3.

4 Fundamentals of Spatial Filtering 153

for x = 0, 1, 2, … , M − 1 and y = 0, 1, 2, … , N − 1, where, as indicated above, C, k0 , k1 , k2 , and k3 are

specified constants, mG is the global mean of the input image, and sG is its standard deviation. Param-
eters mSxy and sSxy are the local mean and standard deviation, respectively, which change for every loca-
tion ( x, y). As usual, M and N are the number of rows and columns in the input image.
Factors such as the values of the global mean and variance relative to values in the areas to be
enhanced play a key role in selecting the parameters in Eq. (3-29), as does the range of differences
between the intensities of the areas to be enhanced and their background. In the case of Fig. 3.27(a),
mG = 161, sG = 103, the maximum intensity values of the image and areas to be enhanced are 228 and
10, respectively, and the minimum values are 0 in both cases.
We would like for the maximum value of the enhanced features to be the same as the maximum value
of the image, so we select C = 22.8. The areas to be enhanced are quite dark relative to the rest of the
image, and they occupy less than a third of the image area; thus, we expect the mean intensity in the
dark areas to be much less than the global mean. Based on this, we let k0 = 0 and k1 = 0.1. Because the
areas to be enhanced are of very low contrast, we let k2 = 0. For the upper limit of acceptable values
of standard deviation we set k3 = 0.1, which gives us one-tenth of the global standard deviation. Figure
3.27(b) is the result of using Eq. (3-29) with these parameters. By comparing this figure with Fig. 3.26(c),
we see that the method based on local statistics detected the same hidden features as local histogram
equalization. But the present approach extracted significantly more detail. For example, we see that all
the objects are solid, but only the boundaries were detected by local histogram equalization. In addition,
note that the intensities of the objects are not the same, with the objects in the top-left and bottom-right
being brighter than the others. Also, the horizontal rectangles in the lower left square evidently are of
different intensities. Finally, note that the background in both the image and dark squares in Fig. 3.27(b)
is nearly the same as in the original image; by comparison, the same regions in Fig. 3.26(c) exhibit more
visible noise and have lost their gray-level content. Thus, the additional complexity required to use local
statistics yielded results in this case that are superior to local histogram equalization.

3.4 FUNDAMENTALS OF SPATIAL FILTERING

3.4

In this section, we discuss the use of spatial filters for image processing. Spatial filter-
ing is used in a broad spectrum of image processing applications, so a solid under-
standing of filtering principles is important. As mentioned at the beginning of this
chapter, the filtering examples in this section deal mostly with image enhancement.
Other applications of spatial filtering are discussed in later chapters.
The name filter is borrowed from frequency domain processing (the topic of
Chapter 4) where “filtering” refers to passing, modifying, or rejecting specified fre-
quency components of an image. For example, a filter that passes low frequencies
is called a lowpass filter. The net effect produced by a lowpass filter is to smooth an
image by blurring it. We can accomplish similar smoothing directly on the image
itself by using spatial filters.
Spatial filtering modifies an image by replacing the value of each pixel by a func-
tion of the values of the pixel and its neighbors. If the operation performed on the
See Section 2.6 regarding
image pixels is linear, then the filter is called a linear spatial filter. Otherwise, the
linearity. filter is a nonlinear spatial filter. We will focus attention first on linear filters and then
introduce some basic nonlinear filters. Section 5.3 contains a more comprehensive
list of nonlinear filters and their application.

www.EBooksWorld.ir

DIP4E_GLOBAL_Print_Ready.indb 153 6/16/2017 2:03:34 PM

 154 Chapter 3 Intensity Transformations and Spatial Filtering

THE MECHANICS OF LINEAR SPATIAL FILTERING

A linear spatial filter performs a sum-of-products operation between an image f and a
filter kernel, w. The kernel is an array whose size defines the neighborhood of opera-
tion, and whose coefficients determine the nature of the filter. Other terms used to
refer to a spatial filter kernel are mask, template, and window. We use the term filter
kernel or simply kernel.
Figure 3.28 illustrates the mechanics of linear spatial filtering using a 3 × 3 ker-
nel. At any point ( x, y) in the image, the response, g( x, y), of the filter is the sum of
products of the kernel coefficients and the image pixels encompassed by the kernel:

g( x, y) = w(−1, −1) f ( x − 1, y − 1) + w(−1, 0) f ( x − 1, y) + …

(3-30)
+ w(0, 0) f ( x, y) + … + w(1, 1) f ( x + 1, y + 1)

