3.

2 ■ Some Basic Intensity Transformation Functions 107

higher contrast than the original by darkening the intensity levels below k
and brightening the levels above k. In this technique, sometimes called
contrast stretching (see Section 3.2.4), values of r lower than k are com-
pressed by the transformation function into a narrow range of s, toward
black. The opposite is true for values of r higher than k. Observe how an in-
tensity value r0 is mapped to obtain the corresponding value s0. In the limit-
ing case shown in Fig. 3.2(b), T(r) produces a two-level (binary) image. A
mapping of this form is called a thresholding function. Some fairly simple, yet
powerful, processing approaches can be formulated with intensity transfor-
mation functions. In this chapter, we use intensity transformations principally
for image enhancement. In Chapter 10, we use them for image segmentation.
Approaches whose results depend only on the intensity at a point sometimes
are called point processing techniques, as opposed to the neighborhood pro-
cessing techniques discussed earlier in this section.

3.1.2 About the Examples in This Chapter

Although intensity transformations and spatial filtering span a broad range of
applications, most of the examples in this chapter are applications to image
enhancement. Enhancement is the process of manipulating an image so that
the result is more suitable than the original for a specific application. The
word specific is important here because it establishes at the outset that en-
hancement techniques are problem oriented. Thus, for example, a method
that is quite useful for enhancing X-ray images may not be the best approach
for enhancing satellite images taken in the infrared band of the electromag-
netic spectrum. There is no general “theory” of image enhancement. When an
image is processed for visual interpretation, the viewer is the ultimate judge
of how well a particular method works. When dealing with machine percep-
tion, a given technique is easier to quantify. For example, in an automated
character-recognition system, the most appropriate enhancement method is
the one that results in the best recognition rate, leaving aside other consider-
ations such as computational requirements of one method over another.
Regardless of the application or method used, however, image enhancement
is one of the most visually appealing areas of image processing. By its very na-
ture, beginners in image processing generally find enhancement applications in-
teresting and relatively simple to understand. Therefore, using examples from
image enhancement to illustrate the spatial processing methods developed in
this chapter not only saves having an extra chapter in the book dealing with
image enhancement but, more importantly, is an effective approach for intro-
ducing newcomers to the details of processing techniques in the spatial domain.
As you will see as you progress through the book, the basic material developed in
this chapter is applicable to a much broader scope than just image enhancement.

3.2 Some Basic Intensity Transformation Functions

Intensity transformations are among the simplest of all image processing tech-
niques. The values of pixels, before and after processing, will be denoted by r
and s, respectively. As indicated in the previous section, these values are related

www.EBooksWorld.ir
108 Chapter 3 ■ Intensity Transformations and Spatial Filtering

by an expression of the form s = T(r), where T is a transformation that maps a

pixel value r into a pixel value s. Because we are dealing with digital quantities,
values of a transformation function typically are stored in a one-dimensional
array and the mappings from r to s are implemented via table lookups. For an
8-bit environment, a lookup table containing the values of T will have 256 entries.
As an introduction to intensity transformations, consider Fig. 3.3, which
shows three basic types of functions used frequently for image enhance-
ment: linear (negative and identity transformations), logarithmic (log and
inverse-log transformations), and power-law (nth power and nth root trans-
formations). The identity function is the trivial case in which output intensi-
ties are identical to input intensities. It is included in the graph only for
completeness.

3.2.1 Image Negatives

The negative of an image with intensity levels in the range [0, L - 1] is ob-
tained by using the negative transformation shown in Fig. 3.3, which is given by
the expression

s = L - 1 - r (3.2-1)

Reversing the intensity levels of an image in this manner produces the

equivalent of a photographic negative. This type of processing is particularly
suited for enhancing white or gray detail embedded in dark regions of an

FIGURE 3.3 Some L
1

basic intensity
transformation Negative
functions. All
curves were nth root
scaled to fit in the 3L/4
range shown.
Output intensity level, s

Log
nth power
L/2

L/4

Identity Inverse log

0
0 L/4 L/2 3L/4 L
1
Input intensity level, r

www.EBooksWorld.ir
 3.2 ■ Some Basic Intensity Transformation Functions 109

a b
FIGURE 3.4
(a) Original digital
mammogram.
(b) Negative
image obtained
using the negative
transformation
in Eq. (3.2-1).
(Courtesy of G.E.
Medical Systems.)

image, especially when the black areas are dominant in size. Figure 3.4
shows an example. The original image is a digital mammogram showing a
small lesion. In spite of the fact that the visual content is the same in both
images, note how much easier it is to analyze the breast tissue in the nega-
tive image in this particular case.

3.2.2 Log Transformations

The general form of the log transformation in Fig. 3.3 is
s = c log (1 + r) (3.2-2)
where c is a constant, and it is assumed that r Ú 0. The shape of the log curve
in Fig. 3.3 shows that this transformation maps a narrow range of low intensity
values in the input into a wider range of output levels. The opposite is true of
higher values of input levels. We use a transformation of this type to expand
the values of dark pixels in an image while compressing the higher-level val-
ues. The opposite is true of the inverse log transformation.
Any curve having the general shape of the log functions shown in Fig. 3.3
would accomplish this spreading/compressing of intensity levels in an image,
but the power-law transformations discussed in the next section are much
more versatile for this purpose. The log function has the important character-
istic that it compresses the dynamic range of images with large variations in
pixel values. A classic illustration of an application in which pixel values have
a large dynamic range is the Fourier spectrum, which will be discussed in
Chapter 4. At the moment, we are concerned only with the image characteris-
tics of spectra. It is not unusual to encounter spectrum values that range from 0
to 10 6 or higher. While processing numbers such as these presents no problems
for a computer, image display systems generally will not be able to reproduce

www.EBooksWorld.ir
110 Chapter 3 ■ Intensity Transformations and Spatial Filtering

a b
FIGURE 3.5
(a) Fourier
spectrum.
(b) Result of
applying the log
transformation in
Eq. (3.2-2) with
c = 1.

faithfully such a wide range of intensity values. The net effect is that a signifi-
cant degree of intensity detail can be lost in the display of a typical Fourier
spectrum.
As an illustration of log transformations, Fig. 3.5(a) shows a Fourier spec-
trum with values in the range 0 to 1.5 * 106. When these values are scaled lin-
early for display in an 8-bit system, the brightest pixels will dominate the
display, at the expense of lower (and just as important) values of the spec-
trum. The effect of this dominance is illustrated vividly by the relatively small
area of the image in Fig. 3.5(a) that is not perceived as black. If, instead of dis-
playing the values in this manner, we first apply Eq. (3.2-2) (with c = 1 in this
case) to the spectrum values, then the range of values of the result becomes 0
to 6.2, which is more manageable. Figure 3.5(b) shows the result of scaling this
new range linearly and displaying the spectrum in the same 8-bit display. The
wealth of detail visible in this image as compared to an unmodified display of
the spectrum is evident from these pictures. Most of the Fourier spectra seen
in image processing publications have been scaled in just this manner.

3.2.3 Power-Law (Gamma) Transformations

Power-law transformations have the basic form

s = c rg (3.2-3)

where c and g are positive constants. Sometimes Eq. (3.2-3) is written as

s = c(r + e)g to account for an offset (that is, a measurable output when the
input is zero). However, offsets typically are an issue of display calibration
and as a result they are normally ignored in Eq. (3.2-3). Plots of s versus r for
various values of g are shown in Fig. 3.6. As in the case of the log transforma-
tion, power-law curves with fractional values of g map a narrow range of dark
input values into a wider range of output values, with the opposite being true
for higher values of input levels. Unlike the log function, however, we notice

www.EBooksWorld.ir
 3.2 ■ Some Basic Intensity Transformation Functions 111
L
1 FIGURE 3.6 Plots
of the equation
g  0.04 s = crg for
various values of
g  0.10
g (c = 1 in all
3L/4 g  0.20 cases). All curves
were scaled to fit
Output intensity level, s

in the range
g  0.40
shown.
g  0.67
L/2 g1
g  1.5

g  2.5

L/4 g  5.0

g  10.0

g  25.0

0
0 L/4 L/ 2 3L/4 L
1
Input intensity level, r

here a family of possible transformation curves obtained simply by varying g.

As expected, we see in Fig. 3.6 that curves generated with values of g 7 1
have exactly the opposite effect as those generated with values of g 6 1.
Finally, we note that Eq. (3.2-3) reduces to the identity transformation when
c = g = 1.
A variety of devices used for image capture, printing, and display respond
according to a power law. By convention, the exponent in the power-law equa-
tion is referred to as gamma [hence our use of this symbol in Eq. (3.2-3)].
The process used to correct these power-law response phenomena is called
gamma correction. For example, cathode ray tube (CRT) devices have an
intensity-to-voltage response that is a power function, with exponents vary-
ing from approximately 1.8 to 2.5. With reference to the curve for g = 2.5 in
Fig. 3.6, we see that such display systems would tend to produce images that
are darker than intended. This effect is illustrated in Fig. 3.7. Figure 3.7(a)
shows a simple intensity-ramp image input into a monitor. As expected, the
output of the monitor appears darker than the input, as Fig. 3.7(b) shows.
Gamma correction in this case is straightforward. All we need to do is pre-
process the input image before inputting it into the monitor by performing
the transformation s = r1>2.5 = r0.4. The result is shown in Fig. 3.7(c). When
input into the same monitor, this gamma-corrected input produces an out-
put that is close in appearance to the original image, as Fig. 3.7(d) shows. A
similar analysis would apply to other imaging devices such as scanners and
printers. The only difference would be the device-dependent value of
gamma (Poynton [1996]).

www.EBooksWorld.ir
112 Chapter 3 ■ Intensity Transformations and Spatial Filtering

a b
c d
FIGURE 3.7
(a) Intensity ramp
image. (b) Image
as viewed on a
simulated monitor
with a gamma of
2.5. (c) Gamma-
corrected image.
(d) Corrected
image as viewed
on the same Original image Gamma Original image as viewed
monitor. Compare correction on monitor
(d) and (a).

Gamma-corrected image Gamma-corrected image as

viewed on the same monitor

Gamma correction is important if displaying an image accurately on a

computer screen is of concern. Images that are not corrected properly can
look either bleached out, or, what is more likely, too dark. Trying to reproduce
colors accurately also requires some knowledge of gamma correction because
varying the value of gamma changes not only the intensity, but also the ratios
of red to green to blue in a color image. Gamma correction has become in-
creasingly important in the past few years, as the use of digital images for
commercial purposes over the Internet has increased. It is not unusual that
images created for a popular Web site will be viewed by millions of people,
the majority of whom will have different monitors and/or monitor settings.
Some computer systems even have partial gamma correction built in. Also,
current image standards do not contain the value of gamma with which an
image was created, thus complicating the issue further. Given these con-
straints, a reasonable approach when storing images in a Web site is to pre-
process the images with a gamma that represents an “average” of the types of
monitors and computer systems that one expects in the open market at any
given point in time.

www.EBooksWorld.ir
 3.2 ■ Some Basic Intensity Transformation Functions 113

■ In addition to gamma correction, power-law transformations are useful for EXAMPLE 3.1:
general-purpose contrast manipulation. Figure 3.8(a) shows a magnetic reso- Contrast
enhancement
nance image (MRI) of an upper thoracic human spine with a fracture disloca-
using power-law
tion and spinal cord impingement. The fracture is visible near the vertical transformations.
center of the spine, approximately one-fourth of the way down from the top of
the picture. Because the given image is predominantly dark, an expansion of
intensity levels is desirable. This can be accomplished with a power-law trans-
formation with a fractional exponent. The other images shown in the figure
were obtained by processing Fig. 3.8(a) with the power-law transformation

a b
c d
FIGURE 3.8
(a) Magnetic
resonance
image (MRI) of a
fractured human
spine.
(b)–(d) Results of
applying the
transformation in
Eq. (3.2-3) with
c = 1 and
g = 0.6, 0.4, and
0.3, respectively.
(Original image
courtesy of Dr.
David R. Pickens,
Department of
Radiology and
Radiological
Sciences,
Vanderbilt
University
Medical Center.)

www.EBooksWorld.ir
114 Chapter 3 ■ Intensity Transformations and Spatial Filtering

function of Eq. (3.2-3). The values of gamma corresponding to images (b)

through (d) are 0.6, 0.4, and 0.3, respectively (the value of c was 1 in all cases).
We note that, as gamma decreased from 0.6 to 0.4, more detail became visible.
A further decrease of gamma to 0.3 enhanced a little more detail in the back-
ground, but began to reduce contrast to the point where the image started to
have a very slight “washed-out” appearance, especially in the background. By
comparing all results, we see that the best enhancement in terms of contrast
and discernable detail was obtained with g = 0.4. A value of g = 0.3 is an ap-
proximate limit below which contrast in this particular image would be
reduced to an unacceptable level. ■

EXAMPLE 3.2: ■ Figure 3.9(a) shows the opposite problem of Fig. 3.8(a). The image to be
Another processed now has a washed-out appearance, indicating that a compression
illustration of
of intensity levels is desirable. This can be accomplished with Eq. (3.2-3)
power-law
transformations. using values of g greater than 1. The results of processing Fig. 3.9(a) with
g = 3.0, 4.0, and 5.0 are shown in Figs. 3.9(b) through (d). Suitable results
were obtained with gamma values of 3.0 and 4.0, the latter having a slightly

a b
c d
FIGURE 3.9
(a) Aerial image.
(b)–(d) Results of
applying the
transformation in
Eq. (3.2-3) with
c = 1 and
g = 3.0, 4.0, and
5.0, respectively.
(Original image
for this example
courtesy of
NASA.)

www.EBooksWorld.ir
 3.2 ■ Some Basic Intensity Transformation Functions 115

more appealing appearance because it has higher contrast. The result obtained
with g = 5.0 has areas that are too dark, in which some detail is lost. The dark
region to the left of the main road in the upper left quadrant is an example of
such an area. ■

3.2.4 Piecewise-Linear Transformation Functions

A complementary approach to the methods discussed in the previous three sec-
tions is to use piecewise linear functions. The principal advantage of piecewise
linear functions over the types of functions we have discussed thus far is that
the form of piecewise functions can be arbitrarily complex. In fact, as you will
see shortly, a practical implementation of some important transformations can
be formulated only as piecewise functions. The principal disadvantage of piece-
wise functions is that their specification requires considerably more user input.

