Point Operations and Spatial Filtering

Ranga Rodrigo

November 13, 2011

1/102
1 Point Operations
Histogram Processing

2 Spatial Filtering
Smoothing Spatial Filters
Sharpening Spatial Filters

3 Edge Detection
Line Detection Using the Hough Transform

2/102
Outline

1 Point Operations
Histogram Processing

2 Spatial Filtering
Smoothing Spatial Filters
Sharpening Spatial Filters

3 Edge Detection
Line Detection Using the Hough Transform

3/102
We can process images either in spatial domain or in transform
domains. Spatial domain refers to image plane itself. In spatial domain
processing, we directly manipulate pixels. Spatial domain processing
is common. Frequency domain and wavelets are examples of
transform domains.

4/102
Domains

Image enhancement

Spatial domain Transform domain

5/102
Spatial Domain Processing

Spatial domain processing

Range operations Domain operations

Point operations Neighborhood operations

6/102
Point Operations

£ ¤
g (x 0 , y 0 ) = T f (x 0 , y 0 ) .
These are known as gray level transformations.
Examples:
Gamma correction.
Window-center correction.
Histogram equalization.

7/102
Neighborhood Operations

g (x 0 , y 0 ) = T(x,y)∈N f (x, y) , where N is some neighborhood of

£ ¤

(x 0 , y 0 ).
Examples:
Filtering: mean, Gaussian, median etc.
Image gradients etc.

8/102
Domain Operations

g (x 0 , y 0 ) = f (T x (x 0 , y 0 ), T y (x 0 , y 0 )).
Examples:
Warping.
Flipping, rotating, etc.
Image registration.

9/102
Point Operations: Recapitulation

£ ¤
g (x 0 , y 0 ) = T f (x 0 , y 0 ) .
These are known as gray level transformations.
The enhanced value of a pixel depends only on the original value
of the pixel.
If we denote the value of the pixel before and after the
transformation as r and s , respectively, the above expression
reduces to
s = T (r ). (1)
For example, s = 255 − r gives the negative of the image.

10/102
 L −1

3
4L

1
2L

1
4L

0
0 1 1 3 L −1
4L 2L 4L

Figure 1: Intensity transformation function for image negative.

11/102
Example
Write a program to generate the negative of an image.

12/102
im = imread ( ‘ image . jpg ’ ) ;
imneg = 255 − im ;
subplot (1 ,2 ,1)
imshow ( im )
t i t l e ( ‘ Original ’ )
subplot (1 ,2 ,2)
imshow ( imneg )
t i t l e ( ‘ Negative ’ )

13/102
Power-Law Transformations

s = cr γ , (2)
where c and γ are positive constants.
Values of γ such that 0 < γ < 1 map a narrow range of dark input pixels
vales into a wider range of output values, with opposite
being true for higher values of input levels.
Values γ > 1 have the opposite behavior to above.
c = γ = 1 gives the identity transformation.
Gamma correction is an application.

14/102
 L −1
γ = 0.1
3
4L γ = 0.5

1
2L

1 γ=2
4L

0
0 1 1 3 L −1
4L 2L 4L

Figure 2: Intensity transformation function using power law.

15/102
Example
Write a program to carry out power-law transformation on an image.

16/102
r = imread ( ‘ image . jpg ’ ) ;
c = 1;
gamma = 0 . 9 ;
s1 = c * r . ^gamma ;
gamma = 1 . 2 ;
s2 = c * r . ^gamma ;
subplot (1 ,3 ,1)
imshow ( r )
t i t l e ( ’Original ’ )
subplot (1 ,3 ,2)
imshow ( s1 )
t i t l e ( ‘Gamma = 0 . 9 ’ )
subplot (1 ,3 ,3)
imshow ( s2 )
t i t l e ( ‘Gamma = 1 . 2 ’ )

17/102
Piecewise-Linear Transformation Functions

These functions can be used to do gray level operations such as these:

Contrast stretching.
Window-center correction to enhance a portion of levels.
Gray-level slicing.

18/102
L −1

3
4L

1
2L

1
4L T (r )

0
0 1 1 3 L −1
4L 2L 4L

Figure 3: Contrast stretching function.

19/102
Here is an example of a piecewise linear transformation.
im = imread ( ’Images / airplane .jpg ’ ) ;
r = rgb2gray ( im ) ;
line1 = 0:0.5:100;
l i n e 2 = 1 5 5 / 5 5 * ( [ 2 0 1 : 1 : 2 5 5 ] − 200) + 100;
t = [ line1 , line2 ]
plot ( t )
s = t ( r + 1);
subplot (1 ,2 ,1)
imshow ( r )
t i t l e ( ’Original ’ )
subplot (1 ,2 ,2)
imshow ( mat2gray ( s ) )
t i t l e ( ’Output ’ )

20/102
Example
Write a program to carry out contrast stretching.