As coordinates x and y are varied, the center of the kernel moves from pixel to pixel,
generating the filtered image, g, in the process.†
Observe that the center coefficient of the kernel, w(0, 0) , aligns with the pixel at
location ( x, y). For a kernel of size m × n, we assume that m = 2a + 1 and n = 2b + 1,
It certainly is possible where a and b are nonnegative integers. This means that our focus is on kernels of
to work with kernels of
even size, or mixed even odd size in both coordinate directions. In general, linear spatial filtering of an image
and odd sizes. However, of size M × N with a kernel of size m × n is given by the expression
working with odd sizes
simplifies indexing and a b
is also more intuitive
because the kernels have g( x, y) = ∑ ∑ w(s, t ) f ( x + s, y + t )
s = −a t = −b
(3-31)
centers falling on integer
values, and they are
spatially symmetric. where x and y are varied so that the center (origin) of the kernel visits every pixel in
f once. For a fixed value of ( x, y), Eq. (3-31) implements the sum of products of the
form shown in Eq. (3-30), but for a kernel of arbitrary odd size. As you will learn in
the following section, this equation is a central tool in linear filtering.

SPATIAL CORRELATION AND CONVOLUTION

Spatial correlation is illustrated graphically in Fig. 3.28, and it is described mathemati-
cally by Eq. (3-31). Correlation consists of moving the center of a kernel over an
image, and computing the sum of products at each location. The mechanics of spatial
convolution are the same, except that the correlation kernel is rotated by 180°. Thus,
when the values of a kernel are symmetric about its center, correlation and convolu-
tion yield the same result. The reason for rotating the kernel will become clear in
the following discussion. The best way to explain the differences between the two
concepts is by example.
We begin with a 1-D illustration, in which case Eq. (3-31) becomes
a
g( x ) = ∑ w(s) f ( x + s)
s = −a
(3-32)

†
A filtered pixel value typically is assigned to a corresponding location in a new image created to hold the results
of filtering. It is seldom the case that filtered pixels replace the values of the corresponding location in the origi-
nal image, as this would change the content of the image while filtering is being performed.

www.EBooksWorld.ir

DIP4E_GLOBAL_Print_Ready.indb 154 6/16/2017 2:03:36 PM

 3.4 Fundamentals of Spatial Filtering 155

FIGURE 3.28 Image origin

The mechanics y
of linear spatial
filtering Kernel origin
using a 3 × 3
kernel. The pixels
are shown as Magnified view showing filter kernel
squares to sim- Filter kernel coefficients and corresponding pixels
plify the graph- in the image
ics. Note that
the origin of the
image is at the top Image pixels
left, but the origin
of the kernel is at
its center. Placing
w (1, 1) w(1, 0) w (1, 1)
the origin at the Image f
center of spatially
symmetric kernels
simplifies writing w(0, 1) w (0, 0) w(0, 1) Filter kernel, w(s, t)
x
expressions for
linear filtering.
w (1,1) w (1, 0) w(1, 1)

f (x  1, y  1) f (x  1, y) f (x  1, y 1)
Kernel coefficients

f (x, y  1) f (x, y) f(x, y 1)

f(x 1, y  1) f(x 1, y) f(x 1, y 1)

Pixel values under kernel

when it is centered on (x, y)

Figure 3.29(a) shows a 1-D function, f, and a kernel, w. The kernel is of size 1 × 5, so
a = 2 and b = 0 in this case. Figure 3.29(b) shows the starting position used to per-
form correlation, in which w is positioned so that its center coefficient is coincident
with the origin of f.
The first thing we notice is that part of w lies outside f, so the summation is
Zero padding is not the
undefined in that area. A solution to this problem is to pad function f with enough
only padding option, as 0’s on either side. In general, if the kernel is of size 1 × m, we need (m − 1) 2 zeros
we will discuss in detail
later in this chapter.
on either side of f in order to handle the beginning and ending configurations of w
with respect to f. Figure 3.29(c) shows a properly padded function. In this starting
configuration, all coefficients of the kernel overlap valid values.