Contrast stretching
One of the simplest piecewise linear functions is a contrast-stretching trans-
formation. Low-contrast images can result from poor illumination, lack of dy-
namic range in the imaging sensor, or even the wrong setting of a lens aperture
during image acquisition. Contrast stretching is a process that expands the
range of intensity levels in an image so that it spans the full intensity range of
the recording medium or display device.
Figure 3.10(a) shows a typical transformation used for contrast stretching. The
locations of points (r1, s1) and (r2, s2) control the shape of the transformation func-
tion. If r1 = s1 and r2 = s2, the transformation is a linear function that produces no
changes in intensity levels. If r1 = r2, s1 = 0 and s2 = L - 1, the transformation
becomes a thresholding function that creates a binary image, as illustrated in
Fig. 3.2(b). Intermediate values of (r1, s1) and (r2, s2) produce various degrees of
spread in the intensity levels of the output image, thus affecting its contrast. In gen-
eral, r1 … r2 and s1 … s2 is assumed so that the function is single valued and mo-
notonically increasing. This condition preserves the order of intensity levels, thus
preventing the creation of intensity artifacts in the processed image.
Figure 3.10(b) shows an 8-bit image with low contrast. Figure 3.10(c) shows
the result of contrast stretching, obtained by setting (r1, s1) = (rmin, 0) and
(r2, s2) = (rmax, L - 1), where rmin and rmax denote the minimum and maxi-
mum intensity levels in the image, respectively. Thus, the transformation func-
tion stretched the levels linearly from their original range to the full range
[0, L - 1]. Finally, Fig. 3.10(d) shows the result of using the thresholding func-
tion defined previously, with (r1, s1) = (m, 0) and (r2, s2) = (m, L - 1),
where m is the mean intensity level in the image. The original image on which
these results are based is a scanning electron microscope image of pollen, mag-
nified approximately 700 times.

Intensity-level slicing
Highlighting a specific range of intensities in an image often is of interest.Appli-
cations include enhancing features such as masses of water in satellite imagery
and enhancing flaws in X-ray images. The process, often called intensity-level

www.EBooksWorld.ir
116 Chapter 3 ■ Intensity Transformations and Spatial Filtering

a b
c d
L
1
FIGURE 3.10

Output intensity level, s

Contrast stretching. (r2, s2)
(a) Form of 3L/4
transformation
function. (b) A
L/ 2 T(r)
low-contrast image.
(c) Result of
contrast stretching. L/4
(d) Result of (r1, s1)
thresholding.
(Original image 0
0 L/4 L/2 3L/4 L
1
courtesy of Dr.
Roger Heady, Input intensity level, r
Research School of
Biological Sciences,
Australian National
University,
Canberra,
Australia.)

slicing, can be implemented in several ways, but most are variations of two basic
themes. One approach is to display in one value (say, white) all the values in the
range of interest and in another (say, black) all other intensities. This transfor-
mation, shown in Fig. 3.11(a), produces a binary image. The second approach,
based on the transformation in Fig. 3.11(b), brightens (or darkens) the desired
range of intensities but leaves all other intensity levels in the image unchanged.

a b L
1 L
1

FIGURE 3.11 (a) This

transformation
highlights intensity
range [A, B] and
reduces all other
s s T (r)
intensities to a lower T(r)
level. (b) This
transformation
highlights range
[A, B] and preserves
all other intensity
levels. r r
0 A B L
1 0 A B L
1

www.EBooksWorld.ir
 3.2 ■ Some Basic Intensity Transformation Functions 117

■ Figure 3.12(a) is an aortic angiogram near the kidney area (see Section EXAMPLE 3.3:
1.3.2 for a more detailed explanation of this image). The objective of this ex- Intensity-level
slicing.
ample is to use intensity-level slicing to highlight the major blood vessels that
appear brighter as a result of an injected contrast medium. Figure 3.12(b)
shows the result of using a transformation of the form in Fig. 3.11(a), with the
selected band near the top of the scale, because the range of interest is brighter
than the background. The net result of this transformation is that the blood
vessel and parts of the kidneys appear white, while all other intensities are
black. This type of enhancement produces a binary image and is useful for
studying the shape of the flow of the contrast medium (to detect blockages, for
example).
If, on the other hand, interest lies in the actual intensity values of the region
of interest, we can use the transformation in Fig. 3.11(b). Figure 3.12(c) shows
the result of using such a transformation in which a band of intensities in the
mid-gray region around the mean intensity was set to black, while all other in-
tensities were left unchanged. Here, we see that the gray-level tonality of the
major blood vessels and part of the kidney area were left intact. Such a result
might be useful when interest lies in measuring the actual flow of the contrast
medium as a function of time in a series of images. ■

Bit-plane slicing
Pixels are digital numbers composed of bits. For example, the intensity of each
pixel in a 256-level gray-scale image is composed of 8 bits (i.e., one byte). In-
stead of highlighting intensity-level ranges, we could highlight the contribution

a b c
FIGURE 3.12 (a) Aortic angiogram. (b) Result of using a slicing transformation of the type illustrated in Fig.
3.11(a), with the range of intensities of interest selected in the upper end of the gray scale. (c) Result of
using the transformation in Fig. 3.11(b), with the selected area set to black, so that grays in the area of the
blood vessels and kidneys were preserved. (Original image courtesy of Dr. Thomas R. Gest, University of
Michigan Medical School.)

www.EBooksWorld.ir
118 Chapter 3 ■ Intensity Transformations and Spatial Filtering

FIGURE 3.13 One 8-bit byte Bit plane 8

Bit-plane (most significant)
representation of
an 8-bit image.

Bit plane 1
(least significant)

made to total image appearance by specific bits. As Fig. 3.13 illustrates, an 8-bit
image may be considered as being composed of eight 1-bit planes, with plane 1
containing the lowest-order bit of all pixels in the image and plane 8 all the
highest-order bits.
Figure 3.14(a) shows an 8-bit gray-scale image and Figs. 3.14(b) through (i)
are its eight 1-bit planes, with Fig. 3.14(b) corresponding to the lowest-order bit.
Observe that the four higher-order bit planes, especially the last two, contain a
significant amount of the visually significant data. The lower-order planes con-
tribute to more subtle intensity details in the image. The original image has a
gray border whose intensity is 194. Notice that the corresponding borders of some
of the bit planes are black (0), while others are white (1). To see why, consider a

a b c
d e f
g h i
FIGURE 3.14 (a) An 8-bit gray-scale image of size 500 * 1192 pixels. (b) through (i) Bit planes 1 through 8,
with bit plane 1 corresponding to the least significant bit. Each bit plane is a binary image.

www.EBooksWorld.ir
 3.2 ■ Some Basic Intensity Transformation Functions 119

pixel in, say, the middle of the lower border of Fig. 3.14(a). The corresponding
pixels in the bit planes, starting with the highest-order plane, have values 1 1 0 0
0 0 1 0, which is the binary representation of decimal 194. The value of any pixel
in the original image can be similarly reconstructed from its corresponding
binary-valued pixels in the bit planes.
In terms of intensity transformation functions, it is not difficult to show that
the binary image for the 8th bit plane of an 8-bit image can be obtained by
processing the input image with a thresholding intensity transformation func-
tion that maps all intensities between 0 and 127 to 0 and maps all levels be-
tween 128 and 255 to 1. The binary image in Fig. 3.14(i) was obtained in just
this manner. It is left as an exercise (Problem 3.4) to obtain the intensity trans-
formation functions for generating the other bit planes.
Decomposing an image into its bit planes is useful for analyzing the rela-
tive importance of each bit in the image, a process that aids in determining
the adequacy of the number of bits used to quantize the image. Also, this type
of decomposition is useful for image compression (the topic of Chapter 8), in
which fewer than all planes are used in reconstructing an image. For example,
Fig. 3.15(a) shows an image reconstructed using bit planes 8 and 7. The recon-
struction is done by multiplying the pixels of the nth plane by the constant
2n - 1. This is nothing more than converting the nth significant binary bit to
decimal. Each plane used is multiplied by the corresponding constant, and all
planes used are added to obtain the gray scale image. Thus, to obtain
Fig. 3.15(a), we multiplied bit plane 8 by 128, bit plane 7 by 64, and added the
two planes. Although the main features of the original image were restored,
the reconstructed image appears flat, especially in the background. This is not
surprising because two planes can produce only four distinct intensity levels.
Adding plane 6 to the reconstruction helped the situation, as Fig. 3.15(b)
shows. Note that the background of this image has perceptible false contour-
ing. This effect is reduced significantly by adding the 5th plane to the recon-
struction, as Fig. 3.15(c) illustrates. Using more planes in the reconstruction
would not contribute significantly to the appearance of this image. Thus, we
conclude that storing the four highest-order bit planes would allow us to re-
construct the original image in acceptable detail. Storing these four planes in-
stead of the original image requires 50% less storage (ignoring memory
architecture issues).

a b c
FIGURE 3.15 Images reconstructed using (a) bit planes 8 and 7; (b) bit planes 8, 7, and 6; and (c) bit planes 8,
7, 6, and 5. Compare (c) with Fig. 3.14(a).

www.EBooksWorld.ir
120 Chapter 3 ■ Intensity Transformations and Spatial Filtering

3.3 Histogram Processing

The histogram of a digital image with intensity levels in the range [0, L - 1]
is a discrete function h(rk) = nk, where rk is the kth intensity value and nk is
the number of pixels in the image with intensity rk. It is common practice to
normalize a histogram by dividing each of its components by the total num-
ber of pixels in the image, denoted by the product MN, where, as usual, M
and N are the row and column dimensions of the image. Thus, a normalized
histogram is given by p(rk) = rk >MN, for k = 0, 1, 2, Á , L - 1. Loosely
speaking, p(rk) is an estimate of the probability of occurrence of intensity
level rk in an image. The sum of all components of a normalized histogram is
Consult the book Web equal to 1.
site for a review of basic Histograms are the basis for numerous spatial domain processing tech-
probability theory.
niques. Histogram manipulation can be used for image enhancement, as
shown in this section. In addition to providing useful image statistics, we shall
see in subsequent chapters that the information inherent in histograms also is
quite useful in other image processing applications, such as image compression
and segmentation. Histograms are simple to calculate in software and also
lend themselves to economic hardware implementations, thus making them a
popular tool for real-time image processing.
As an introduction to histogram processing for intensity transformations,
consider Fig. 3.16, which is the pollen image of Fig. 3.10 shown in four basic in-
tensity characteristics: dark, light, low contrast, and high contrast. The right
side of the figure shows the histograms corresponding to these images. The
horizontal axis of each histogram plot corresponds to intensity values, rk. The
vertical axis corresponds to values of h(rk) = nk or p(rk) = nk>MN if the val-
ues are normalized. Thus, histograms may be viewed graphically simply as
plots of h(rk) = nk versus rk or p(rk) = nk>MN versus rk.
We note in the dark image that the components of the histogram are con-
centrated on the low (dark) side of the intensity scale. Similarly, the compo-
nents of the histogram of the light image are biased toward the high side of
the scale. An image with low contrast has a narrow histogram located typi-
cally toward the middle of the intensity scale. For a monochrome image this
implies a dull, washed-out gray look. Finally, we see that the components of
the histogram in the high-contrast image cover a wide range of the intensity
scale and, further, that the distribution of pixels is not too far from uniform,
with very few vertical lines being much higher than the others. Intuitively, it
is reasonable to conclude that an image whose pixels tend to occupy the entire
range of possible intensity levels and, in addition, tend to be distributed uni-
formly, will have an appearance of high contrast and will exhibit a large vari-
ety of gray tones. The net effect will be an image that shows a great deal of
gray-level detail and has high dynamic range. It will be shown shortly that it
is possible to develop a transformation function that can automatically
achieve this effect, based only on information available in the histogram of
the input image.