21/102
Outline

1 Point Operations
Histogram Processing

2 Spatial Filtering
Smoothing Spatial Filters
Sharpening Spatial Filters

3 Edge Detection
Line Detection Using the Hough Transform

22/102
Histogram

The histogram of a digital image with gray levels in the range

[0, L − 1] is a discrete function h(r k ) = n k , where r k is the k th gray
level and n k is the number of pixels having gray level r k .
We can normalize the histogram by dividing by the total number of
pixels n . Then we have an estimate of the probability of
occurrence of level r k , i.e., p(r k ) = n k /n .

23/102
Number of Bins in a Histogram

The histogram that we described above has L bins.

We can construct a coarser histogram by selecting a smaller
number of bins than L . Then several adjacent values of r will be
counted for a bin.

24/102
Example
Write a program to compute the histogram with a given number of bins.

25/102
 close a l l
im = imread ( ’fruits .jpg ’ ) ;
img = rgb2gray ( im ) ;
imshow ( img )
numbins = 1 0 ;
binbounds = l i n s p a c e ( 0 , 2 5 5 , numbins + 1 ) ;
cumhist = zeros ( numbins + 1 , 1 ) ;
f o r i = 2 : numbins + 1
cumhist ( i ) = sum ( sum ( img <= binbounds ( i ) ) ) ;
endfor
h i s t = cumhist ( 2 : end ) − cumhist ( 1 : end − 1 ) ;
b i n c e n t e r s = ( binbounds ( 2 : end )
+ binbounds ( 1 : end − 1 ) ) / 2 ;
bar ( b i n c e n t e r s ’ , h i s t , 0 . 2 )
d

26/102
Histogram Equalization

We saw that a given image may have a peaked histogram. Such a

peaked or non-flat histogram shows that pixels do not equally occupy
all the gray levels. This results in a low contrast image. We can remedy
this by carrying out the operation called histogram equalization.
Histogram equalization is a gray-level transformation that results in an
image with a more or less flat histogram.

27/102
Consider, for now, that continuous intensity values of an image are to
be processed. We assume that r ∈ [0, L − 1]. Lets consider the intensity
transformation
s = T (r ) 0 ≤ r ≤ L − 1 (3)
that produces an output intensity level s for every pixel in the input
image having intensity r . We assume that
T (r ) is monotonically increasing in the interval 0 ≤ r ≤ L − 1, and
0 ≤ T (r ) ≤ L − 1 for 0 ≤ r ≤ L − 1.
The intensity levels in the image may be viewed as random variables in
the interval [0, L − 1]. Let p r (r ) and p s (s) denote the probability density
functions (PDFs) of r and s . A fundamental result from basic probability
theory is that if p r (r ) and T (r ) are known, and T (r ) is continuous and
differentiable over the range of values of interest, then the PDF of the
transformed variable s can be obtained using the simple formula
¯ ¯
¯dr ¯
p s (s) = p r (r ) ¯¯ ¯¯ . (4)
ds

28/102
Now let’s consider the following transform function:
Z r
s = T (r ) = (L − 1) p r (w)d w, (5)
0

where w is the dummy variable of integration. The right-hand side of

this equation is the cumulative distribution function (CDF) of random
variable r . This function satisfies the two conditions that we
mentioned.
Because PDFs are always positive and the integral of a function is
the area under the curve, Equation 5 satisfies the first condition.
This is because the area under the curve cannot decrease as r
increases.
When the upper limit in this equation is r = L − 1, the integral
evaluates to 1, because the area under a PDF curve is always 1.
So the maximum value of s is L − 1, and the second condition is
also satisfied.

29/102
To find p s (s) corresponding to this transformation we use Equation 3.

d s d T (r )
= ,
dr dr
d
·Z r ¸
= (L − 1) p r (w)d w , (6)
dr 0
= (L − 1)p r (r ).

Substituting this result in Equation 3, and keeping in mind that all

probability values are positive, yields
¯ ¯
¯dr ¯
p s (s) = p r (r ) ¯¯ ¯¯ ,
ds
¯ ¯
¯ 1 ¯
= p r (r ) ¯
¯ ¯, (7)
(L − 1)p r (r ) ¯
1
= 0 ≤ s ≤ L − 1.
L −1
We recognize the form of p s (s) in the last line of this equation as a
uniform probability density function.
30/102
For discrete values, we deal with probabilities (histogram values) and
the summation instead of probability density functions and integrals.
The probability of occurrence of intensity level r k in a digital image is
approximated by
nk
p r (r k ) = k = 0, 1, . . . , L − 1, (8)
MN

where M N is the total number of pixels in the image, n k is the number

of pixels that have intensity r k , and L is the number of possible
intensity levels in the image (e.g., 256). Recall that a plot of p r (r k )
versus r k is commonly referred to as a histogram.
The discrete form of the Equation 4 is
k
X
s k = T (r k ) = (L − 1) p r (r j ),
j =0
(9)
k
(L − 1) X
= nj k = 0, 1, . . . , L − 1.
M N j =0

31/102
Thus the output image is obtained by mapping each pixel in the input
image with intensity r k into a corresponding pixel level s k in the output
image using Equation 9.