www.EBooksWorld.ir

DIP4E_GLOBAL_Print_Ready.indb 155 6/16/2017 2:03:36 PM

 156 Chapter 3 Intensity Transformations and Spatial Filtering

FIGURE 3.29 Correlation Convolution

Illustration of 1-D
correlation and Origin f w Origin f w rotated 180
convolution of a (a) 0 0 0 1 0 0 0 0 1 2 4 2 8 0 0 0 1 0 0 0 0 8 2 4 2 1 (i)
kernel, w, with a
function f (b) 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 (j)
consisting of a 1 2 4 2 8 8 2 4 2 1
discrete unit
impulse. Note that Starting position alignment Starting position alignment
correlation and
convolution are Zero padding Zero padding
functions of the
(c) 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 (k)
variable x, which
acts to displace 1 2 4 2 8 8 2 4 2 1
one function with Starting position Starting position
respect to the (d) 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 (l)
other. For the 1 2 4 2 8 8 2 4 2 1
extended
correlation and Position after 1 shift Position after 1 shift
convolution (e) 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 (m)
results, the
1 2 4 2 8 8 2 4 2 1
starting
configuration Position after 3 shifts Position after 3 shifts
places the right-
most element of (f) 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 (n)
the kernel to be 1 2 4 2 8 8 2 4 2 1
coincident with Final position Final position
the origin of f.
Additional Correlation result Convolution result
padding must be (g) 0 8 2 4 2 1 0 0 0 1 2 4 2 8 0 0 (o)
used.

Extended (full) correlation result Extended (full) convolution result

(h) 0 0 0 8 2 4 2 1 0 0 0 0 0 0 0 1 2 4 2 8 0 0 0 0 (p)

The first correlation value is the sum of products in this initial position, computed
using Eq. (3-32) with x = 0 :
2
g( 0 ) = ∑ w(s) f (s + 0) = 0
s = −2

This value is in the leftmost location of the correlation result in Fig. 3.29(g).
To obtain the second value of correlation, we shift the relative positions of w and
f one pixel location to the right [i.e., we let x = 1 in Eq. (3-32)] and compute the sum
of products again. The result is g(1) = 8, as shown in the leftmost, nonzero location
in Fig. 3.29(g). When x = 2, we obtain g(2) = 2. When x = 3, we get g(3) = 4 [see Fig.
3.29(e)]. Proceeding in this manner by varying x one shift at a time, we “build” the
correlation result in Fig. 3.29(g). Note that it took 8 values of x (i.e., x = 0, 1, 2, … , 7 )
to fully shift w past f so the center coefficient in w visited every pixel in f. Sometimes,
it is useful to have every element of w visit every pixel in f. For this, we have to start

www.EBooksWorld.ir

DIP4E_GLOBAL_Print_Ready.indb 156 6/16/2017 2:03:38 PM

 3.4 Fundamentals of Spatial Filtering 157

with the rightmost element of w coincident with the origin of f, and end with the
leftmost element of w being coincident the last element of f (additional padding
would be required). Figure Fig. 3.29(h) shows the result of this extended, or full, cor-
relation. As Fig. 3.29(g) shows, we can obtain the “standard” correlation by cropping
the full correlation in Fig. 3.29(h).
There are two important points to note from the preceding discussion. First, cor-
relation is a function of displacement of the filter kernel relative to the image. In
other words, the first value of correlation corresponds to zero displacement of the
kernel, the second corresponds to one unit displacement, and so on.† The second
thing to notice is that correlating a kernel w with a function that contains all 0’s and
a single 1 yields a copy of w, but rotated by 180°. A function that contains a single 1
Rotating a 1-D kernel
by 180° is equivalent to with the rest being 0’s is called a discrete unit impulse. Correlating a kernel with a dis-
flipping the kernel about
its axis.
crete unit impulse yields a rotated version of the kernel at the location of the impulse.
The right side of Fig. 3.29 shows the sequence of steps for performing convolution
(we will give the equation for convolution shortly). The only difference here is that
the kernel is pre-rotated by 180° prior to performing the shifting/sum of products
operations. As the convolution in Fig. 3.29(o) shows, the result of pre-rotating the
kernel is that now we have an exact copy of the kernel at the location of the unit
impulse. In fact, a foundation of linear system theory is that convolving a function
with an impulse yields a copy of the function at the location of the impulse. We will
use this property extensively in Chapter 4.
The 1-D concepts just discussed extend easily to images, as Fig. 3.30 shows. For a
kernel of size m × n, we pad the image with a minimum of (m − 1) 2 rows of 0’s at
the top and bottom and (n − 1) 2 columns of 0’s on the left and right. In this case,
m and n are equal to 3, so we pad f with one row of 0’s above and below and one
column of 0’s to the left and right, as Fig. 3.30(b) shows. Figure 3.30(c) shows the
initial position of the kernel for performing correlation, and Fig. 3.30(d) shows the
final result after the center of w visits every pixel in f, computing a sum of products
at each location. As before, the result is a copy of the kernel, rotated by 180°. We will
discuss the extended correlation result shortly.
In 2-D, rotation by 180°
For convolution, we pre-rotate the kernel as before and repeat the sliding sum of
is equivalent to flipping products just explained. Figures 3.30(f) through (h) show the result. You see again
the kernel about one axis
and then the other.
that convolution of a function with an impulse copies the function to the location
of the impulse. As noted earlier, correlation and convolution yield the same result if
the kernel values are symmetric about the center.
The concept of an impulse is fundamental in linear system theory, and is used in
numerous places throughout the book. A discrete impulse of strength (amplitude) A
located at coordinates ( x0 , y0 ) is defined as