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 121

Histogram of dark image

Histogram of light image

Histogram of low-contrast image

Histogram of high-contrast image

FIGURE 3.16 Four basic image types: dark, light, low contrast, high
contrast, and their corresponding histograms.

www.EBooksWorld.ir
122 Chapter 3 ■ Intensity Transformations and Spatial Filtering

3.3.1 Histogram Equalization

Consider for a moment continuous intensity values and let the variable r de-
note the intensities of an image to be processed. As usual, we assume that r is
in the range [0, L - 1], with r = 0 representing black and r = L - 1 repre-
senting white. For r satisfying these conditions, we focus attention on transfor-
mations (intensity mappings) of the form
s = T(r) 0 … r … L - 1 (3.3-1)
that produce an output intensity level s for every pixel in the input image hav-
ing intensity r. We assume that:
(a) T(r) is a monotonically† increasing function in the interval 0 … r … L - 1;
and
(b) 0 … T(r) … L - 1 for 0 … r … L - 1.
In some formulations to be discussed later, we use the inverse
r = T -1(s) 0 … s … L - 1 (3.3-2)
in which case we change condition (a) to
(a ¿ ) T(r) is a strictly monotonically increasing function in the interval
0 … r … L - 1.
The requirement in condition (a) that T(r) be monotonically increasing
guarantees that output intensity values will never be less than corresponding
input values, thus preventing artifacts created by reversals of intensity. Condi-
tion (b) guarantees that the range of output intensities is the same as the
input. Finally, condition (a ¿ ) guarantees that the mappings from s back to r
will be one-to-one, thus preventing ambiguities. Figure 3.17(a) shows a function

a b T(r) T (r)
FIGURE 3.17
(a) Monotonically L
1 L
1
increasing
Single
function, showing value, sk
how multiple T (r)
T(r)
values can map to
a single value. Single sk
(b) Strictly value, sq
monotonically
increasing ...
function. This is a
one-to-one
r r
mapping, both 0 rk L
1
ways. 0 Multiple Single L
1
values value

†
Recall that a function T(r) is monotonically increasing if T(r2) Ú T(r1) for r2 7 r1. T(r) is a strictly mo-
notonically increasing function if T(r2) 7 T(r1) for r2 7 r1. Similar definitions apply to monotonically
decreasing functions.

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 123

that satisfies conditions (a) and (b). Here, we see that it is possible for multi-
ple values to map to a single value and still satisfy these two conditions. That
is, a monotonic transformation function performs a one-to-one or many-to-
one mapping. This is perfectly fine when mapping from r to s. However,
Fig. 3.17(a) presents a problem if we wanted to recover the values of r unique-
ly from the mapped values (inverse mapping can be visualized by reversing
the direction of the arrows). This would be possible for the inverse mapping
of sk in Fig. 3.17(a), but the inverse mapping of sq is a range of values, which,
of course, prevents us in general from recovering the original value of r that
resulted in sq. As Fig. 3.17(b) shows, requiring that T(r) be strictly monotonic
guarantees that the inverse mappings will be single valued (i.e., the mapping
is one-to-one in both directions). This is a theoretical requirement that allows
us to derive some important histogram processing techniques later in this
chapter. Because in practice we deal with integer intensity values, we are
forced to round all results to their nearest integer values. Therefore, when
strict monotonicity is not satisfied, we address the problem of a nonunique in-
verse transformation by looking for the closest integer matches. Example 3.8
gives an illustration of this.
The intensity levels in an image may be viewed as random variables in the
interval [0, L - 1]. A fundamental descriptor of a random variable is its prob-
ability density function (PDF). Let pr (r) and ps (s) denote the PDFs of r and s,
respectively, where the subscripts on p are used to indicate that pr and ps are
different functions in general. A fundamental result from basic probability
theory is that if pr(r) and T(r) are known, and T(r) is continuous and differen-
tiable over the range of values of interest, then the PDF of the transformed
(mapped) variable s can be obtained using the simple formula

ps (s) = pr(r) ` `
dr
(3.3-3)
ds

Thus, we see that the PDF of the output intensity variable, s, is determined by
the PDF of the input intensities and the transformation function used [recall
that r and s are related by T(r)].
A transformation function of particular importance in image processing has
the form
r
s = T(r) = (L - 1) pr (w) dw (3.3-4)
L0
where w is a dummy variable of integration. The right side of this equation is
recognized as the cumulative distribution function (CDF) of random variable
r. Because PDFs always are positive, and recalling that the integral of a func-
tion is the area under the function, it follows that the transformation function
of Eq. (3.3-4) satisfies condition (a) because the area under the function can-
not decrease as r increases. When the upper limit in this equation is
r = (L - 1), the integral evaluates to 1 (the area under a PDF curve always
is 1), so the maximum value of s is (L - 1) and condition (b) is satisfied also.

www.EBooksWorld.ir
124 Chapter 3 ■ Intensity Transformations and Spatial Filtering

To find the ps(s) corresponding to the transformation just discussed, we use

Eq. (3.3-3). We know from Leibniz’s rule in basic calculus that the derivative of
a definite integral with respect to its upper limit is the integrand evaluated at
the limit. That is,

ds dT(r)
=
dr dr
r
d
= (L - 1) B p (w) dw R (3.3-5)
dr L0 r

= (L - 1)pr(r)

Substituting this result for dr> ds in Eq. (3.3-3), and keeping in mind that all
probability values are positive, yields

ps (s) = pr (r) ` `
dr
ds

= pr(r) ` `
1
(3.3-6)
(L - 1)pr (r)

1
= 0 … s … L - 1
L - 1
We recognize the form of ps (s) in the last line of this equation as a uniform
probability density function. Simply stated, we have demonstrated that per-
forming the intensity transformation in Eq. (3.3-4) yields a random variable, s,
characterized by a uniform PDF. It is important to note from this equation that
T(r) depends on pr (r) but, as Eq. (3.3-6) shows, the resulting ps (s) always is
uniform, independently of the form of pr(r). Figure 3.18 illustrates these
concepts.

pr (r) ps (s)

A
Eq. (3.3-4)
1
L
1

r s
0 L
1 0 L
1

a b
FIGURE 3.18 (a) An arbitrary PDF. (b) Result of applying the transformation in
Eq. (3.3-4) to all intensity levels, r. The resulting intensities, s, have a uniform PDF,
independently of the form of the PDF of the r’s.

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 125

■ To fix ideas, consider the following simple example. Suppose that the (con- EXAMPLE 3.4:
tinuous) intensity values in an image have the PDF Illustration of
Eqs. (3.3-4) and
2r (3.3-6).
for 0 … r … L - 1
pr(r) = c (L - 1)2
0 otherwise
From Eq. (3.3-4),
r r
2 r2
s = T(r) = (L - 1) pr (w) dw = w dw =
L0 L - 1 L0 L - 1
Suppose next that we form a new image with intensities, s, obtained using
this transformation; that is, the s values are formed by squaring the corre-
sponding intensity values of the input image and dividing them by (L - 1).
For example, consider an image in which L = 10, and suppose that a pixel
in an arbitrary location (x, y) in the input image has intensity r = 3. Then
the pixel in that location in the new image is s = T(r) = r 2>9 = 1. We can
verify that the PDF of the intensities in the new image is uniform simply by
substituting pr(r) into Eq. (3.3-6) and using the fact that s = r 2>(L - 1);
that is,
-1

ps(s) = pr(r) ` ` = `B R `
dr 2r ds
ds 2 dr
(L - 1)

` `
-1
2r d r2
= B R
(L - 1)2 dr L - 1

2 ` ` =
2r (L - 1) 1
=
(L - 1) 2 r L - 1
where the last step follows from the fact that r is nonnegative and we assume
that L 7 1. As expected, the result is a uniform PDF. ■

For discrete values, we deal with probabilities (histogram values) and sum-
mations instead of probability density functions and integrals.† As mentioned
earlier, the probability of occurrence of intensity level rk in a digital image is
approximated by
nk
pr(rk) = k = 0, 1, 2, Á , L - 1 (3.3-7)
MN
where MN is the total number of pixels in the image, nk is the number of pix-
els that have intensity rk, and L is the number of possible intensity levels in the
image (e.g., 256 for an 8-bit image). As noted in the beginning of this section, a
plot of pr(rk) versus rk is commonly referred to as a histogram.

†
The conditions of monotonicity stated earlier apply also in the discrete case. We simply restrict the val-
ues of the variables to be discrete.

www.EBooksWorld.ir
126 Chapter 3 ■ Intensity Transformations and Spatial Filtering

The discrete form of the transformation in Eq. (3.3-4) is

k
sk = T(rk) = (L - 1) a pr(rj)
j=0
(3.3-8)
(L - 1) k
MN ja
= nj k = 0, 1, 2, Á , L - 1
=0

Thus, a processed (output) image is obtained by mapping each pixel in the

input image with intensity rk into a corresponding pixel with level sk in the
output image, using Eq. (3.3-8). The transformation (mapping) T(rk) in this
equation is called a histogram equalization or histogram linearization trans-
formation. It is not difficult to show (Problem 3.10) that this transformation
satisfies conditions (a) and (b) stated previously in this section.

EXAMPLE 3.5: ■ Before continuing, it will be helpful to work through a simple example.
A simple Suppose that a 3-bit image (L = 8) of size 64 * 64 pixels (MN = 4096) has
illustration of
the intensity distribution shown in Table 3.1, where the intensity levels are in-
histogram
equalization. tegers in the range [0, L - 1] = [0, 7].
The histogram of our hypothetical image is sketched in Fig. 3.19(a). Values
of the histogram equalization transformation function are obtained using
Eq. (3.3-8). For instance,
0
s0 = T(r0) = 7 a pr(rj) = 7pr (r0) = 1.33
j=0

Similarly,
1
s1 = T(r1) = 7 a pr (rj) = 7pr (r0) + 7pr(r1) = 3.08
j=0

and s2 = 4.55, s3 = 5.67, s4 = 6.23, s5 = 6.65, s6 = 6.86, s7 = 7.00. This trans-

formation function has the staircase shape shown in Fig. 3.19(b).

pr(rk) = nk>MN
TABLE 3.1
rk nk
Intensity
distribution and r0 = 0 790 0.19
histogram values r1 = 1 1023 0.25
for a 3-bit, r2 = 2 850 0.21
64 * 64 digital r3 = 3 656 0.16
image. r4 = 4 329 0.08
r5 = 5 245 0.06
r6 = 6 122 0.03
r7 = 7 81 0.02

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 127

pr (rk) sk ps (sk)

.25 7.0 .25

.20 5.6 .20
.15 4.2 .15
T(r)
.10 2.8 .10
.05 1.4 .05
rk rk sk
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

a b c
FIGURE 3.19 Illustration of histogram equalization of a 3-bit (8 intensity levels) image. (a) Original
histogram. (b) Transformation function. (c) Equalized histogram.

At this point, the s values still have fractions because they were generated
by summing probability values, so we round them to the nearest integer:

s0 = 1.33 : 1 s4 = 6.23 : 6
s1 = 3.08 : 3 s5 = 6.65 : 7
s2 = 4.55 : 5 s6 = 6.86 : 7
s3 = 5.67 : 6 s7 = 7.00 : 7

These are the values of the equalized histogram. Observe that there are only
five distinct intensity levels. Because r0 = 0 was mapped to s0 = 1, there are
790 pixels in the histogram equalized image with this value (see Table 3.1).
Also, there are in this image 1023 pixels with a value of s1 = 3 and 850 pixels
with a value of s2 = 5. However both r3 and r4 were mapped to the same
value, 6, so there are (656 + 329) = 985 pixels in the equalized image with this
value. Similarly, there are (245 + 122 + 81) = 448 pixels with a value of 7 in
the histogram equalized image. Dividing these numbers by MN = 4096 yielded
the equalized histogram in Fig. 3.19(c).
Because a histogram is an approximation to a PDF, and no new allowed in-
tensity levels are created in the process, perfectly flat histograms are rare in
practical applications of histogram equalization. Thus, unlike its continuous
counterpart, it cannot be proved (in general) that discrete histogram equaliza-
tion results in a uniform histogram. However, as you will see shortly, using Eq.
(3.3-8) has the general tendency to spread the histogram of the input image so
that the intensity levels of the equalized image span a wider range of the in-
tensity scale. The net result is contrast enhancement. ■

We discussed earlier in this section the many advantages of having intensity

values that cover the entire gray scale. In addition to producing intensities that
have this tendency, the method just derived has the additional advantage that
it is fully “automatic.” In other words, given an image, the process of histogram
equalization consists simply of implementing Eq. (3.3-8), which is based on in-
formation that can be extracted directly from the given image, without the

www.EBooksWorld.ir
128 Chapter 3 ■ Intensity Transformations and Spatial Filtering

need for further parameter specifications. We note also the simplicity of the
computations required to implement the technique.
The inverse transformation from s back to r is denoted by

rk = T-1(sk) k = 0, 1, 2, Á , L - 1 (3.3-9)

It can be shown (Problem 3.10) that this inverse transformation satisfies con-
ditions (a¿) and (b) only if none of the levels, rk, k = 0, 1, 2, Á , L - 1, are
missing from the input image, which in turn means that none of the components
of the image histogram are zero. Although the inverse transformation is not
used in histogram equalization, it plays a central role in the histogram-matching
scheme developed in the next section.