32/102
Example
Suppose that a 3-bit image (L = 8) of size 64 × 64 pixels (M N = 4096)
has the intensity distribution shown in Table 1, where the intensity
levels are in the range [0, L − 1] = [0, 7]. carry out histogram equalization.

33/102
 rk nk p r (r k ) = n k /M N
r0 = 0 790 0.19
r1 = 1 1023 0.25
r2 = 2 850 0.21
r3 = 3 656 0.16
r4 = 4 329 0.08
r5 = 5 245 0.06
r6 = 6 122 0.03
r7 = 7 81 0.02
Table 1: Table for Example 5.

34/102
Example
White a program to carry out histogram equalization.

35/102
img = imread ( ’Images /Fig3 .15(a)1 top.jpg ’ ) ;
L = 256;
cumhist = zeros ( L , 1 ) ;
f o r k = 0 : L−1
s ( k +1) = sum ( sum ( img <= k ) ) ;
endfor
n = s i z e ( img , 1 ) * s i z e ( img , 2 )
s = s/n;
imeq = mat2gray ( s ( img + 1 ) ) ;
subplot (2 ,2 ,1)
imshow ( img )
subplot (2 ,2 ,2)
imshow ( imeq )
subplot (2 ,2 ,3)
i m h i s t ( im2double ( img ) )
subplot (2 ,2 ,4)
i m h i s t ( imeq )

36/102
Outline

1 Point Operations
Histogram Processing

2 Spatial Filtering
Smoothing Spatial Filters
Sharpening Spatial Filters

3 Edge Detection
Line Detection Using the Hough Transform

37/102
Spatial filtering is one of the main tools that we use in image
processing.
There are many applications including image enhancement, e.g.,
smoothing.
We can accomplish effects such as smoothing by applying a
spatial filter directly on the image.
Spatial filters are also called spatial masks, kernels, templates,
and windows.

38/102
Applying a Spatial Filter

In the spatial filter operation we apply a predefined operation on the

image pixels included in the neighborhood of the pixel in question.
Filtering gives the value of the corresponding pixel at the coordinates
of the center of the neighborhood, in the resultant image. If the
operations performed on the imaged pixels is linear, then the filter is
called a linear filter. Otherwise, the filter is non-linear. Figure 4 shows
a sketch of the neighborhood of a particular pixel. Figure 5 shows the
weights of a 5 × 5 kernel.

39/102
 0 j0 w −1
j
0

i0

h −1

Figure 4: A 3 × 3 neighborhood for spatial filtering. The center if the

neighborhood is the pixel (i 0 , j 0 ).

40/102
w(−2, −2) w(−2, −1) w(−2, 0) w(−2, 1) w(−2, 2)

w(−1, −2) w(−1, −1) w(−1, 0) w(−1, 1) w(−1, 2)

w(0, −2) w(0, −1) w(0, 0) w(0, 1) w(0, 2)

w(1, −2) w(1, −1) w(1, 0) w(1, 1) w(1, 2)

w(2, −2) w(2, −1) w(2, 0) w(2, 1) w(2, 2)

Figure 5: A 5 × 5 kernel showing weights.

41/102
Consider the 3 × 3 kernel shown in Figure 6.

w(−1, −1) w(−1, 0) w(−1, 1)

w(0, −1) w(0, 0) w(0, 1)

w(1, −1) w(1, 0) w(1, 1)

Figure 6: A 3 × 3 kernel showing weights.

42/102
At any pixel (i , j ) in the image, the response g (i , j ) of the filter is the
sum of products of the filter coefficients and the image pixels
encompassed by the filter:

g (i , j ) = w(−1, −1) f (i − 1, j − 1) + w(−1, 0) f (i − 1, j )

(10)
+ · · · + w(1, 1) f (i + 1, j + 1).

Observe that the center coefficient of of the filter, w(0, 0) aligns with the
pixel at the location (i , j ). For a mask of size m × n , we assume that
m = 2a + 1 and n = 2b + 1, where a and b are positive integers. This
means that we always choose filters of odd dimensions for
convenience. If general, linear spatial filtering for an image of size
M × N , with the filter of size m × n is given by the expression
a
X b
X
g (i , j ) = w(s, t ) f (i + s, j + t ), (11)
s=−a t =−b

where i and j are varied so that each pixel in f is visited. Of course,

when the kernel is not fully within the image, we have to use methods
such as zero padding.
43/102
Example
Give a 3 × 3 kernel that can be used for averaging a 3 × 3 neighborhood.