⎧⎪ A if x = x0 and y = y0
d( x − x0 , y − y0 ) = ⎨ (3-33)
⎩⎪0 otherwise
†
In reality, we are shifting f to the left of w every time we increment x in Eq. (3-32). However, it is more intuitive
to think of the smaller kernel moving right over the larger array f. The motion of the two is relative, so either
way of looking at the motion is acceptable. The reason we increment f and not w is that indexing the equations
for correlation and convolution is much easier (and clearer) this way, especially when working with 2-D arrays.

www.EBooksWorld.ir

DIP4E_GLOBAL_Print_Ready.indb 157 6/16/2017 2:03:39 PM

 158 Chapter 3 Intensity Transformations and Spatial Filtering

FIGURE 3.30 Padded f

Correlation 0 0 0 0 0 0 0
(middle row) and Origin f 0 0 0 0 0 0 0
convolution (last 0 0 0 0 0 0 0 0 0 0 0 0
row) of a 2-D 0 0 0 0 0 w 0 0 0 1 0 0 0
kernel with an 0 0 1 0 0 1 2 3 0 0 0 0 0 0 0
image consisting 0 0 0 0 0 4 5 6 0 0 0 0 0 0 0
of a discrete unit 0 0 0 0 0 7 8 9 0 0 0 0 0 0 0
impulse. The 0’s
(a) (b)
are shown in gray
to simplify visual
Initial position for w Correlation result Full correlation result
analysis. Note that
correlation and 1 2 3 0 0 0 0 0 0 0 0 0 0 0
convolution are 4 5 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
functions of x and 7 8 9 0 0 0 0 0 9 8 7 0 0 0 9 8 7 0 0
y. As these 0 0 0 1 0 0 0 0 6 5 4 0 0 0 6 5 4 0 0
variable change, 0 0 0 0 0 0 0 0 3 2 1 0 0 0 3 2 1 0 0
they 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
displace one 0 0 0 0 0 0 0 0 0 0 0 0 0 0
function with (c) (d) (e)
respect to the
other. See the Rotated w Convolution result Full convolution result
discussion of Eqs. 9 8 7 0 0 0 0 0 0 0 0 0 0 0
(3-36) and (3-37) 6 5 4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
regarding full 3 2 1 0 0 0 0 0 1 2 3 0 0 0 1 2 3 0 0
correlation and 0 0 0 1 0 0 0 0 4 5 6 0 0 0 4 5 6 0 0
convolution. 0 0 0 0 0 0 0 0 7 8 9 0 0 0 7 8 9 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0
(f) (g) (h)

For example, the unit impulse in Fig. 3.29(a) is given by d( x − 3) in the 1-D version of
Recall that A = 1 for a
unit impulse.
the preceding equation. Similarly, the impulse in Fig. 3.30(a) is given by d( x − 2, y − 2)
[remember, the origin is at (0, 0) ].
Summarizing the preceding discussion in equation form, the correlation of a
kernel w of size m × n with an image f ( x, y), denoted as (w 夽 f )( x, y), is given by
Eq. (3-31), which we repeat here for convenience:
a b
(w 夽 f )( x, y) = ∑ ∑ w(s, t ) f ( x + s, y + t )
s = −a t = −b
(3-34)

Because our kernels do not depend on ( x, y), we will sometimes make this fact explic-
it by writing the left side of the preceding equation as w 夽 f ( x, y). Equation (3-34) is
evaluated for all values of the displacement variables x and y so that the center point
of w visits every pixel in f,† where we assume that f has been padded appropriately.

