
The kernel of a Gaussian
filter with a standard
deviation of σ = 1.4

The original grayscale image is smoothed with

a Gaussian filter to suppress noise.
Step 2
Apply first difference gradient operator to compute the edge strength &
direction. Any of the filter masks (Roberts, Prewitt etc.) can be used.

The gradient magnitudes in the

smoothed image as well as their
directions are determined by applying
e.g. the Sobel-operator
Step 3
Apply non-maximal suppression to the gradient
magnitude.
 The purpose of this step is to convert the
“blurred” edges in the image of the gradient
magnitudes to “sharp” edges.
 Basically this is done by preserving all local
maxima in the gradient image, and deleting
everything else.
The algorithm is for each
pixel in the gradient image:
1. Round the gradient direction
θ to nearest 45°
 in a 3 × 3 region four
orientations can be defined Two possible orientations of a
horizontal edge
for an edge passing
through the center point The angle ranges of
of the region: horizontal, the edge normals for
vertical, +45°& −45°. the four types of edge
directions in a 3 × 3
 Every edge has two neighborhood. Each
possible orientations. edge direction has
two ranges
2. Compare the edge strength of the
current pixel with the edge
strength of the pixel in the positive
and negative gradient direction.
 i.e. if the gradient direction is north
(θ = 180º ), compare with the pixels
to the north and south Almost all pixels have gradient
 If the edge strength of the current directions pointing north. They are
pixel is largest; preserve the value of therefore compared with the
the edge strength. If not, suppress pixels above and below. The pixels
(i.e. remove) the value. that turn out to be maximal in this
 gN(x, y) is the non maximal edge comparison are marked with
suppressed image white borders. All other pixels will
be suppressed.
Non-maximum suppression. Edge-pixels are only preserved
where the gradient has local maxima.
The final operation is to threshold gN(x, y) to reduce false edge
points.
False edges can be reduced by applying a threshold T
 all values below T are changed to 0
However,
 selecting a good value for T is difficult
 some false edges will remain if T is too low
 some edges will disappear if T is too high
Step 4
Hysteresis thresholding / Edge Linking
 Canny’s approach employs double thresholding, known as hysteresis.

 a low threshold, TL and a high threshold, TH.

 Experimental evidence (Canny [1986]) suggests that the ratio of the high
to low threshold should be in the range of 2:1 to 3:1.
 Edge pixels stronger than TH are marked as strong; edge
pixels weaker than TL are suppressed and edge pixels between
the two thresholds are marked as weak.
Thresholding of edges. In the second image strong edges are
white, while weak edges are grey. Edges with a strength
below both thresholds are suppressed.
 Strong edges are interpreted as “certain edges” and can immediately be
included in the final edge image.
 Weak edges are included if and only if they are connected to strong edges.
 The logic is that noise and other small variations are unlikely to result in
a strong edge. Thus strong edges will (almost) only be due to true edges
in the original image.
 The weak edges can either be due to true edges or noise/color
variations. The latter type will probably be distributed independently of
edges on the entire image, and thus only a small amount will be located
adjacent to strong edges.
 Weak edges due to true edges are much more likely to be connected
directly to strong edges.
Edge tracking and final output. The middle image shows strong edges in white, weak
edges connected to strong edges in blue, and other weak edges in red
An edge element is deemed present at (x,y) if f’(x, y) exceeds a
predefined threshold

But we haven’t found edge segments, only edge points

How can we find and describe more complex features?
The Hough transform is a common approach to finding
parameterised line segments (here straight lines) or other
shapes
HOUGH TRANSFORM
 Performed after Edge Detection
 It is a technique to isolate the curve of a given shape in a given
image
 Classical Hough Transform can locate regular curves like straight
lines, circles, parabolas, ellipses, etc.
o Requires that the curve be specified in some parametric form

 Generalized Hough Transform can be used where a simple

analytic description of feature is not possible
• Given marked edge pixels, find examples of specific
shapes
– Line segments
– Circles
– Generalized shapes
• Basic idea - Patented 1962
– Every edge pixel is a point that votes for all shapes that pass
through it.
– Votes are collected in “parameter space” - look for peaks
 y = mx + c where m is the slope & c the y-intercept
The equation can be rewritten in c – m space as: c = -xm + y
Suppose we have several edge points (x1, y1), ..., (xn, yn) in x – y space that we
want to fit in a line.
Each point in x – y space maps to a line in c – m space.