44/102
Instead of using equal 1/9 entries, we can have a 3 × 3 kernel of all
ones and then divide the filter output by 9.

45/102
Example
Write a program to average filter an image using a 3 × 3 kernel.

46/102
w = 3;
h = 1 / 9 * ones (w, w ) ;
hw = f l o o r (w / 2 ) ;
imrgb = imread ( ’Images / airplane .jpg ’ ) ;
im = im2double ( rgb2gray ( imrgb ) ) ;
[ row , c o l ] = s i z e ( im )
r e s u l t = zeros ( row , c o l ) ;
f o r i =hw+1: row −hw
f o r j =hw+1: c o l −hw
result ( i , j ) =
sum ( sum ( h . * im ( i −hw : i +hw , j − hw : j + hw ) ) ) ;
end
end
figure ;
imshow ( r e s u l t ) ;

47/102
A faster and convenient implementation of the aforementioned loops is
as follows:
r e s u l t = conv2 ( im , h , " same " ) ;

48/102
Outline

1 Point Operations
Histogram Processing

2 Spatial Filtering
Smoothing Spatial Filters
Sharpening Spatial Filters

3 Edge Detection
Line Detection Using the Hough Transform

49/102
Smoothing filters are used for blurring and noise reduction. Blurring is
used as a preprocessing operation to remove small noise-like objects
before large object extraction and bridging small gaps in lines and
curves. Noise reduction can be achieved by blurring with a linear filter
and by nonlinear filtering.

50/102
Examples of Smoothing Filters

The averaging filter that we studied earlier is a smoothing filter. This

reduces sharp transitions and noise. Figure 7 shows this kernel. Note
that using 1 as filter coefficients instead of 1/9 and later multiplying the
sum by 1/9 is computationally efficient.

1 1 1

1
× 1 1 1
9

1 1 1

Figure 7: 3 × 3 averaging smoothing filter.

51/102
The weighted averaging kernel shown in Figure 8 gives more
importance to pixels close to the center.

1 2 1

1
× 2 4 2
16

1 2 1

Figure 8: 3 × 3 averaging smoothing filter.

52/102
Example
Write a program to carry out weighted averaging using the kernel in
Figure 8.

53/102
Order-Statistics (Non-Linear) Filters

Order statistics filters determine the value of the resultant pixel by

a value related to the ranking or order of the pixels within a kernel.
The best known in this category is the median filter, which
computes the value of the resultant pixel as the median of the
neighborhood.
Median filters are very effective in the presence of impulse noise
called salt-and-pepper noise.
In order to do median filtering, we sort the pixels in the
neighborhood, find the median, and set the pixel of the resultant
image with the median value.

54/102
Outline

1 Point Operations
Histogram Processing

2 Spatial Filtering
Smoothing Spatial Filters
Sharpening Spatial Filters

3 Edge Detection
Line Detection Using the Hough Transform

55/102
We were able to achieve image blurring by using pixel averaging,
an operation similar to integration.
Sharpening, the opposite of blurring, can be achieved by spatial
differentiation.
The strength of the response of the derivative operator is
proportional to the degree of intensity discontinuity in the image at
the point of interest.
Thus, image differentiation enhances the edges and other
discontinuities, while deemphasizing areas of low intensity
variations.

56/102
Using the Second Derivative for Sharpening—The
Laplacian

We can use a 2-D, second order derivative for image sharpening.

We will use the discrete formulation of the second-order derivative
and construct a filter mask.
We are interested in isotropic filters, whose response is
independent of the direction of discontinuities.
Isotropic filters are invariant to the image rotation, or simply,
rotation invariant.

57/102
Laplacian

A simple isotropic second-order derivative operator is the Laplacian.

For a function (image) f (x, y) of two variables it is defined as

∂2 f ∂2 f
∇2 f = + . (12)
∂x 2 ∂y 2

The second-order derivative in x -direction is

∂2 f
= f (x + 1, y) + f (x − 1, y) − 2 f (x, y). (13)
∂x 2
In y -direction, we have

58/102
Laplacian

A simple isotropic second-order derivative operator is the Laplacian.