EXAMPLE 3.6: ■ The left column in Fig. 3.20 shows the four images from Fig. 3.16, and the
Histogram center column shows the result of performing histogram equalization on each
equalization.
of these images. The first three results from top to bottom show significant im-
provement. As expected, histogram equalization did not have much effect on
the fourth image because the intensities of this image already span the full in-
tensity scale. Figure 3.21 shows the transformation functions used to generate the
equalized images in Fig. 3.20. These functions were generated using Eq. (3.3-8).
Observe that transformation (4) has a nearly linear shape, indicating that the
inputs were mapped to nearly equal outputs.
The third column in Fig. 3.20 shows the histograms of the equalized images. It
is of interest to note that, while all these histograms are different, the histogram-
equalized images themselves are visually very similar.This is not unexpected be-
cause the basic difference between the images on the left column is one of
contrast, not content. In other words, because the images have the same con-
tent, the increase in contrast resulting from histogram equalization was
enough to render any intensity differences in the equalized images visually in-
distinguishable. Given the significant contrast differences between the original
images, this example illustrates the power of histogram equalization as an
adaptive contrast enhancement tool. ■

3.3.2 Histogram Matching (Specification)

As indicated in the preceding discussion, histogram equalization automati-
cally determines a transformation function that seeks to produce an output
image that has a uniform histogram. When automatic enhancement is de-
sired, this is a good approach because the results from this technique are
predictable and the method is simple to implement. We show in this section
that there are applications in which attempting to base enhancement on a
uniform histogram is not the best approach. In particular, it is useful some-
times to be able to specify the shape of the histogram that we wish the
processed image to have. The method used to generate a processed image
that has a specified histogram is called histogram matching or histogram
specification.

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 129

FIGURE 3.20 Left column: images from Fig. 3.16. Center column: corresponding histogram-
equalized images. Right column: histograms of the images in the center column.

www.EBooksWorld.ir
130 Chapter 3 ■ Intensity Transformations and Spatial Filtering

FIGURE 3.21 255

Transformation
functions for
histogram
equalization. 192
Transformations
(4)
(1) through (4) (1)
were obtained from 128
the histograms of (2)
the images (from
top to bottom) in (3)
the left column of 64
Fig. 3.20 using
Eq. (3.3-8).
0
0 64 128 192 255

Let us return for a moment to continuous intensities r and z (considered con-

tinuous random variables), and let pr(r) and pz (z) denote their corresponding
continuous probability density functions. In this notation, r and z denote the in-
tensity levels of the input and output (processed) images, respectively. We can
estimate pr(r) from the given input image, while pz (z) is the specified probabili-
ty density function that we wish the output image to have.
Let s be a random variable with the property
r
s = T(r) = (L - 1) pr(w) dw (3.3-10)
L0
where, as before, w is a dummy variable of integration. We recognize this expres-
sion as the continuous version of histogram equalization given in Eq. (3.3-4).
Suppose next that we define a random variable z with the property
z
G(z) = (L - 1) pz (t) dt = s (3.3-11)
3
0

where t is a dummy variable of integration. It then follows from these two

equations that G(z) = T(r) and, therefore, that z must satisfy the condition

z = G-1[T(r)] = G-1(s) (3.3-12)

The transformation T(r) can be obtained from Eq. (3.3-10) once pr(r) has
been estimated from the input image. Similarly, the transformation function
G(z) can be obtained using Eq. (3.3-11) because pz(z) is given.
Equations (3.3-10) through (3.3-12) show that an image whose intensity
levels have a specified probability density function can be obtained from a
given image by using the following procedure:
1. Obtain pr(r) from the input image and use Eq. (3.3-10) to obtain the val-
ues of s.
2. Use the specified PDF in Eq. (3.3-11) to obtain the transformation function
G(z).

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 131

3. Obtain the inverse transformation z = G-1(s); because z is obtained from

s, this process is a mapping from s to z, the latter being the desired values.
4. Obtain the output image by first equalizing the input image using Eq.
(3.3-10); the pixel values in this image are the s values. For each pixel with
value s in the equalized image, perform the inverse mapping z = G-1(s) to
obtain the corresponding pixel in the output image. When all pixels have
been thus processed, the PDF of the output image will be equal to the
specified PDF.

■ Assuming continuous intensity values, suppose that an image has the inten- EXAMPLE 3.7:
sity PDF pr(r) = 2 r>(L - 1)2 for 0 … r … (L - 1) and pr(r) = 0 for other Histogram
specification.
values of r. Find the transformation function that will produce an image whose
intensity PDF is pz (z) = 3z2>(L - 1)3 for 0 … z … (L - 1) and pz (z) = 0 for
other values of z.
First, we find the histogram equalization transformation for the interval
[0, L - 1]:
r r
2 r2
s = T(r) = (L - 1) pr(w) dw = w dw =
L0 (L - 1) L0 (L - 1)

By definition, this transformation is 0 for values outside the range [0, L - 1].
Squaring the values of the input intensities and dividing them by (L - 1)2 will
produce an image whose intensities, s, have a uniform PDF because this is a
histogram-equalization transformation, as discussed earlier.
We are interested in an image with a specified histogram, so we find next
z z
3 z3
G(z) = (L - 1) pz (w) dw = w2 dw =
L0 (L - 1) L0
2
(L - 1)2

over the interval [0, L - 1]; this function is 0 elsewhere by definition. Finally,
we require that G(z) = s, but G(z) = z3>(L - 1)2; so z3>(L - 1)2 = s, and
we have

z = C (L - 1)2s D
1>3

So, if we multiply every histogram equalized pixel by (L - 1)2 and raise the
product to the power 1>3, the result will be an image whose intensities, z, have
the PDF pz(z) = 3z2>(L - 1)3 in the interval [0, L - 1], as desired.
Because s = r2>(L - 1) we can generate the z’s directly from the intensi-
ties, r, of the input image:
1/3

z = C (L - 1) s D = C (L - 1)r 2 D
1/3 r2 1/3
2
= B (L - 1)2 R
(L - 1)

Thus, squaring the value of each pixel in the original image, multiplying the re-
sult by (L - 1), and raising the product to the power 1>3 will yield an image

www.EBooksWorld.ir
132 Chapter 3 ■ Intensity Transformations and Spatial Filtering

whose intensity levels, z, have the specified PDF. We see that the intermedi-
ate step of equalizing the input image can be skipped; all we need is to obtain
the transformation function T(r) that maps r to s. Then, the two steps can be
combined into a single transformation from r to z. ■
As the preceding example shows, histogram specification is straightforward
in principle. In practice, a common difficulty is finding meaningful analytical
expressions for T(r) and G-1. Fortunately, the problem is simplified signifi-
cantly when dealing with discrete quantities. The price paid is the same as for
histogram equalization, where only an approximation to the desired histogram
is achievable. In spite of this, however, some very useful results can be ob-
tained, even with crude approximations.
The discrete formulation of Eq. (3.3-10) is the histogram equalization trans-
formation in Eq. (3.3-8), which we repeat here for convenience:
k
sk = T(rk) = (L - 1) a pr (rj)
j=0
(3.3-13)
k
(L - 1)
MN ja
= nj k = 0, 1, 2, Á , L - 1
=0

where, as before, MN is the total number of pixels in the image, nj is the num-
ber of pixels that have intensity value rj, and L is the total number of possible
intensity levels in the image. Similarly, given a specific value of sk, the discrete
formulation of Eq. (3.3-11) involves computing the transformation function
q
G(zq) = (L - 1) a pz (zi) (3.3-14)
i=0

for a value of q, so that

G(zq) = sk (3.3-15)
where pz (zi), is the ith value of the specified histogram. As before, we find the
desired value zq by obtaining the inverse transformation:
zq = G-1(sk) (3.3-16)

In other words, this operation gives a value of z for each value of s; thus, it per-
forms a mapping from s to z.
In practice, we do not need to compute the inverse of G. Because we deal
with intensity levels that are integers (e.g., 0 to 255 for an 8-bit image), it is a
simple matter to compute all the possible values of G using Eq. (3.3-14) for
q = 0, 1, 2, Á , L - 1. These values are scaled and rounded to their nearest
integer values spanning the range [0, L - 1]. The values are stored in a table.
Then, given a particular value of sk, we look for the closest match in the values
stored in the table. If, for example, the 64th entry in the table is the closest to
sk, then q = 63 (recall that we start counting at 0) and z63 is the best solution
to Eq. (3.3-15). Thus, the given value sk would be associated with z63 (i.e., that

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 133

specific value of sk would map to z63). Because the zs are intensities used
as the basis for specifying the histogram pz(z), it follows that z0 = 0,
z1 = 1, Á , zL - 1 = L - 1, so z63 would have the intensity value 63. By re-
peating this procedure, we would find the mapping of each value of sk to the
value of zq that is the closest solution to Eq. (3.3-15). These mappings are the
solution to the histogram-specification problem.
Recalling that the sks are the values of the histogram-equalized image, we
may summarize the histogram-specification procedure as follows:

1. Compute the histogram pr(r) of the given image, and use it to find the his-
togram equalization transformation in Eq. (3.3-13). Round the resulting
values, sk, to the integer range [0, L - 1].
2. Compute all values of the transformation function G using the Eq. (3.3-14)
for q = 0, 1, 2, Á , L - 1, where pz (zi) are the values of the specified his-
togram. Round the values of G to integers in the range [0, L - 1]. Store
the values of G in a table.
3. For every value of sk, k = 0, 1, 2, Á , L - 1, use the stored values of G
from step 2 to find the corresponding value of zq so that G(zq) is closest to
sk and store these mappings from s to z. When more than one value of zq
satisfies the given sk (i.e., the mapping is not unique), choose the smallest
value by convention.
4. Form the histogram-specified image by first histogram-equalizing the
input image and then mapping every equalized pixel value, sk, of this
image to the corresponding value zq in the histogram-specified image
using the mappings found in step 3. As in the continuous case, the inter-
mediate step of equalizing the input image is conceptual. It can be skipped
by combining the two transformation functions, T and G-1, as Example 3.8
shows.

As mentioned earlier, for G-1 to satisfy conditions (a¿) and (b), G has to be
strictly monotonic, which, according to Eq. (3.3-14), means that none of the val-
ues pz(zi) of the specified histogram can be zero (Problem 3.10). When working
with discrete quantities, the fact that this condition may not be satisfied is not a
serious implementation issue, as step 3 above indicates. The following example
illustrates this numerically.

■ Consider again the 64 * 64 hypothetical image from Example 3.5, whose EXAMPLE 3.8:
histogram is repeated in Fig. 3.22(a). It is desired to transform this histogram A simple example
of histogram
so that it will have the values specified in the second column of Table 3.2.
specification.
Figure 3.22(b) shows a sketch of this histogram.
The first step in the procedure is to obtain the scaled histogram-equalized
values, which we did in Example 3.5:

s0 = 1 s2 = 5 s4 = 7 s6 = 7

s1 = 3 s3 = 6 s5 = 7 s7 = 7

www.EBooksWorld.ir
134 Chapter 3 ■ Intensity Transformations and Spatial Filtering

a b pr (rk) pz (zq)
c d
FIGURE 3.22 .30 .30
(a) Histogram of a .25 .25
3-bit image. (b) .20 .20
Specified .15 .15
histogram.
(c) Transformation .10 .10
function obtained .05 .05
from the specified rk zq
histogram. 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
(d) Result of
G (zq) pz (zq)
performing
histogram
7 .25
specification. 6
Compare .20
5
(b) and (d). 4 .15
3 .10
2
1 .05
zq zq
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

In the next step, we compute all the values of the transformation function, G,
using Eq. (3.3-14):
0
G(z0) = 7 a pz (zj) = 0.00
j=0

Similarly,

G(z1) = 7 a pz (zj) = 7 C p(z0) + p(z1) D = 0.00

1

j=0

and
G(z2) = 0.00 G(z4) = 2.45 G(z6) = 5.95

G(z3) = 1.05 G(z5) = 4.55 G(z7) = 7.00

TABLE 3.2
Specified Actual
Specified and
zq pz (zq) pz(zk)
actual histograms
(the values in the z0 = 0 0.00 0.00
third column are z1 = 1 0.00 0.00
from the z2 = 2 0.00 0.00
computations z3 = 3 0.15 0.19
performed in the z4 = 4 0.20 0.25
body of Example z5 = 5 0.30 0.21
3.8). z6 = 6 0.20 0.24
z7 = 7 0.15 0.11

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 135

As in Example 3.5, these fractional values are converted to integers in our

valid range, [0, 7]. The results are:

G(z0) = 0.00 : 0 G(z4) = 2.45 : 2

G(z1) = 0.00 : 0 G(z5) = 4.55 : 5
G(z2) = 0.00 : 0 G(z6) = 5.95 : 6
G(z3) = 1.05 : 1 G(z7) = 7.00 : 7

These results are summarized in Table 3.3, and the transformation function is
sketched in Fig. 3.22(c). Observe that G is not strictly monotonic, so condition
(a¿) is violated. Therefore, we make use of the approach outlined in step 3 of
the algorithm to handle this situation.
In the third step of the procedure, we find the smallest value of zq so that
the value G(zq) is the closest to sk. We do this for every value of sk to create
the required mappings from s to z. For example, s0 = 1, and we see that
G(z3) = 1, which is a perfect match in this case, so we have the correspon-
dence s0 : z3. That is, every pixel whose value is 1 in the histogram equalized
image would map to a pixel valued 3 (in the corresponding location) in the
histogram-specified image. Continuing in this manner, we arrive at the map-
pings in Table 3.4.
In the final step of the procedure, we use the mappings in Table 3.4 to map
every pixel in the histogram equalized image into a corresponding pixel in the
newly created histogram-specified image. The values of the resulting his-
togram are listed in the third column of Table 3.2, and the histogram is
sketched in Fig. 3.22(d). The values of pz (zq) were obtained using the same
procedure as in Example 3.5. For instance, we see in Table 3.4 that s = 1 maps
to z = 3, and there are 790 pixels in the histogram-equalized image with a
value of 1. Therefore, pz (z3) = 790>4096 = 0.19.
Although the final result shown in Fig. 3.22(d) does not match the specified
histogram exactly, the general trend of moving the intensities toward the high
end of the intensity scale definitely was achieved. As mentioned earlier, ob-
taining the histogram-equalized image as an intermediate step is useful for ex-
plaining the procedure, but this is not necessary. Instead, we could list the
mappings from the rs to the ss and from the ss to the zs in a three-column