Aug 01, 2016 96

Properties in slope-intercept space
 Each point (xi , yi) defines a line in the c − m space (parameter
space).

 Points lying on the same line in the x − y space, define lines in

the parameter space which all intersect at the same point.

 The coordinates of the point of intersection define the

parameters of the line in the x − y space.
ALGORITHM
1. Quantize parameter space (c, m):
P [mmin, . . . , mmax] [cmin, . . . , cmax]
2. For each edge point (x, y)
For(m = mmin; m ≤ mmax; m++) {
c = − xm + y; /* round off if needed *
(P[m][c])++; /* voting */ }
3. Find local maxima in P[m][c]

if P[m][
][cc] = M, then M points lie on the line y = m x + c
Find the equation for the edge for the following edge
points
• (2,2) c = -2m + 2
• (4,7) c = -4m + 7
• (5,5) c = -5m + 5
• (6,2) c = -6m + 2
• (6,6) c = -6m + 6

Let mmin = -2 & mmax = 2

 (2,2) (4, 7) (6, 2)
(5,5) (6,6)
m c = -2m + c = -4m + c = -6m +
c = -5m +5 c = -6m +6
2 7 2
-2 6 15 15 14 18
-1 4 11 10 8 12
0 2 7 5 2 6
1 0 3 0 -4 0
2 -2 -1 -5 -10 -6
Point Votes So, the equation becomes
(-2, 15) 2 y=x
(0, 2) 2
(1, 0) 3
 ♦ But there is a Problem
♦ The slope of line parallel to y –
axis is infinity, the computer can
not handle it.
♦ So, we follow another
parameterization of a line

p = x cos θ + y sinθ
sinθ
Aug 01, 2016 101
ALGORITHM
1. Quantize the parameter space (p, θ)
P[pmin, . . . , pmax][θmin, . . . ,θmax]
2. For each edge point (x, y)
For(θ = θmin; θ ≤ θmax; θ ++) {
p = x cos(θ) + y sin(θ) ; /* round off if needed *
(P[p][θ])++; /* voting */
}
3. Find local maxima in P[p][θ]
 y
6
5
Use
4 Hough Transform to form the
3
2
main edge.
1 Work with polar coordinates
0
0 1 2 3 4 5 6 x

p = x cos θ + y sinθ
Edgels – (0,5), (1,4), (3,2), (3,3), (5,0)
Aug 01, 2016 103
 In the given question, the curve corresponding to the point (3,3)
does not intersect at the common point (3.5, 45), where the other
curves meet.

The equation for the edge will be:

3.5 = x cos 45 + y sin 45

Aug 01, 2016 104

SEGMENTATION
Different objects or parts of the same object
have different characteristics

So

Feature values recorded at pixels belonging to different

regions should be different
 BASIC IDEA OF IMAGE SEGMENTATION

All the image segmentation methods assume that:

1. the intensity values are different in different
regions, and,
2. within each region, which represents the
corresponding object in a scene, the intensity
values are similar.
THRESHOLDING
Gray value
Simplest feature in a gray level
image
♦ Image can be segmented by gray level thresholding
♦ The simplest segmentation process
♦ Computationally inexpensive and fast.
♦ Can easily be done in real time
♦ Correct thresholding leads to better segmentation.
 THE BASICS OF INTENSITY THRESHOLDING
Suppose image f (x, y) is composed of
light objects on a dark background.
How will its Histogram look?