For a function (image) f (x, y) of two variables it is defined as

∂2 f ∂2 f
∇2 f = + . (12)
∂x 2 ∂y 2

The second-order derivative in x -direction is

∂2 f
= f (x + 1, y) + f (x − 1, y) − 2 f (x, y). (13)
∂x 2
In y -direction, we have

∂2 f
= f (x, y + 1) + f (x, y − 1) − 2 f (x, y). (14)
∂y 2

58/102
Therefore, the discrete version of the Laplacian is

∇2 f = f (x + 1, y) + f (x − 1, y) + f (x, y + 1) + f (x, y − 1) − 4 f (x, y). (15)

Figure 9 shows some implementations of the Laplacian.

59/102
0 1 0 1 1 1

1 -4 1 1 -8 1

0 1 0 1 1 1

0 -1 0 -1 -1 -1

-1 4 -1 -1 8 -1

0 -1 0 -1 -1 -1

Figure 9: Laplacian kernels.

60/102
Because the Laplacian is a derivative operator, its use highlights the
intensity discontinuities in an image and deemphasizes regions with
slowly varying intensity levels. We can produce a sharpened image by
combining the Laplacian with the original image. If we use the first
kernel in Figure 9, we can add the image and the Laplacian. If g (x, y) is
the sharpened image

g (x, y) = f (x, y) + ∇2 f (x, y). (16)

Usually, we need to re-scale the graylevels to the [0, L − 1] after this

operation.

61/102
Unsharp Masking

A process used in printing industry is unsharp masking. It sharpens

the image by subtracting an unsharp (smoothed) version of the image
from the original. Unsharp masking consists of the following steps:
1 Blur the original image.
2 Subtract the blurred image from the original. The resulting
difference is called the mask.
3 Add the mask to the original.

62/102
Using the First-Order Derivatives

First-order derivative based operations are implemented using the

magnitude of the gradient. For a function f (x, y), the gradient of f at
the coordinates (x, y) is defined as the two-dimensional column vector

∂f
 
· ¸
fx  ∂x 
∇ f = grad( f ) = ∂f .
=  (17)
fy
∂y

This vector has the important property that it points in the direction of
the greatest rate of change of f at location (x, y).

63/102
The magnitude (length) of vector ∇ f , denoted as M (x, y), where,
q
M (x, y) = mag(∇ f ) = f x2 + f y2 , (18)

is the value at (x, y) of the rate of change in the direction of the gradient
vector. Note that M (x, y) is an image of the same size as the original,
created when x and y are allowed to vary over all the pixel locations in
f . This is commonly referred to as the gradient image of simply the
gradient.

64/102
Sobel operators are discrete approximations to the gradient. Figure 10
shows the Sobel operators.

-1 0 1 1 2 1

-2 0 2 0 0 0

-1 0 1 -1 -2 -1

(a) For f y (b) For f x

Figure 10: Sobel operators.

65/102
(a) Input (b) f x

(c) f y (d) Magnitude

Figure 11: Sobel filtering.

66/102
Outline

1 Point Operations
Histogram Processing

2 Spatial Filtering
Smoothing Spatial Filters
Sharpening Spatial Filters

3 Edge Detection
Line Detection Using the Hough Transform

67/102
Segmentation subdivides an image into its constituent regions or
objects. the level to which the subdivision is carried depends on the
problem being solved. Segmentation of nontrivial images is one of the
most difficult tasks in image processing. Segmentation accuracy
determines the eventual success or failure of computerized analysis
procedures. For this reason, considerable care should be taken to
improve the probability of rugged segmentation. In some situations,
such as industrial inspection applications, at least some measure of
control over the environment is possible at times. In others, as in
remote sensing, user control over image acquisition is limited
principally to the choice of image sensors.
Segmentation algorithms for monochrome images generally are based
on one of two basic properties of image intensity values: discontinuity
and similarity. In the first category, the approach is to partition an
image based on abrupt changes in intensity, such as edges in an
image. The principal approaches in the second category are based on
partitioning an image into regions that are similar according to a set of
predefined criteria.
68/102
There are three basic types of intensity discontinuities in a digital
image:
1 points,
2 lines, and
3 edges.
The most common way to look for discontinuities is to run a mask
through the image.

69/102
For a 3 × 3 mask this procedure involves computing the sum of
products of the coefficients with the intensity levels contained in the
region encompassed by the mask. That is the response, R , of the
mask at any point in the image is given by
9
X
R= w i zi
i =1

where z i is the intensity of the pixel associated with the mask

coefficient w i .

-1 -1 -1

-1 8 -1

-1 -1 -1

Figure 12: A mask for point detection.

70/102
Point Detection I

The detection of isolated points embedded in areas of constant or

nearly constant intensity in an image is straightforward in principle.
Using the mask shown in Figure 12 we say that an isolated point has
been detected at the location on which the mask is centered if

R ≥ T,

where T is a nonnegative threshold. Point detection is implemented in

MATLAB using functions imfilter , with a mask. The important
requirements are that the strongest response of a mask must be when
the mask is centered on an isolated point, and that the response be 0
in areas of constant intensity.