TABLE 3.3
zq G(zq)
All possible
z0 = 0 0 values of the
z1 = 1 0 transformation
z2 = 2 0 function G scaled,
z3 = 3 1 rounded, and
z4 = 4 2 ordered with
z5 = 5 5 respect to z.
z6 = 6 6
z7 = 7 7

www.EBooksWorld.ir
136 Chapter 3 ■ Intensity Transformations and Spatial Filtering

TABLE 3.4
sk : zq
Mappings of all
the values of sk 1 : 3
into corresponding 3 : 4
values of zq. 5 : 5
6 : 6
7 : 7

table. Then, we would use those mappings to map the original pixels directly
into the pixels of the histogram-specified image. ■

EXAMPLE 3.9: ■ Figure 3.23(a) shows an image of the Mars moon, Phobos, taken by NASA’s
Comparison Mars Global Surveyor. Figure 3.23(b) shows the histogram of Fig. 3.23(a). The
between
image is dominated by large, dark areas, resulting in a histogram characterized
histogram
equalization and by a large concentration of pixels in the dark end of the gray scale. At first
histogram glance, one might conclude that histogram equalization would be a good ap-
matching. proach to enhance this image, so that details in the dark areas become more
visible. It is demonstrated in the following discussion that this is not so.
Figure 3.24(a) shows the histogram equalization transformation [Eq. (3.3-8)
or (3.3-13)] obtained from the histogram in Fig. 3.23(b). The most relevant
characteristic of this transformation function is how fast it rises from intensity
level 0 to a level near 190. This is caused by the large concentration of pixels in
the input histogram having levels near 0. When this transformation is applied
to the levels of the input image to obtain a histogram-equalized result, the net
effect is to map a very narrow interval of dark pixels into the upper end of the
gray scale of the output image. Because numerous pixels in the input image
have levels precisely in this interval, we would expect the result to be an image
with a light, washed-out appearance. As Fig. 3.24(b) shows, this is indeed the

a b
FIGURE 3.23
(a) Image of the
Mars moon
7.00
Phobos taken by
Number of pixels (  10 4)

NASA’s Mars
Global Surveyor. 5.25
(b) Histogram.
(Original image
3.50
courtesy of
NASA.)
1.75

0
0 64 128 192 255
Intensity

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 137

255 a b
c
192 FIGURE 3.24
Output intensity

(a) Transformation
function for
128 histogram
equalization.
(b) Histogram-
64
equalized image
(note the washed-
0 out appearance).
0 64 128 192 255 (c) Histogram
Input intensity of (b).
7.00
Number of pixels (  10 4)

5.25

3.50

1.75

0
0 64 128 192 255
Intensity

case. The histogram of this image is shown in Fig. 3.24(c). Note how all the in-
tensity levels are biased toward the upper one-half of the gray scale.
Because the problem with the transformation function in Fig. 3.24(a) was
caused by a large concentration of pixels in the original image with levels near
0, a reasonable approach is to modify the histogram of that image so that it
does not have this property. Figure 3.25(a) shows a manually specified function
that preserves the general shape of the original histogram, but has a smoother
transition of levels in the dark region of the gray scale. Sampling this function
into 256 equally spaced discrete values produced the desired specified his-
togram. The transformation function G(z) obtained from this histogram using
Eq. (3.3-14) is labeled transformation (1) in Fig. 3.25(b). Similarly, the inverse
transformation G-1(s) from Eq. (3.3-16) (obtained using the step-by-step pro-
cedure discussed earlier) is labeled transformation (2) in Fig. 3.25(b). The en-
hanced image in Fig. 3.25(c) was obtained by applying transformation (2) to
the pixels of the histogram-equalized image in Fig. 3.24(b). The improvement
of the histogram-specified image over the result obtained by histogram equal-
ization is evident by comparing these two images. It is of interest to note that a
rather modest change in the original histogram was all that was required to
obtain a significant improvement in appearance. Figure 3.25(d) shows the his-
togram of Fig. 3.25(c). The most distinguishing feature of this histogram is
how its low end has shifted right toward the lighter region of the gray scale
(but not excessively so), as desired. ■

www.EBooksWorld.ir
138 Chapter 3 ■ Intensity Transformations and Spatial Filtering

a c 7.00
b

Number of pixels (  10 4)
d
5.25
FIGURE 3.25
(a) Specified
histogram. 3.50
(b) Transformations.
(c) Enhanced image
1.75
using mappings
from curve (2).
(d) Histogram of (c). 0
0 64 128 192 255
Intensity
255

192
Output intensity

(1)
128
(2)
64

0
0 64 128 192 255
Input intensity
7.00
Number of pixels (  104)

5.25

3.50

1.75

0
0 64 128 192 255
Intensity

Although it probably is obvious by now, we emphasize before leaving this

section that histogram specification is, for the most part, a trial-and-error
process. One can use guidelines learned from the problem at hand, just as we
did in the preceding example. At times, there may be cases in which it is possi-
ble to formulate what an “average” histogram should look like and use that as
the specified histogram. In cases such as these, histogram specification be-
comes a straightforward process. In general, however, there are no rules for
specifying histograms, and one must resort to analysis on a case-by-case basis
for any given enhancement task.

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 139

3.3.3 Local Histogram Processing

The histogram processing methods discussed in the previous two sections are
global, in the sense that pixels are modified by a transformation function
based on the intensity distribution of an entire image. Although this global ap-
proach is suitable for overall enhancement, there are cases in which it is neces-
sary to enhance details over small areas in an image. The number of pixels in
these areas may have negligible influence on the computation of a global
transformation whose shape does not necessarily guarantee the desired local
enhancement. The solution is to devise transformation functions based on the
intensity distribution in a neighborhood of every pixel in the image.
The histogram processing techniques previously described are easily adapted
to local enhancement. The procedure is to define a neighborhood and move
its center from pixel to pixel. At each location, the histogram of the points in
the neighborhood is computed and either a histogram equalization or his-
togram specification transformation function is obtained. This function is
then used to map the intensity of the pixel centered in the neighborhood. The
center of the neighborhood region is then moved to an adjacent pixel location
and the procedure is repeated. Because only one row or column of the neigh-
borhood changes during a pixel-to-pixel translation of the neighborhood, up-
dating the histogram obtained in the previous location with the new data
introduced at each motion step is possible (Problem 3.12). This approach has
obvious advantages over repeatedly computing the histogram of all pixels in
the neighborhood region each time the region is moved one pixel location.
Another approach used sometimes to reduce computation is to utilize
nonoverlapping regions, but this method usually produces an undesirable
“blocky” effect.

■ Figure 3.26(a) shows an 8-bit, 512 * 512 image that at first glance appears EXAMPLE 3.10:
to contain five black squares on a gray background. The image is slightly noisy, Local histogram
equalization.
but the noise is imperceptible. Figure 3.26(b) shows the result of global his-
togram equalization. As often is the case with histogram equalization of
smooth, noisy regions, this image shows significant enhancement of the noise.
Aside from the noise, however, Fig. 3.26(b) does not reveal any new significant
details from the original, other than a very faint hint that the top left and bot-
tom right squares contain an object. Figure 3.26(c) was obtained using local
histogram equalization with a neighborhood of size 3 * 3. Here, we see signif-
icant detail contained within the dark squares. The intensity values of these ob-
jects were too close to the intensity of the large squares, and their sizes were
too small, to influence global histogram equalization significantly enough to
show this detail. ■

3.3.4 Using Histogram Statistics for Image Enhancement

Statistics obtained directly from an image histogram can be used for image en-
hancement. Let r denote a discrete random variable representing intensity val-
ues in the range [0, L - 1], and let p(ri) denote the normalized histogram

www.EBooksWorld.ir
140 Chapter 3 ■ Intensity Transformations and Spatial Filtering

a b c
FIGURE 3.26 (a) Original image. (b) Result of global histogram equalization. (c) Result of local
histogram equalization applied to (a), using a neighborhood of size 3 * 3.

component corresponding to value ri. As indicated previously, we may view

p(ri) as an estimate of the probability that intensity ri occurs in the image from
which the histogram was obtained.
As we discussed in Section 2.6.8, the nth moment of r about its mean is de-
fined as
L-1
mn(r) = a (ri - m)n p(ri) (3.3-17)
i=0

We follow convention in where m is the mean (average intensity) value of r (i.e., the average intensity
using m for the mean
value. Do not confuse it
of the pixels in the image):
with the same symbol L-1
used to denote the num-
ber of rows in an m * n m = a ri p(ri) (3.3-18)
neighborhood, in which i=0
we also follow notational
convention. The second moment is particularly important:
L-1
m2(r) = a (ri - m)2 p(ri) (3.3-19)
i=0

We recognize this expression as the intensity variance, normally denoted by s2

(recall that the standard deviation is the square root of the variance). Whereas
the mean is a measure of average intensity, the variance (or standard devia-
tion) is a measure of contrast in an image. Observe that all moments are com-
puted easily using the preceding expressions once the histogram has been
obtained from a given image.
When working with only the mean and variance, it is common practice to es-
timate them directly from the sample values, without computing the histogram.
Appropriately, these estimates are called the sample mean and sample variance.
They are given by the following familiar expressions from basic statistics:
1 M-1 N-1
MN xa a f(x, y)
m = (3.3-20)
=0 y=0

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 141

and The denominator in

Eq. (3.3-21) is written

C f(x, y) - m D
sometimes as MN - 1
1 M-1 N-1 2 instead of MN. This is
s2 = a a (3.3-21) done to obtain a so-
MN x = 0 y = 0 called unbiased estimate
of the variance. Howev-
for x = 0, 1, 2, Á , M - 1 and y = 0, 1, 2, Á , N - 1. In other words, as we er, we are more interest-
ed in Eqs. (3.3-21) and
know, the mean intensity of an image can be obtained simply by summing the (3.3-19) agreeing when
values of all its pixels and dividing the sum by the total number of pixels in the the histogram in the lat-
ter equation is computed
image.A similar interpretation applies to Eq. (3.3-21).As we illustrate in the fol- from the same image
lowing example, the results obtained using these two equations are identical to used in Eq. (3.3-21). For
this we require the MN
the results obtained using Eqs. (3.3-18) and (3.3-19), provided that the histogram term. The difference is
used in these equations is computed from the same image used in Eqs. (3.3-20) negligible for any image
of practical size.
and (3.3-21).

■ Before proceeding, it will be useful to work through a simple numerical ex- EXAMPLE 3.11:
ample to fix ideas. Consider the following 2-bit image of size 5 * 5: Computing
histogram
statistics.
0 0 1 1 2
1 2 3 0 1
3 3 2 2 0
2 3 1 0 0
1 1 3 2 2

The pixels are represented by 2 bits; therefore, L = 4 and the intensity levels
are in the range [0, 3]. The total number of pixels is 25, so the histogram has the
components
6 7
p(r0) = = 0.24; p(r1) = = 0.28;
25 25

7 5
p(r2) = = 0.28; p(r3) = = 0.20
25 25
where the numerator in p(ri) is the number of pixels in the image with intensity
level ri. We can compute the average value of the intensities in the image using
Eq. (3.3-18):
3
m = a ri p(ri)
i=0

= (0)(0.24) + (1)(0.28) + (2)(0.28) + (3)(0.20)

= 1.44
Letting f(x, y) denote the preceding 5 * 5 array and using Eq. (3.3-20), we obtain
1 4 4
25 xa a f(x, y)
m =
=0 y=0

= 1.44

www.EBooksWorld.ir
142 Chapter 3 ■ Intensity Transformations and Spatial Filtering

As expected, the results agree. Similarly, the result for the variance is the same
(1.1264) using either Eq. (3.3-19) or (3.3-21). ■
We consider two uses of the mean and variance for enhancement purposes.
The global mean and variance are computed over an entire image and are use-
ful for gross adjustments in overall intensity and contrast. A more powerful
use of these parameters is in local enhancement, where the local mean and
variance are used as the basis for making changes that depend on image char-
acteristics in a neighborhood about each pixel in an image.
Let (x, y) denote the coordinates of any pixel in a given image, and let Sxy
denote a neighborhood (subimage) of specified size, centered on (x, y). The
mean value of the pixels in this neighborhood is given by the expression
L-1
mSxy = a ri pSxy (ri) (3.3-22)
i=0

where pSxy is the histogram of the pixels in region Sxy. This histogram has L
components, corresponding to the L possible intensity values in the input image.
However, many of the components are 0, depending on the size of Sxy. For ex-
ample, if the neighborhood is of size 3 * 3 and L = 256, only between 1 and 9
of the 256 components of the histogram of the neighborhood will be nonzero.
These non-zero values will correspond to the number of different intensities in
Sxy (the maximum number of possible different intensities in a 3 * 3 region is 9,
and the minimum is 1).
The variance of the pixels in the neighborhood similarly is given by
L-1
s2Sxy = a (ri - mSxy)2 pSxy(ri) (3.3-23)
i=0

As before, the local mean is a measure of average intensity in neighborhood

Sxy, and the local variance (or standard deviation) is a measure of intensity
contrast in that neighborhood. Expressions analogous to (3.3-20) and (3.3-21)
can be written for neighborhoods. We simply use the pixel values in the neigh-
borhoods in the summations and the number of pixels in the neighborhood in
the denominator.
As the following example illustrates, an important aspect of image process-
ing using the local mean and variance is the flexibility they afford in developing
simple, yet powerful enhancement techniques based on statistical measures
that have a close, predictable correspondence with image appearance.