To extract the objects from the

background select a threshold, T, that
separates these modes.
 Global Thresholding
 T is a constant applicable over an entire image.
 Apply the same threshold to the whole image
 Local Thresholding
 The value of T changes over an image.
 The value of T at any point (x, y) in an image depends on properties of
a neighborhood of (x, y) (for example, the average intensity of the
pixels in the neighborhood).
 Dynamic (Adaptive) Thresholding
 The threshold depends on spatial coordinates (x,y)
 Muliple Thresholding
 If an image has more than two dominant modes, then multiple
thresholds will be required.
 Imagine a poker playing robot that needs to visually
interpret the cards in its hand

Original Image Thresholded Image

 If you get the threshold wrong the results can be
disastrous

Threshold Too Low Threshold Too High

BASIC GLOBAL THRESHOLDING

1. Select an initial estimate for T (typically the average grey

level in the image)
2. Segment the image using T to produce two groups of
pixels: G1 consisting of pixels with grey levels >T and G2
consisting pixels with grey levels ≤ T
3. Compute the average (mean) intensity values m1 and m2
for the pixels in G1 and G2 , respectively.
4. Compute a new threshold value midway between m1
and m2 :
1
T m1  m2 
2

5. Repeat Steps 2 through 4 until the difference between

values of T in successive iterations is smaller than a
predefined value, T.

 This algorithm works very well for finding thresholds

when the histogram is suitable
The objective of segmentation is to partition an image into
regions.
We did it by
 attempting to find boundaries between regions based on
discontinuities in intensity levels,
 using thresholds based on the distribution of pixel properties,
such as intensity values or color.

We will now discuss segmentation techniques that find the

regions directly.
SEGMENTATION BY REGION GROWING

Groups pixels or subregions into larger regions based on

predefined criteria for growth.

The basic approach is to start with a set of “seed” points, and

from these grow regions by appending to each seed those
neighboring pixels that have predefined properties similar to
the seed (such as ranges of intensity or color or texture etc).
Let
f (x, y) denote an input image;
S(x, y) denote a seed array containing 1’s at the locations of
seed points and 0’s elsewhere; and
Q denote a predicate to be applied at each location (x, y).
Arrays f and S are assumed to be of the same size.
A basic region-growing algorithm based on 8-connectivity may be
stated as follows.
1. Start from a seed marking it as a new pixel for the region.

2. Check for 8 neighbours of the newly added pixels. If they satisfy the
homogeneity criteria, add them to the region

3. Once a pixel is accepted as a member of the current region, the

nearest neighbours of this new pixel are examined.

4. The process goes on recursively until no more pixels are accepted.

5. All pixels in the current region are marked with a unique label.

6. Then another seed pixel is picked up & the procedure is repeated.

MAIN PROBLEMS OF REGION GROWING
Large execution time

The selection of the property to be satisfied

The selection of the seed points

Homogeneity property for a region can be defined as

9 6 6 6 7 7 6 6
Find the regions for 6 7 6 7 5 5 4 7
the given seeds. 6 6 4 4 3 2 5 6
Threshold value = 3
5 4 5 4 2 3 4 6
1 3 2 3 3 2 4 7
0 0 1 0 2 2 5 6
1 1 0 1 0 3 4 4
1 0 1 0 2 3 5 6
REGION SPLITTING (TOP DOWN APPROACH)
REGION MERGING (BOTTOM UP METHOD)
 Start from pixel level. Consider each of them as a homogeneous region.

 At any level of merging, check if four adjacent homogeneous regions

arranged in 2 x 2 fashion, together satisfy the homogeneity property. If yes,
then merge them into a single region; otherwise leave the regions as such.