71/102
If T is given, the following command implements the point-detection
approach just discussed:
g = abs ( i m f i l t e r ( double ( f ) , w ) ) >= T ;
where f is the input image, w is an appropriate point-detection mask
and g is the resulting image.

72/102
Line Detection

The next level of complexity is line detection. Consider the mask in

Figure13.

73/102
-1 -1 -1 -1 -1 2 -1 2 -1 2 -1 -1
2 2 2 -1 2 -1 -1 2 -1 -1 2 -1
-1 -1 -1 2 -1 -1 -1 2 -1 -1 -1 2

(a) Hor (b) +45◦ (c) Ver (d) −45◦

Figure 13: Line detector masks.

74/102
If the first mask were moved around an image, it would respond more
strongly to lines (one pixel thick) oriented horizontally. With a constant
background, the maximum response would result when the line
passed through the middle row of the mask. Similarly, the second
mask in Figure13 responds best to lines oriented at +45◦ ; the third
mask to vertical lines; and the fourth mask to lines in the −45◦
direction. Note that the preferred direction of each mask is weighted
with a larger coefficient (i.e., 2) than other possible directions. The
coefficients of each mask sum to zero, indicating a zero response from
the mask in areas of constant intensity.

75/102
Edge Detection Using Function edge

Although point and line detection certainly are important in any

discussion on image segmentation, edge detection is by far the most
common approach for detecting meaningful in intensity values. Such
discontinuities are detected by using first-and second-order derivatives.
The first-order derivative of choice in image processing is the gradient.
The gradient of a 2-D function, f (x, y), is defined as the vector
· ¸ " ∂f #
Gx
∇f = = ∂∂xf
Gy ∂y .

76/102
The magnitude of this vector is

|∇ f | = [G x2 +G 2y ]1/2 .

To simplify computation, this quantity is approximated sometimes by

omitting the square-root operation,

|∇ f | ≈ G x2 +G 2y ,

or by using absolute values,

|∇ f | ≈ |G x | + |G y |.

77/102
These approximations still behave as derivatives; that is, they are zero
in areas of constant intensity and their values are proportional to the
degree of intensity change in areas whose pixel values are variable. It
is common practice to refer to a magnitude of the gradient or its
approximations simply as “the gradient.” A fundamental property of the
gradient vector is that it points in the direction of the maximum rate of
change of f at coordinates (x, y). The angle which this maximum rate
of change occurs is
Gy
µ ¶
α(x, y) = tan−1 .
Gx
One of the key issues is how to estimate the derivative G x and G y
digitally. The various approaches used by function edge are discussed
later in this section.

78/102
Second order derivatives in image processing are generally computed
using the Laplacian. That is the Laplacian of a 2-D function f (x, y) is
formed from second-order derivatives, as follows:

∂2 f (x, y) ∂2 f (x, y)
∇2 f (x, y) = + .
∂x 2 ∂y 2

The Laplacian is seldom used by itself for edge detection because, as

a second-order derivative, it is unacceptably sensitive to noise, its
magnitude produces double edges, and it is unable to detect edge
direction. However, the Laplacian can be powerful complement when
used in combination with other edge-detection techniques. For
example, although its double edges make it unsuitably for edge
detection directly, this property can be used for edge location.

79/102
With the preceding discussion as background, the basic idea behind
edge detection is to find places in an image where the intensity
changes rapidly, using one of two general criteria:
1 find places where the first derivative of the intensity is grater in
magnitude than a specified threshold, or
2 find places where the second derivative of the intensity has a zero
crossing.

80/102
IPT’s function edge provides several derivative estimators based on the
criteria just discussed. For some of these estimators, it is possible to
specify whether the edge detector is sensitive to horizontal or vertical
edges or to both. The general syntax for this function is
[ g , t ] = edge ( f , ‘ method ’ , parameters )
where f is the input image, method is one of the approaches listed in its
help and parameters are additional parameters explained in the
following discussion. In the output, g is a logical array with 1s at the
locations where edge points where detected in f and 0s elsewhere.
Parameter t is a optional; it gives the threshold used by edge to
determine which gradient values are strong enough to be called edge
points.

81/102
Sobel Edge Detector

The Sobel edge detector uses the masks in Figure 14 to approximate

digitally the first derivatives G x and G y . In other words, the gradient at
the center point in a neighborhood is computed as follows by the Sobel
detector:
g = |∇ f | = [G x2 +G 2y ]1/2 .