†
As we mentioned earlier, the minimum number of required padding elements for a 2-D correlation is (m − 1) 2
rows above and below f, and (n − 1) 2 columns on the left and right. With this padding, and assuming that f
is of size M × N , the values of x and y required to obtain a complete correlation are x = 0, 1, 2, … , M − 1 and
y = 0, 1, 2, … , N − 1. This assumes that the starting configuration is such that the center of the kernel coincides
with the origin of the image, which we have defined to be at the top, left (see Fig. 2.19).

www.EBooksWorld.ir

DIP4E_GLOBAL_Print_Ready.indb 158 6/16/2017 2:03:40 PM

 3.4 Fundamentals of Spatial Filtering 159

As explained earlier, a = (m − 1) 2, b = (n − 1) 2, and we assume that m and n are

odd integers.
In a similar manner, the convolution of a kernel w of size m × n with an image
f ( x, y), denoted by (w 夹 f )( x, y), is defined as
a b
(w 夹 f )( x, y) = ∑ ∑ w(s, t ) f ( x − s, y − t )
s = −a t = −b
(3-35)

where the minus signs align the coordinates of f and w when one of the functions is
rotated by 180° (see Problem 3.17). This equation implements the sum of products
process to which we refer throughout the book as linear spatial filtering. That is, lin-
ear spatial filtering and spatial convolution are synonymous.
Because convolution is commutative (see Table 3.5), it is immaterial whether w
or f is rotated, but rotation of the kernel is used by convention. Our kernels do not
depend on ( x, y), a fact that we sometimes make explicit by writing the left side
of Eq. (3-35) as w 夹 f ( x, y). When the meaning is clear, we let the dependence of
the previous two equations on x and y be implied, and use the simplified notation
w 夽 f and w 夹 f . As with correlation, Eq. (3-35) is evaluated for all values of the
displacement variables x and y so that the center of w visits every pixel in f, which
we assume has been padded. The values of x and y needed to obtain a full convolu-
tion are x = 0, 1, 2, … , M − 1 and y = 0, 1, 2, … , N − 1. The size of the result is M × N .
We can define correlation and convolution so that every element of w (instead of
just its center) visits every pixel in f. This requires that the starting configuration be
such that the right, lower corner of the kernel coincides with the origin of the image.
Similarly, the ending configuration will be with the top left corner of the kernel coin-
ciding with the lower right corner of the image. If the kernel and image are of sizes
m × n and M × N, respectively, the padding would have to increase to (m − 1) pad-
ding elements above and below the image, and (n − 1) elements to the left and right.
Under these conditions, the size of the resulting full correlation or convolution array
will be of size Sv × Sh , where (see Figs. 3.30(e) and (h), and Problem 3.19),

Sv = m + M − 1 (3-36)

and

Sh = n + N − 1 (3-37)

Often, spatial filtering algorithms are based on correlation and thus implement
Eq. (3-34) instead. To use the algorithm for correlation, we input w into it; for con-
volution, we input w rotated by 180°. The opposite is true for an algorithm that
implements Eq. (3-35). Thus, either Eq. (3-34) or Eq. (3-35) can be made to perform
the function of the other by rotating the filter kernel. Keep in mind, however, that
the order of the functions input into a correlation algorithm does make a difference,
because correlation is neither commutative nor associative (see Table 3.5).

www.EBooksWorld.ir

DIP4E_GLOBAL_Print_Ready.indb 159 6/16/2017 2:03:42 PM

 160 Chapter 3 Intensity Transformations and Spatial Filtering

TABLE 3.5
Property Convolution Correlation
Some fundamen-
tal properties of Commutative f 夹g = g夹 f —
convolution and
correlation. A Associative f 夹 ( g 夹 h) = ( f 夹 g ) 夹 h —
dash means that
the property does
Distributive f 夹 ( g + h) = ( f 夹 g ) + ( f 夹 h) f 夽 ( g + h) = ( f 夽 g ) + ( f 夽 h)
not hold.