 Repeat the operation recursively until there are no more regions that can
be merged.
 5 6 6 6 7 7 6 6 Find the regions based
6 7 6 7 5 5 4 7 on
6 6 4 4 3 2 5 6
5 4 5 4 2 3 4 6 1. Region splitting
1 3 2 3 3 2 4 7 technique
0 0 1 0 2 2 5 6
1 1 0 1 0 3 4 4 2. Region merging
1 0 1 0 2 3 5 6 technique

An 8x8 image with gray levels Threshold value = 3

ranging from 0 to 7
Nov 11, 2013 124
 FEATURE EXTRACTION
The resulting sets of segmented pixels have to be converted into a form suitable
for further computer processing.
 Step after segmentation - Feature Extraction, which consists of
 Feature detection - finding the features in an image, region, or boundary.
 For example, detect corners in a region boundary
 Feature description - assigning quantitative attributes to the detected
features.
 E.g. describe those corners by their orientation and location, both of
which are quantitative attributes
BOUNDARY FOLLOWING (TRACING)
♦ Many algorithms require that the points in the boundary of a
region be ordered in a clockwise or counterclockwise
direction.
♦ Boundary-following algorithm gives an ordered sequence of points.
 The image should be binary in which object and background points
are labeled 1 and 0, respectively
 The image is padded with a border of 0’s to eliminate the
possibility of an object merging with the image border.
 MOORE BOUNDARY TRACING ALGORITHM
Input: A binary image containing a connected
component P of black cells.
Output: A sequence B (b0, b1, ..., bk) of boundary
pixels i.e. the contour.

1. Let the starting point, b0 , be the uppermost-leftmost point

in the image that is labeled 1. Let c0 the west neighbor of b0
(it is always a background point). Examine the 8-neighbors
of b0 , starting at c0 and proceeding in a clockwise direction.
Let b1 denote the first neighbor encountered whose value
is 1, and let c1 be the (background) point immediately
preceding b1 in the sequence.
2. Let b = b0 and c = c0. Insert b in B
3. Let the 8-neighbors of b, starting at c and proceeding in a clockwise
direction, be denoted by n1,n2 ,…, n8. Find the first neighbor labeled 1 and
denote it by nk .
4. Let b = nk and c = nk –1. Insert b in B
5. Repeat Steps 3 and 4 until b = b0.
Examples of boundaries that can be processed by the boundary-following algorithm.
(a) Closed boundary with a branch. (b) Self-intersecting boundary. (c) Multiple
boundaries (processed one at a time).

If we start with a binary region instead of a boundary, the algorithm extracts

the outer boundary of the region. Typically, the resulting boundary will be one
pixel thick, but not always
 FREEMAN CHAIN CODES
Why focus on a boundary?
The boundary is a good
representation of an object shape and
also requires less memory.

Chain codes: represent an object

boundary by a connected sequence of Direction numbers for
straight line segments of specified length (a) 4-directional chain code
and direction. (b) 8-directional chain code
The direction of each segment is coded by
using a numbering scheme shown below
Find the 4 directional Chain Code

1101101030332330322212

3
PROBLEMS WITH CHAIN CODES
 the resulting chain of codes tends To use boundary representation
to be quite long in object recognition, we need to
 any small disturbances along the achieve invariance to starting
boundary due to noise or imperfect point and orientation
segmentation cause changes in the  Normalized codes
code that may not be related to the
shape of the boundary  Differential codes
 Dependent on the starting point
 Dependent on the orientation
NORMALIZATION STRATEGY
TO OVERCOME THE STARTING POINT PROBLEM

33001122 00112233
30011223 Sort 01122330
00112233 rows 11223300
01122330 12233001
11223300 22330011
12233001 23300112
22330011 30011223
23300112 33001122

First row gives the

normalized chain code 00112233
33001122
 DIFFERENTIAL STRATEGY
NORMALIZATION FOR ROTATION
 Counting (counterclockwise) the number of direction changes that separate two
adjacent element of the code
 Assuming the first difference code represents a closed path, rotation normalization
can be achieved by circularly shifting the number of the code so that the list of
numbers forms the smallest possible integer.
90o

33001212 33010122
01011311
01011311 01131101
SHAPE NUMBERS

 The shape number of a boundary obtained from a chain

code is defined as the smallest magnitude of the circular
first difference.

 The order of the shape number is defined as the

number of digits in its representation.
Find the shape number & order
of the given boundary.
4-direction chain code 03032211

First difference 3133030

Circular first 33133030

difference

Shape number 03033133

Order 8