EXAMPLE 3.12: ■ Figure 3.27(a) shows an SEM (scanning electron microscope) image of a
Local enhance- tungsten filament wrapped around a support. The filament in the center of
ment using
the image and its support are quite clear and easy to study. There is another
histogram
statistics. filament structure on the right, dark side of the image, but it is almost imper-
ceptible, and its size and other characteristics certainly are not easily discern-
able. Local enhancement by contrast manipulation is an ideal approach to
problems such as this, in which parts of an image may contain hidden features.

www.EBooksWorld.ir
 3.3 ■ Histogram Processing 143

a b c
FIGURE 3.27 (a) SEM image of a tungsten filament magnified approximately 130 *.
(b) Result of global histogram equalization. (c) Image enhanced using local histogram
statistics. (Original image courtesy of Mr. Michael Shaffer, Department of Geological
Sciences, University of Oregon, Eugene.)

In this particular case, the problem is to enhance dark areas while leaving
the light area as unchanged as possible because it does not require enhance-
ment. We can use the concepts presented in this section to formulate an en-
hancement method that can tell the difference between dark and light and, at
the same time, is capable of enhancing only the dark areas. A measure of
whether an area is relatively light or dark at a point (x, y) is to compare the av-
erage local intensity, mSxy, to the average image intensity, called the global
mean and denoted mG. This quantity is obtained with Eq. (3.3-18) or (3.3-20)
using the entire image. Thus, we have the first element of our enhancement
scheme: We will consider the pixel at a point (x, y) as a candidate for processing
if mSxy … k0 mG, where k0 is a positive constant with value less than 1.0.
Because we are interested in enhancing areas that have low contrast, we also
need a measure to determine whether the contrast of an area makes it a candi-
date for enhancement. We consider the pixel at a point (x, y) as a candidate for
enhancement if sSxy … k2sG, where sG is the global standard deviation
obtained using Eqs. (3.3-19) or (3.3-21) and k2 is a positive constant. The value
of this constant will be greater than 1.0 if we are interested in enhancing light
areas and less than 1.0 for dark areas.
Finally, we need to restrict the lowest values of contrast we are willing to ac-
cept; otherwise the procedure would attempt to enhance constant areas, whose
standard deviation is zero. Thus, we also set a lower limit on the local standard
deviation by requiring that k1sG … sSxy, with k1 6 k2. A pixel at (x, y) that
meets all the conditions for local enhancement is processed simply by multi-
plying it by a specified constant, E, to increase (or decrease) the value of its in-
tensity level relative to the rest of the image. Pixels that do not meet the
enhancement conditions are not changed.

www.EBooksWorld.ir
144 Chapter 3 ■ Intensity Transformations and Spatial Filtering

We summarize the preceding approach as follows. Let f(x, y) represent the

value of an image at any image coordinates (x, y), and let g(x, y) represent the
corresponding enhanced value at those coordinates. Then,

E # f(x, y) if mSxy … k0 mG AND k1sG … sSxy … k2sG

g(x, y) = c (3.3-24)
f(x, y) otherwise

for x = 0, 1, 2, Á , M - 1 and y = 0, 1, 2, Á , N - 1, where, as indicated

above, E, k0, k1, and k2 are specified parameters, mG is the global mean of the
input image, and sG is its standard deviation. Parameters mSxy and sSxy are the
local mean and standard deviation, respectively. As usual, M and N are the row
and column image dimensions.
Choosing the parameters in Eq. (3.3-24) generally requires a bit of experi-
mentation to gain familiarity with a given image or class of images. In this
case, the following values were selected: E = 4.0, k0 = 0.4, k1 = 0.02, and
k2 = 0.4. The relatively low value of 4.0 for E was chosen so that, when it was
multiplied by the levels in the areas being enhanced (which are dark), the re-
sult would still tend toward the dark end of the scale, and thus preserve the
general visual balance of the image. The value of k0 was chosen as less than
half the global mean because we can see by looking at the image that the areas
that require enhancement definitely are dark enough to be below half the
global mean. A similar analysis led to the choice of values for k1 and k2.
Choosing these constants is not difficult in general, but their choice definitely
must be guided by a logical analysis of the enhancement problem at hand. Fi-
nally, the size of the local area Sxy should be as small as possible in order to
preserve detail and keep the computational burden as low as possible. We
chose a region of size 3 * 3.
As a basis for comparison, we enhanced the image using global histogram
equalization. Figure 3.27(b) shows the result. The dark area was improved but
details still are difficult to discern, and the light areas were changed, something
we did not want to do. Figure 3.27(c) shows the result of using the local statis-
tics method explained above. In comparing this image with the original in Fig.
3.27(a) or the histogram equalized result in Fig. 3.27(b), we note the obvious
detail that has been brought out on the right side of Fig. 3.27(c). Observe, for
example, the clarity of the ridges in the dark filaments. It is noteworthy that
the light-intensity areas on the left were left nearly intact, which was one of
our initial objectives. ■

3.4 Fundamentals of Spatial Filtering

In this section, we introduce several basic concepts underlying the use of spa-
tial filters for image processing. Spatial filtering is one of the principal tools
used in this field for a broad spectrum of applications, so it is highly advisable
that you develop a solid understanding of these concepts. As mentioned at the
beginning of this chapter, the examples in this section deal mostly with the use
of spatial filters for image enhancement. Other applications of spatial filtering
are discussed in later chapters.

www.EBooksWorld.ir
 3.4 ■ Fundamentals of Spatial Filtering 145

The name filter is borrowed from frequency domain processing, which is

the topic of the next chapter, where “filtering” refers to accepting (passing) or
rejecting certain frequency components. For example, a filter that passes low
frequencies is called a lowpass filter. The net effect produced by a lowpass fil-
ter is to blur (smooth) an image. We can accomplish a similar smoothing di-
rectly on the image itself by using spatial filters (also called spatial masks,
kernels, templates, and windows). In fact, as we show in Chapter 4, there is a
one-to-one correspondence between linear spatial filters and filters in the fre- See Section 2.6.2
regarding linearity.
quency domain. However, spatial filters offer considerably more versatility be-
cause, as you will see later, they can be used also for nonlinear filtering,
something we cannot do in the frequency domain.

3.4.1 The Mechanics of Spatial Filtering

In Fig. 3.1, we explained briefly that a spatial filter consists of (1) a
neighborhood, (typically a small rectangle), and (2) a predefined operation that
is performed on the image pixels encompassed by the neighborhood. Filtering
creates a new pixel with coordinates equal to the coordinates of the center of
the neighborhood, and whose value is the result of the filtering operation.† A
processed (filtered) image is generated as the center of the filter visits each
pixel in the input image. If the operation performed on the image pixels is lin-
ear, then the filter is called a linear spatial filter. Otherwise, the filter is
nonlinear. We focus attention first on linear filters and then illustrate some
simple nonlinear filters. Section 5.3 contains a more comprehensive list of non-
linear filters and their application.
Figure 3.28 illustrates the mechanics of linear spatial filtering using a 3 * 3
neighborhood. At any point (x, y) in the image, the response, g(x, y), of the fil-
ter is the sum of products of the filter coefficients and the image pixels encom-
passed by the filter:
g(x, y) = w(-1, -1)f(x - 1, y - 1) + w(-1, 0)f(x - 1, y) + Á
+ w(0, 0)f(x, y) + Á + w(1, 1)f(x + 1, y + 1)

Observe that the center coefficient of the filter, w(0, 0), aligns with the pixel at
location (x, y). 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 our focus in
the following discussion is on filters of odd size, with the smallest being of size It certainly is possible to
work with filters of even
3 * 3. In general, linear spatial filtering of an image of size M * N with a fil- size or mixed even and
ter of size m * n is given by the expression: odd sizes. However,
working with odd sizes
a b simplifies indexing and
g(x, y) = a a w(s, t)f(x + s, y + t) also is more intuitive
because the filters have
s = -a t = -b
centers falling on integer
values.
where x and y are varied so that each pixel in w visits every pixel in f.

†
The 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 original image, as this would change the content of the image while filtering still is being
performed.

www.EBooksWorld.ir
146 Chapter 3 ■ Intensity Transformations and Spatial Filtering

Image origin
y

Filter mask

Image pixels

w (
1,
1) w(
1, 0) w (
1, 1)

Image

w (0,
1) w (0,0) w(0,1)

x

w (1,
1) w (1,0) w(1,1)

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

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)

Pixels of image
section under filter

FIGURE 3.28 The mechanics of linear spatial filtering using a 3 * 3 filter mask. The form chosen to denote
the coordinates of the filter mask coefficients simplifies writing expressions for linear filtering.

3.4.2 Spatial Correlation and Convolution

There are two closely related concepts that must be understood clearly when
performing linear spatial filtering. One is correlation and the other is
convolution. Correlation is the process of moving a filter mask over the image
and computing the sum of products at each location, exactly as explained in
the previous section. The mechanics of convolution are the same, except that
the filter is first rotated by 180°. The best way to explain the differences be-
tween the two concepts is by example. We begin with a 1-D illustration.
Figure 3.29(a) shows a 1-D function, f, and a filter, w, and Fig. 3.29(b) shows
the starting position to perform correlation. The first thing we note is that there

www.EBooksWorld.ir
 3.4 ■ Fundamentals of Spatial Filtering 147

Correlation Convolution

Origin f w Origin f w rotated 180

(a) 0 0 0 1 0 0 0 0 1 2 3 2 8 0 0 0 1 0 0 0 0 8 2 3 2 1 (i)

(b) 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 (j)
1 2 3 2 8 8 2 3 2 1
Starting position alignment

Zero padding

(c) 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 (k)
1 2 3 2 8 8 2 3 2 1

(d) 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 (l)
1 2 3 2 8 8 2 3 2 1
Position after one shift

(e) 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 (m)
1 2 3 2 8 8 2 3 2 1
Position after four shifts

(f) 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 (n)
1 2 3 2 8 8 2 3 2 1
Final position

Full correlation result Full convolution result

(g) 0 0 0 8 2 3 2 1 0 0 0 0 0 0 0 1 2 3 2 8 0 0 0 0 (o)

Cropped correlation result Cropped convolution result

(h) 0 8 2 3 2 1 0 0 0 1 2 3 2 8 0 0 (p)

FIGURE 3.29 Illustration of 1-D correlation and convolution of a filter with a discrete unit impulse. Note that
correlation and convolution are functions of displacement.

are parts of the functions that do not overlap. The solution to this problem is to
pad f with enough 0s on either side to allow each pixel in w to visit every pixel in Zero padding is not the
only option. For example,
f. If the filter is of size m, we need m - 1 0s on either side of f. Figure 3.29(c) we could duplicate the
shows a properly padded function. The first value of correlation is the sum of value of the first and last
element m - 1 times on
products of f and w for the initial position shown in Fig. 3.29(c) (the sum of each side of f, or mirror
products is 0). This corresponds to a displacement x = 0. To obtain the second the first and last m - 1
elements and use the
value of correlation, we shift w one pixel location to the right (a displacement of mirrored values for
x = 1) and compute the sum of products. The result again is 0. In fact, the first padding.

nonzero result is when x = 3, in which case the 8 in w overlaps the 1 in f and the
result of correlation is 8. Proceeding in this manner, we obtain the full correlation
result in Fig. 3.29(g). Note that it took 12 values of x (i.e., x = 0, 1, 2, Á , 11) to
fully slide w past f so that each pixel in w visited every pixel in f. Often, we like
to work with correlation arrays that are the same size as f, in which case we crop
the full correlation to the size of the original function, as Fig. 3.29(h) shows.

www.EBooksWorld.ir
148 Chapter 3 ■ Intensity Transformations and Spatial Filtering