82/102
 z1 z2 z3

z4 z5 z6

z7 z8 z9

(a) General mask

-1 -2 -1 -1 0 1

0 0 0 -2 0 2

1 2 -1 -1 0 1

(b) Sobel, Verti- (c) Sobel, Hori-

cal zontal

-1 -1 -1 -1 0 1

0 0 0 -1 0 1

1 1 1 -1 0 83/102
1
Sobel edge detection can be implemented by filtering an image, f ,
(using imfilter ) with the left mask in Figure 14 filtering f again with the
other mask, squaring the pixels values of each filtered image, adding
the two results, and computing their square root. Similar comments
apply to the other and third entries in Figure 14. Function edge simply
packages the preceding operations into the function call and adds
other features, such as accepting a threshold value or determining a
threshold automatically. In addition, edge contains edge detection
techniques that are not implementable directly with imfilter .

84/102
The general calling syntax for the Sobel detector is
[ g , t ] = edge ( f , ‘ sobel ’ , T , d i r )

where f is the input image, T is specified threshold, and dir specifies

the preferred direction of the edges detected: ‘ horizontal ’ , ‘ vertical ’ or
‘ both’ (the default). As noted earlier, g is a logical image containing 1s
at locations where edges were detected and 0s elsewhere. Parameter
t in the output is optional. It is the threshold value used by edge. If T is
specified, then t = T. Otherwise, if T is not specified (or is empty, [ ]),
edge sets t equal to a threshold it determines automatically and then
uses for edge detection. One of the principal reason for including t in
the output argument is to get an initial value for the threshold. Function
edge uses the Sobel detector as a default if the syntax g = edge(f), is
used.

85/102
Prewitt Edge Detector

The Prewitt edge detector uses the masks in Figure 14. to

approximate digitally the first derivatives. Its general calling syntax is
[ g , t ] = edge ( f , ‘ prewitt ’ , T, dir )
The parameters of this function are identical to the Sobel parameters.
The Perwitt detector is slightly simpler to implement computationally
than the Sobel detector, but it tends to produce somewhat noisier
results. (It can be shown that the coefficient with value 2 in the Sobel
detector provides smoothing.)

86/102
Laplacian of a Gaussian (LoG) Detector

Consider the Gaussian function

r2
−
h(r ) = −e 2σ2

where r 2 = x 2 + y 2 and σ is the standard deviation. This is a smoothing

function which, if convolved with an image, will blur it. The degree of
blurring is determined by the value of σ. The Laplacian of this function
(the second derivative with respect to r ) is

r 2 − σ2 − r 22
· ¸
2
∇ h(r ) = − e 2σ
σ4

87/102
This function is called the Laplacian of Gaussian (LoG). Because the
second derivative is a linear operation, convolving (filtering) an image
with ∇2 h(r ) is the same as convolving the image with the smoothing
function first and then computing the Laplacian of the result. This is the
key concept underlying the LoG detector. We convolve the image with
∇2 h(r ) knowing that it has two effects: It smoothes the image (thus
reducing noise), and it computes the Laplacian, which yields a
double-edge image. Locating edges then consists of finding the zero
crossings between the double edges.

88/102
The general calling syntax for the LoG detector is
[ g , t ] = edge ( f , ‘ log ’ , T , sigma )
where sigma is the standard deviation and the other parameters are as
explained previously. The default value of sigma is 2. As before, edge
ignores any edges that are not stronger than T. If T is not provided, or
it is empty, [ ], edge chooses the value automatically. Setting T to 0
produces edges that are closed contours, a familiar characteristic of
the LoG method.

89/102
Zero-Crossings Detector

This detector is based on the same concept as the LoG method, but
the convolution is carried out using a specified filter function, H. The
calling syntax is
[ g , t ] = edge ( f , ‘ zerocross ’ , T , H)
The other parameters are as explained for the LoG detector.

90/102
Canny Edge Detector I
The Canny detector [1] is the most powerful edge detector provided by
function edge. The method can be summarized as follows:
1 The image is smoothed using a Gaussian filter with a specified
standard deviation, σ to reduce noise.
2 The local gradient, g (x, y) = [G x2 +G 2y ]1/2 , and edge direction,
α(x, y) = tan−1 (G y /G x ), are computed at each point, using an
operator such as Sobel or Prewitt. An edge point is defined to be
a point whose strength is locally maximum in the direction of the
gradient.
3 The edge point is determined in (2) give rise to ridges in the
gradient magnitude image. The algorithm then tracks along the
top of these ridges and sets to zero all pixels that are not actually
on the ridge top so as to give a thin line in the output, a process
known as nonmaximal suppression. The ridge pixels are then
thresholded using two thresholds, T1 and T2 , with T1 < T2 . Ridge

91/102
Canny Edge Detector II

pixels with values greater than T2 are said to be "strong" edge

pixels. Ridge pixels with values between T1 and T2 are said to be
“weak” edge pixels.
4 Finally the algorithm performs edge linking by incorporating the
weak pixels that are 8-connected to the strong pixels.