Because the values of

Figure 3.31 shows two kernels used for smoothing the intensities of an image. To
these kernels are sym- filter an image using one of these kernels, we perform a convolution of the kernel
metric about the center, with the image in the manner just described. When talking about filtering and ker-
no rotation is required
before convolution. nels, you are likely to encounter the terms convolution filter, convolution mask, or
convolution kernel to denote filter kernels of the type we have been discussing. Typi-
cally, these terms are used in the literature to denote a spatial filter kernel, and not
to imply necessarily that the kernel is used for convolution. Similarly, “convolving a
kernel with an image” often is used to denote the sliding, sum-of-products process
we just explained, and does not necessarily differentiate between correlation and
convolution. Rather, it is used generically to denote either of the two operations.
This imprecise terminology is a frequent source of confusion. In this book, when we
use the term linear spatial filtering, we mean convolving a kernel with an image.
Sometimes an image is filtered (i.e., convolved) sequentially, in stages, using a dif-
ferent kernel in each stage. For example, suppose than an image f is filtered with a
kernel w1 , the result filtered with kernel w2 , that result filtered with a third kernel,
and so on, for Q stages. Because of the commutative property of convolution, this
multistage filtering can be done in a single filtering operation, w 夹 f , where
We could not write a
similar equation for
correlation because it is
w = w1 夹 w2 夹 w3 夹  夹 wQ (3-38)
not commutative.
The size of w is obtained from the sizes of the individual kernels by successive
applications of Eqs. (3-36) and (3-37). If all the individual kernels are of size m × n,
it follows from these equations that w will be of size Wv × Wh , where
Wv = Q × (m − 1) + m (3-39)
and
Wh = Q × (n − 1) + n (3-40)
These equations assume that every value of a kernel visits every value of the array
resulting from the convolution in the previous step. That is, the initial and ending
configurations, are as described in connection with Eqs. (3-36) and (3-37).

a b
1 1 1 0.3679 0.6065 0.3679
FIGURE 3.31
Examples of 1 1
smoothing kernels:  1 1 1  0.6065 1.0000 0.6065
9 4.8976
(a) is a box kernel;
(b) is a Gaussian
1 1 1 0.3679 0.6065 0.3679
kernel.

www.EBooksWorld.ir

DIP4E_GLOBAL_Print_Ready.indb 160 6/16/2017 2:03:43 PM

 3.4 Fundamentals of Spatial Filtering 161

SEPARABLE FILTER KERNELS

As noted in Section 2.6, a 2-D function G( x, y) is said to be separable if it can be written
as the product of two 1-D functions, G1 ( x) and G2 ( x); that is, G( x, y) = G1 ( x)G2 ( y).
A spatial filter kernel is a matrix, and a separable kernel is a matrix that can be
expressed as the outer product of two vectors. For example, the 2 * 3 kernel
To be strictly consistent
in notation, we should
use uppercase, bold ⎡1 1 1 ⎤
symbols for kernels when w=⎢ ⎥
we refer to them as ⎣1 1 1 ⎦
matrices. However,
kernels are mostly
treated in the book as is separable because it can be expressed as the outer product of the vectors
2-D functions, which we
denote in italics. To avoid
⎡1 ⎤
confusion, we continue ⎡1⎤
to use italics for kernels c=⎢ ⎥ and r = ⎢⎢1 ⎥⎥
in this short section, with
the understanding that
⎣1⎦ ⎢⎣1 ⎥⎦
the two notations are
intended to be equivalent
in this case. That is,

⎡1⎤ ⎡1 1 1 ⎤
c rT = ⎢ ⎥ [ 1 1 1 ] = ⎢1 1 1 ⎥ = w
⎣1⎦ ⎣ ⎦
A separable kernel of size m × n can be expressed as the outer product of two vec-
tors, v and w:

w = vwT (3-41)

where v and w are vectors of size m × 1 and n × 1, respectively. For a square kernel
of size m × m, we write

w = vvT (3-42)

It turns out that the product of a column vector and a row vector is the same as the
2-D convolution of the vectors (see Problem 3.24).
The importance of separable kernels lies in the computational advantages that
result from the associative property of convolution. If we have a kernel w that can
be decomposed into two simpler kernels, such that w = w1 夹 w2 , then it follows
from the commutative and associative properties in Table 3.5 that

w 夹 f = (w1 夹 w2 ) 夹 f = (w2 夹 w1 ) 夹 f = w2 夹 (w1 夹 f ) = (w1 夹 f ) 夹 w2 (3-43)

This equation says that convolving a separable kernel with an image is the same as
convolving w1 with f first, and then convolving the result with w2 .
We assume that the For an image of size M × N and a kernel of size m × n, implementation of Eq.
values of M and N (3-35) requires on the order of MNmn multiplications and additions. This is because
include any padding of
f prior to performing it follows directly from that equation that each pixel in the output (filtered) image
convolution. depends on all the coefficients in the filter kernel. But, if the kernel is separable and
we use Eq. (3-43), then the first convolution, w1 夹 f , requires on the order of MNm

www.EBooksWorld.ir

DIP4E_GLOBAL_Print_Ready.indb 161 6/16/2017 2:03:45 PM