There are two important points to note from the discussion in the preceding
paragraph. First, correlation is a function of displacement of the filter. In other
words, the first value of correlation corresponds to zero displacement of the
filter, the second corresponds to one unit displacement, and so on. The second
thing to notice is that correlating a filter w with a function that contains all 0s
and a single 1 yields a result that is a copy of w, but rotated by 180°. We call a
function that contains a single 1 with the rest being 0s a discrete unit impulse.
So we conclude that correlation of a function with a discrete unit impulse
yields a rotated version of the function at the location of the impulse.
The concept of convolution is a cornerstone of linear system theory. As you
will learn in Chapter 4, a fundamental property of convolution is that convolv-
ing a function with a unit impulse yields a copy of the function at the location
of the impulse. We saw in the previous paragraph that correlation yields a copy
Note that rotation by of the function also, but rotated by 180°. Therefore, if we pre-rotate the filter
180° is equivalent to flip-
ping the function hori-
and perform the same sliding sum of products operation, we should be able to
zontally. obtain the desired result. As the right column in Fig. 3.29 shows, this indeed is
the case. Thus, we see that to perform convolution all we do is rotate one func-
tion by 180° and perform the same operations as in correlation. As it turns out,
it makes no difference which of the two functions we rotate.
The preceding concepts extend easily to images, as Fig. 3.30 shows. For a fil-
ter of size m * n, we pad the image with a minimum of m - 1 rows of 0s at
the top and bottom and n - 1 columns of 0s on the left and right. In this case,
m and n are equal to 3, so we pad f with two rows of 0s above and below and
two columns of 0s to the left and right, as Fig. 3.30(b) shows. Figure 3.30(c)
shows the initial position of the filter mask for performing correlation, and
Fig. 3.30(d) shows the full correlation result. Figure 3.30(e) shows the corre-
In 2-D, rotation by 180° sponding cropped result. Note again that the result is rotated by 180°. For con-
is equivalent to flipping
the mask along one axis
volution, we pre-rotate the mask as before and repeat the sliding sum of
and then the other. products just explained. Figures 3.30(f) through (h) show the result. You see
again that convolution of a function with an impulse copies the function at the
location of the impulse. It should be clear that, if the filter mask is symmetric,
correlation and convolution yield the same result.
If, instead of containing a single 1, image f in Fig. 3.30 had contained a re-
gion identically equal to w, the value of the correlation function (after nor-
malization) would have been maximum when w was centered on that region
of f. Thus, as you will see in Chapter 12, correlation can be used also to find
matches between images.
Summarizing the preceding discussion in equation form, we have that the
correlation of a filter w(x, y) of size m * n with an image f(x, y), denoted as
w(x, y)  f(x, y), is given by the equation listed at the end of the last section,
which we repeat here for convenience:
a b
w(x, y)  f(x, y) = a a w(s, t)f(x + s, y + t) (3.4-1)
s = -a t = -b

This equation is evaluated for all values of the displacement variables x and y
so that all elements of w visit every pixel in f, where we assume that f has been
padded appropriately. As explained earlier, a = (m - 1)>2, b = (n - 1)>2,
and we assume for notational convenience that m and n are odd integers.
www.EBooksWorld.ir
 3.4 ■ Fundamentals of Spatial Filtering 149

Padded f FIGURE 3.30

0 0 0 0 0 0 0 0 0 Correlation
0 0 0 0 0 0 0 0 0 (middle row) and
0 0 0 0 0 0 0 0 0 convolution (last
Origin f(x, y) 0 0 0 0 0 0 0 0 0 row) of a 2-D
0 0 0 0 0 0 0 0 0 1 0 0 0 0 filter with a 2-D
0 0 0 0 0 w(x, y) 0 0 0 0 0 0 0 0 0 discrete, unit
0 0 1 0 0 1 2 3 0 0 0 0 0 0 0 0 0 impulse. The 0s
0 0 0 0 0 4 5 6 0 0 0 0 0 0 0 0 0 are shown in gray
0 0 0 0 0 7 8 9 0 0 0 0 0 0 0 0 0 to simplify visual
(a) (b) analysis.
Initial position for w Full correlation result Cropped correlation result
1 2 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
4 5 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 9 8 7 0
7 8 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 6 5 4 0
0 0 0 0 0 0 0 0 0 0 0 0 9 8 7 0 0 0 0 3 2 1 0
0 0 0 0 1 0 0 0 0 0 0 0 6 5 4 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 3 2 1 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 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
(c) (d) (e)
Rotated w Full convolution result Cropped convolution result
9 8 7 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
6 5 4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 2 3 0
3 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4 5 6 0
0 0 0 0 0 0 0 0 0 0 0 0 1 2 3 0 0 0 0 7 8 9 0
0 0 0 0 1 0 0 0 0 0 0 0 4 5 6 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 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 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)

In a similar manner, the convolution of w(x, y) and f(x, y), denoted by Often, when the mean-
ing is clear, we denote
w(x, y)
f(x, y),† is given by the expression the result of correlation
or convolution by a func-
a b tion g (x, y), instead of
w(x, y)
f(x, y) = a a w(s, t)f(x - s, y - t) (3.4-2) writing w(x, y)  f(x, y)
s = -a t = -b or w(x, y)
f(x, y). For
example, see the equa-
tion at the end of the
where the minus signs on the right flip f (i.e., rotate it by 180°). Flipping and previous section, and
shifting f instead of w is done for notational simplicity and also to follow Eq. (3.5-1).
convention. The result is the same. As with correlation, this equation is eval-
uated for all values of the displacement variables x and y so that every ele-
ment of w visits every pixel in f, which we assume has been padded
appropriately. You should expand Eq. (3.4-2) for a 3 * 3 mask and convince
yourself that the result using this equation is identical to the example in
Fig. 3.30. In practice, we frequently work with an algorithm that implements

†
Because correlation and convolution are commutative, we have that w(x, y)  f(x, y)
= f(x, y)  w(x, y) and w(x, y)
f(x, y) = f(x, y)
w(x, y).

www.EBooksWorld.ir
150 Chapter 3 ■ Intensity Transformations and Spatial Filtering

Eq. (3.4-1). If we want to perform correlation, we input w into the algorithm;

for convolution, we input w rotated by 180°. The reverse is true if an algo-
rithm that implements Eq. (3.4-2) is available instead.
As mentioned earlier, convolution is a cornerstone of linear system theory.
As you will learn in Chapter 4, the property that the convolution of a function
with a unit impulse copies the function at the location of the impulse plays a
central role in a number of important derivations. We will revisit convolution
in Chapter 4 in the context of the Fourier transform and the convolution the-
orem. Unlike Eq. (3.4-2), however, we will be dealing with convolution of
functions that are of the same size. The form of the equation is the same, but
the limits of summation are different.
Using correlation or convolution to perform spatial filtering is a matter of
preference. In fact, because either Eq. (3.4-1) or (3.4-2) can be made to per-
form the function of the other by a simple rotation of the filter, what is impor-
tant is that the filter mask used in a given filtering task be specified in a way
that corresponds to the intended operation. All the linear spatial filtering re-
sults in this chapter are based on Eq. (3.4-1).
Finally, we point out that you are likely to encounter the terms,
convolution filter, convolution mask or convolution kernel in the image pro-
cessing literature. As a rule, these terms are used to denote a spatial filter,
and not necessarily that the filter will be used for true convolution. Similarly,
“convolving a mask 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.

3.4.3 Vector Representation of Linear Filtering

When interest lies in the characteristic response, R, of a mask either for cor-
relation or convolution, it is convenient sometimes to write the sum of
products as

R = w1 z1 + w2 z2 + Á + wmn zmn
mn
= a wk zk (3.4-3)
k=1
Consult the Tutorials sec-
tion of the book Web site = wTz
for a brief review of vec-
tors and matrices.
where the ws are the coefficients of an m * n filter and the zs are the corre-
sponding image intensities encompassed by the filter. If we are interested in
using Eq. (3.4-3) for correlation, we use the mask as given. To use the same
equation for convolution, we simply rotate the mask by 180°, as explained in
the last section. It is implied that Eq. (3.4-3) holds for a particular pair of coor-
dinates (x, y). You will see in the next section why this notation is convenient
for explaining the characteristics of a given linear filter.

www.EBooksWorld.ir
 3.4 ■ Fundamentals of Spatial Filtering 151

FIGURE 3.31
w1 w2 w3 Another
representation of
a general 3 * 3
filter mask.
w4 w5 w6

w7 w8 w9

As an example, Fig. 3.31 shows a general 3 * 3 mask with coefficients la-

beled as above. In this case, Eq. (3.4-3) becomes
R = w1 z1 + w2 z2 + Á + w9 z9
9
= a wk zk (3.4-4)
k=1
T
= w z
where w and z are 9-dimensional vectors formed from the coefficients of the
mask and the image intensities encompassed by the mask, respectively.

3.4.4 Generating Spatial Filter Masks

Generating an m * n linear spatial filter requires that we specify mn mask co-
efficients. In turn, these coefficients are selected based on what the filter is
supposed to do, keeping in mind that all we can do with linear filtering is to im-
plement a sum of products. For example, suppose that we want to replace the
pixels in an image by the average intensity of a 3 * 3 neighborhood centered
on those pixels. The average value at any location (x, y) in the image is the sum
of the nine intensity values in the 3 * 3 neighborhood centered on (x, y) di-
vided by 9. Letting zi, i = 1, 2, Á , 9, denote these intensities, the average is
1 9
9 ia
R = zi
=1

But this is the same as Eq. (3.4-4) with coefficient values wi = 1>9. In other
words, a linear filtering operation with a 3 * 3 mask whose coefficients are 1> 9
implements the desired averaging. As we discuss in the next section, this oper-
ation results in image smoothing. We discuss in the following sections a num-
ber of other filter masks based on this basic approach.
In some applications, we have a continuous function of two variables, and
the objective is to obtain a spatial filter mask based on that function. For ex-
ample, a Gaussian function of two variables has the basic form
x2 + y2
h(x, y) = e - 2s2

where s is the standard deviation and, as usual, we assume that coordinates x

and y are integers. To generate, say, a 3 * 3 filter mask from this function, we

www.EBooksWorld.ir
152 Chapter 3 ■ Intensity Transformations and Spatial Filtering

sample it about its center. Thus, w1 = h(-1, -1), w2 = h(-1, 0), Á ,

w9 = h(1, 1). An m * n filter mask is generated in a similar manner. Recall
that a 2-D Gaussian function has a bell shape, and that the standard deviation
controls the “tightness” of the bell.
Generating a nonlinear filter requires that we specify the size of a neigh-
borhood and the operation(s) to be performed on the image pixels contained
in the neighborhood. For example, recalling that the max operation is nonlin-
ear (see Section 2.6.2), a 5 * 5 max filter centered at an arbitrary point (x, y)
of an image obtains the maximum intensity value of the 25 pixels and assigns
that value to location (x, y) in the processed image. Nonlinear filters are quite
powerful, and in some applications can perform functions that are beyond the
capabilities of linear filters, as we show later in this chapter and in Chapter 5.

3.5 Smoothing Spatial Filters

Smoothing filters are used for blurring and for noise reduction. Blurring is
used in preprocessing tasks, such as removal of small details from an image
prior to (large) object extraction, and bridging of small gaps in lines or curves.
Noise reduction can be accomplished by blurring with a linear filter and also
by nonlinear filtering.
3.5.1 Smoothing Linear Filters
The output (response) of a smoothing, linear spatial filter is simply the average
of the pixels contained in the neighborhood of the filter mask. These filters
sometimes are called averaging filters. As mentioned in the previous section,
they also are referred to a lowpass filters.
The idea behind smoothing filters is straightforward. By replacing the value
of every pixel in an image by the average of the intensity levels in the neigh-
borhood defined by the filter mask, this process results in an image with re-
duced “sharp” transitions in intensities. Because random noise typically
consists of sharp transitions in intensity levels, the most obvious application of
smoothing is noise reduction. However, edges (which almost always are desir-
able features of an image) also are characterized by sharp intensity transitions,
so averaging filters have the undesirable side effect that they blur edges. An-
other application of this type of process includes the smoothing of false con-
tours that result from using an insufficient number of intensity levels, as
discussed in Section 2.4.3. A major use of averaging filters is in the reduction
of “irrelevant” detail in an image. By “irrelevant” we mean pixel regions that
are small with respect to the size of the filter mask. This latter application is il-
lustrated later in this section.
Figure 3.32 shows two 3 * 3 smoothing filters. Use of the first filter yields
the standard average of the pixels under the mask. This can best be seen by
substituting the coefficients of the mask into Eq. (3.4-4):
1 9
9 ia
R = zi
=1

which is the average of the intensity levels of the pixels in the 3 * 3 neighbor-
hood defined by the mask, as discussed earlier. Note that, instead of being 1>9,

www.EBooksWorld.ir
 3.5 ■ Smoothing Spatial Filters 153

a b
1 1 1 1 2 1 FIGURE 3.32 Two
3 * 3 smoothing
(averaging) filter
1 1
 1 1 1  2 4 2 masks. The
9 16
constant multipli-
er in front of each
1 1 1 1 2 1 mask is equal to 1
divided by the
sum of the values
of its coefficients,
as is required to
compute an
average.
the coefficients of the filter are all 1s. The idea here is that it is computationally
more efficient to have coefficients valued 1. At the end of the filtering process
the entire image is divided by 9. An m * n mask would have a normalizing
constant equal to 1> mn. A spatial averaging filter in which all coefficients are
equal sometimes is called a box filter.
The second mask in Fig. 3.32 is a little more interesting. This mask yields a so-
called weighted average, terminology used to indicate that pixels are multiplied by
different coefficients, thus giving more importance (weight) to some pixels at the
expense of others. In the mask shown in Fig. 3.32(b) the pixel at the center of the
mask is multiplied by a higher value than any other, thus giving this pixel more
importance in the calculation of the average. The other pixels are inversely
weighted as a function of their distance from the center of the mask.The diagonal
terms are further away from the center than the orthogonal neighbors (by a fac-
tor of 12) and, thus, are weighed less than the immediate neighbors of the center
pixel. The basic strategy behind weighing the center point the highest and then
reducing the value of the coefficients as a function of increasing distance from the
origin is simply an attempt to reduce blurring in the smoothing process.We could
have chosen other weights to accomplish the same general objective. However,
the sum of all the coefficients in the mask of Fig. 3.32(b) is equal to 16, an attrac-
tive feature for computer implementation because it is an integer power of 2. In
practice, it is difficult in general to see differences between images smoothed by
using either of the masks in Fig. 3.32, or similar arrangements, because the area
spanned by these masks at any one location in an image is so small.
With reference to Eq. (3.4-1), the general implementation for filtering an
M * N image with a weighted averaging filter of size m * n (m and n odd) is
given by the expression
a b

a a w(s, t)f(x + s, y + t)
s = -a t = -b
g(x, y) = a b
(3.5-1)
a a w(s, t)
s = -a t = -b

The parameters in this equation are as defined in Eq. (3.4-1). As before, it is un-
derstood that the complete filtered image is obtained by applying Eq. (3.5-1)
for x = 0, 1, 2, Á , M - 1 and y = 0, 1, 2, Á , N - 1. The denominator in

www.EBooksWorld.ir
154 Chapter 3 ■ Intensity Transformations and Spatial Filtering

Eq. (3.5-1) is simply the sum of the mask coefficients and, therefore, it is a con-
stant that needs to be computed only once.