92/102
The syntax for the Canny edge detector is
[ g , t ] = edge ( f , ‘ canny ’ , T , sigma )
where T is a vector, T = [T1 , T2 ], containing the two thresholds
explained in step 3 of the preceding procedure, and sigma is the
standard deviation of the smoothing filter. If t is included in the output
argument, it is a two-element vector containing the two threshold
values used by the algorithm. The rest of the syntax is as explained for
the other methods, including the automatic computation of thresholds if
T is not supplied. The default value for sigma is 1.

93/102
We can extract and display the vertical edges in the image, f ,
[ gv , t ] = edge ( f , ‘ sobel ’ , ‘ vertical ’ ) ;
imshow ( gv )
t

t = 0.0516

94/102
Outline

1 Point Operations
Histogram Processing

2 Spatial Filtering
Smoothing Spatial Filters
Sharpening Spatial Filters

3 Edge Detection
Line Detection Using the Hough Transform

95/102
Ideally, the methods discussed in the previous section should yield
pixels lying only on edges. In practice, the resulting pixels seldom
characterize an edge complectly because of noise, breaks in the edge
from nonuniform illumination, and other effects that introduce spurious
intensity discontinuities. Thus edge-detection algorithms typically are
followed by linking procedures to assemble edge pixels into meaningful
edges. One approach that can be used to find and link line segments
in an image is the Hough transform.

96/102
Given a set of points in an image (typically a binary image), suppose
that we want to find subsets of these points that lie on straight lines.
One possible solution is to first find all lines determined by every pair of
points and then find all subsets of points that are close to particular
lines. The problem with this procedure is that it involves finding
n(n − 1)/2 ∼ n 2 lines and then performing n(n(n − 1))/2 ∼ n 3
comparisons of every point to all lines. This approach is
computationally prohibitive in all but the most trivial applications.

97/102
With the Hough transform, on the other hand, we consider a point
(x i , y i ) and all the lines that pass through it. Infinitely many lines pass
through (x i , y i ) all of which satisfy the slope-intercept equation
y i = ax i + b for some values of a and b . Writing this equation as
b = −ax i + y i and considering the ab -plane (also called parameter
space) yields the equation of a single line for a fixed pair (x i , y i ).
Furthermore, a second point (x j , y j ) also has a line in parameter space
associated with it, and this line intersects the line associated with
(x i , y i ) at (a 0 , b 0 ). where a 0 is the slope and b 0 the intercept of the line
containing both (x i , y i ) and (x j , y j ) in the x y -plane. In fact, all points
contained on this line have lines in parameter space that intersect at
(a 0 , b 0 ).

98/102
In principle, the parameter-space lines corresponding to all image
points (x i , y i ) could be plotted, and then the image lines could be
identified where large numbers of parameter lines intersect. A practical
difficulty with this approach, however, is that a (the slope of the line)
approaches infinity as the line approaches the vertical direction. One
way around this difficulty is to use the normal representation of a line:

x cos θ + y sin θ = ρ.

99/102
The computational attractiveness of the Hough transform arises from
subdividing theρθ parameter space into so-called accumulator cells.
Usually the maximum range of the values is −90◦ ≤ θ ≤ 90◦ and
−D ≤ ρ ≤ D where D is the distance between corners in the image.
The cell at the coordinates (i , j ) with accumulator value A(i , j ),
corresponds to the square associated with parameter space
coordinates (ρ i , θ j ) Initially, these cells are set to zero. Then, for every
non-background point (x k , y k ) in the image plane, we let θ equal each
of the allowed subdivision values on the θ axis and solve for the
corresponding ρ using the equation ρ k = x k cos θ + y k sin θ . The
resulting ρ -values are then rounded off to the nearest allowed cell
value along the ρ -axis. The corresponding accumulator cell is then
incremented. At the end of this procedure, a value of Q in A(i , j ),
means that Q points in the x y -plane lie on the line x cos θ j + y sin θ j = ρ i
The number of subdivisions in the ρθ -plane determines the accuracy
of the colinearity of these points.

100/102
Using houghlines to Find and Link Line Segments

l i n e s = h o u g h l i n e s ( f , t h e t a , rho , r , c )
f i g u r e , imshow ( f ) , h o l d on
fo r k = 1: length ( lines )
xy = [ l i n e s ( k ) . p o i n t 1 ; l i n e s ( k ) . p o i n t 2 ] ;
p o l t ( xy ( : , 2 ) , xy ( : , 1 ) , ‘ LineWidth ’ , 4 , ’Color ’ , [ . 6
end

101/102
J Canny.
A computational approach to edge detection.
IEEE Transactions on Pattern Analysis and Machine Intelligence,
8(6):679–698, 1986.

102/102