EXAMPLE 3.13: ■ The effects of smoothing as a function of filter size are illustrated in Fig. 3.33,
Image smoothing which shows an original image and the corresponding smoothed results ob-
with masks of
tained using square averaging filters of sizes m = 3, 5, 9, 15, and 35 pixels, re-
various sizes.
spectively. The principal features of these results are as follows: For m = 3, we
note a general slight blurring throughout the entire image but, as expected, de-
tails that are of approximately the same size as the filter mask are affected con-
siderably more. For example, the 3 * 3 and 5 * 5 black squares in the image,
the small letter “a,” and the fine grain noise show significant blurring when com-
pared to the rest of the image. Note that the noise is less pronounced, and the
jagged borders of the characters were pleasingly smoothed.
The result for m = 5 is somewhat similar, with a slight further increase in
blurring. For m = 9 we see considerably more blurring, and the 20% black cir-
cle is not nearly as distinct from the background as in the previous three im-
ages, illustrating the blending effect that blurring has on objects whose
intensities are close to that of its neighboring pixels. Note the significant fur-
ther smoothing of the noisy rectangles. The results for m = 15 and 35 are ex-
treme with respect to the sizes of the objects in the image. This type of
aggresive blurring generally is used to eliminate small objects from an image.
For instance, the three small squares, two of the circles, and most of the noisy
rectangle areas have been blended into the background of the image in
Fig. 3.33(f). Note also in this figure the pronounced black border. This is a re-
sult of padding the border of the original image with 0s (black) and then
trimming off the padded area after filtering. Some of the black was blended
into all filtered images, but became truly objectionable for the images
smoothed with the larger filters. ■

As mentioned earlier, an important application of spatial averaging is to

blur an image for the purpose of getting a gross representation of objects of
interest, such that the intensity of smaller objects blends with the back-
ground and larger objects become “bloblike” and easy to detect. The size of
the mask establishes the relative size of the objects that will be blended with
the background. As an illustration, consider Fig. 3.34(a), which is an image
from the Hubble telescope in orbit around the Earth. Figure 3.34(b) shows
the result of applying a 15 * 15 averaging mask to this image. We see that a
number of objects have either blended with the background or their inten-
sity has diminished considerably. It is typical to follow an operation like this
with thresholding to eliminate objects based on their intensity. The result of
using the thresholding function of Fig. 3.2(b) with a threshold value equal to
25% of the highest intensity in the blurred image is shown in Fig. 3.34(c).
Comparing this result with the original image, we see that it is a reasonable
representation of what we would consider to be the largest, brightest ob-
jects in that image.

www.EBooksWorld.ir
 3.5 ■ Smoothing Spatial Filters 155

FIGURE 3.33 (a) Original image, of size 500 * 500 pixels. (b)–(f) Results of smoothing a b
with square averaging filter masks of sizes m = 3, 5, 9, 15, and 35, respectively. The black c d
squares at the top are of sizes 3, 5, 9, 15, 25, 35, 45, and 55 pixels, respectively; their borders e f
are 25 pixels apart. The letters at the bottom range in size from 10 to 24 points, in
increments of 2 points; the large letter at the top is 60 points. The vertical bars are 5 pixels
wide and 100 pixels high; their separation is 20 pixels. The diameter of the circles is 25
pixels, and their borders are 15 pixels apart; their intensity levels range from 0% to 100%
black in increments of 20%. The background of the image is 10% black. The noisy
rectangles are of size 50 * 120 pixels.

www.EBooksWorld.ir
156 Chapter 3 ■ Intensity Transformations and Spatial Filtering

a b c
FIGURE 3.34 (a) Image of size 528 * 485 pixels from the Hubble Space Telescope. (b) Image filtered with a
15 * 15 averaging mask. (c) Result of thresholding (b). (Original image courtesy of NASA.)

3.5.2 Order-Statistic (Nonlinear) Filters

Order-statistic filters are nonlinear spatial filters whose response is based on or-
dering (ranking) the pixels contained in the image area encompassed by the fil-
ter, and then replacing the value of the center pixel with the value determined
by the ranking result. The best-known filter in this category is the median filter,
which, as its name implies, replaces the value of a pixel by the median of the in-
tensity values in the neighborhood of that pixel (the original value of the pixel is
included in the computation of the median). Median filters are quite popular be-
cause, for certain types of random noise, they provide excellent noise-reduction
capabilities, with considerably less blurring than linear smoothing filters of simi-
lar size. Median filters are particularly effective in the presence of impulse noise,
also called salt-and-pepper noise because of its appearance as white and black
dots superimposed on an image.
The median, j, of a set of values is such that half the values in the set are
less than or equal to j, and half are greater than or equal to j. In order to per-
form median filtering at a point in an image, we first sort the values of the pixel
in the neighborhood, determine their median, and assign that value to the cor-
responding pixel in the filtered image. For example, in a 3 * 3 neighborhood
the median is the 5th largest value, in a 5 * 5 neighborhood it is the 13th
largest value, and so on. When several values in a neighborhood are the same,
all equal values are grouped. For example, suppose that a 3 * 3 neighborhood
has values (10, 20, 20, 20, 15, 20, 20, 25, 100). These values are sorted as (10, 15,
20, 20, 20, 20, 20, 25, 100), which results in a median of 20. Thus, the principal
function of median filters is to force points with distinct intensity levels to be
more like their neighbors. In fact, isolated clusters of pixels that are light or
dark with respect to their neighbors, and whose area is less than m2>2 (one-
half the filter area), are eliminated by an m * m median filter. In this case
“eliminated” means forced to the median intensity of the neighbors. Larger
clusters are affected considerably less.

www.EBooksWorld.ir
 3.6 ■ Sharpening Spatial Filters 157

a b c
FIGURE 3.35 (a) X-ray image of circuit board corrupted by salt-and-pepper noise. (b) Noise reduction with
a 3 * 3 averaging mask. (c) Noise reduction with a 3 * 3 median filter. (Original image courtesy of Mr.
Joseph E. Pascente, Lixi, Inc.)

Although the median filter is by far the most useful order-statistic filter in
image processing, it is by no means the only one. The median represents the
50th percentile of a ranked set of numbers, but recall from basic statistics that See Section 10.3.5 regard-
ing percentiles.
ranking lends itself to many other possibilities. For example, using the 100th
percentile results in the so-called max filter, which is useful for finding the
brightest points in an image. The response of a 3 * 3 max filter is given by
R = max5zk ƒ k = 1, 2, Á , 96. The 0th percentile filter is the min filter, used
for the opposite purpose. Median, max, min, and several other nonlinear filters
are considered in more detail in Section 5.3.

■ Figure 3.35(a) shows an X-ray image of a circuit board heavily corrupted EXAMPLE 3.14:
by salt-and-pepper noise. To illustrate the point about the superiority of medi- Use of median
filtering for noise
an filtering over average filtering in situations such as this, we show in Fig.
reduction.
3.35(b) the result of processing the noisy image with a 3 * 3 neighborhood av-
eraging mask, and in Fig. 3.35(c) the result of using a 3 * 3 median filter. The
averaging filter blurred the image and its noise reduction performance was
poor. The superiority in all respects of median over average filtering in this
case is quite evident. In general, median filtering is much better suited than av-
eraging for the removal of salt-and-pepper noise. ■

3.6 Sharpening Spatial Filters

The principal objective of sharpening is to highlight transitions in intensity.
Uses of image sharpening vary and include applications ranging from electron-
ic printing and medical imaging to industrial inspection and autonomous guid-
ance in military systems. In the last section, we saw that image blurring could be
accomplished in the spatial domain by pixel averaging in a neighborhood. Be-
cause averaging is analogous to integration, it is logical to conclude that sharp-
ening can be accomplished by spatial differentiation. This, in fact, is the case,

www.EBooksWorld.ir
158 Chapter 3 ■ Intensity Transformations and Spatial Filtering

and the discussion in this section deals with various ways of defining and imple-
menting operators for sharpening by digital differentiation. Fundamentally, the
strength of the response of a derivative operator is proportional to the degree
of intensity discontinuity of the image at the point at which the operator is ap-
plied. Thus, image differentiation enhances edges and other discontinuities
(such as noise) and deemphasizes areas with slowly varying intensities.

3.6.1 Foundation
In the two sections that follow, we consider in some detail sharpening filters
that are based on first- and second-order derivatives, respectively. Before pro-
ceeding with that discussion, however, we stop to look at some of the funda-
mental properties of these derivatives in a digital context. To simplify the
explanation, we focus attention initially on one-dimensional derivatives. In
particular, we are interested in the behavior of these derivatives in areas of
constant intensity, at the onset and end of discontinuities (step and ramp dis-
continuities), and along intensity ramps. As you will see in Chapter 10, these
types of discontinuities can be used to model noise points, lines, and edges in
an image. The behavior of derivatives during transitions into and out of these
image features also is of interest.
The derivatives of a digital function are defined in terms of differences.
There are various ways to define these differences. However, we require that
any definition we use for a first derivative (1) must be zero in areas of constant
intensity; (2) must be nonzero at the onset of an intensity step or ramp; and
(3) must be nonzero along ramps. Similarly, any definition of a second deriva-
tive (1) must be zero in constant areas; (2) must be nonzero at the onset and
end of an intensity step or ramp; and (3) must be zero along ramps of constant
slope. Because we are dealing with digital quantities whose values are finite,
the maximum possible intensity change also is finite, and the shortest distance
over which that change can occur is between adjacent pixels.
A basic definition of the first-order derivative of a one-dimensional func-
tion f(x) is the difference
We return to Eq. (3.6-1)
in Section 10.2.1 and
0f
show how it follows from
= f(x + 1) - f(x) (3.6-1)
0x
a Taylor series expansion.
For now, we accept it as a
definition. We used a partial derivative here in order to keep the notation the same as
when we consider an image function of two variables, f(x, y), at which time we
will be dealing with partial derivatives along the two spatial axes. Use of a par-
tial derivative in the present discussion does not affect in any way the nature
of what we are trying to accomplish. Clearly, 0f>0x = df>dx when there is
only one variable in the function; the same is true for the second derivative.
We define the second-order derivative of f(x) as the difference

0 2f
= f(x + 1) + f(x - 1) - 2f(x) (3.6-2)
0x2
It is easily verified that these two definitions satisfy the conditions stated
above. To illustrate this, and to examine the similarities and differences between

www.EBooksWorld.ir
 3.6 ■ Sharpening Spatial Filters 159

Intensity transition a
6 b
5 Constant c
intensity Ramp FIGURE 3.36
4 Step
Intensity

Illustration of the
3
first and second
2 derivatives of a
1 1-D digital
0 x function
representing a
Scan x section of a
6 6 6 6 5 4 3 2 1 1 1 1 1 1 6 6 6 6 6
line horizontal
1st derivative 0 0
1
1
1
1
1 0 0 0 0 0 5 0 0 0 0 intensity profile
2nd derivative 0 0
1 0 0 0 0 1 0 0 0 0 5
5 0 0 0 from an image. In
5 (a) and (c) data
4 points are joined
3 by dashed lines as
a visualization aid.
2
1
Intensity

0 x

1 Zero crossing

2

3 First derivative

4 Second derivative

5

first- and second-order derivatives of a digital function, consider the example

in Fig. 3.36.
Figure 3.36(b) (center of the figure) shows a section of a scan line (inten-
sity profile). The values inside the small squares are the intensity values in
the scan line, which are plotted as black dots above it in Fig. 3.36(a). The
dashed line connecting the dots is included to aid visualization. As the fig-
ure shows, the scan line contains an intensity ramp, three sections of con-
stant intensity, and an intensity step. The circles indicate the onset or end of
intensity transitions. The first- and second-order derivatives computed
using the two preceding definitions are included below the scan line in Fig.
3.36(b), and are plotted in Fig. 3.36(c). When computing the first derivative
at a location x, we subtract the value of the function at that location from
the next point. So this is a “look-ahead” operation. Similarly, to compute the
second derivative at x, we use the previous and the next points in the com-
putation. To avoid a situation in which the previous or next points are out-
side the range of the scan line, we show derivative computations in Fig. 3.36
from the second through the penultimate points in the sequence.
Let us consider the properties of the first and second derivatives as we tra-
verse the profile from left to right. First, we encounter an area of constant inten-
sity and, as Figs. 3.36(b) and (c) show, both derivatives are zero there, so condition
(1) is satisfied for both. Next, we encounter an intensity ramp followed by a step,
and we note that the first-order derivative is nonzero at the onset of the ramp and

www.EBooksWorld.ir
160 Chapter 3 ■ Intensity Transformations and Spatial Filtering

the step; similarly, the second derivative is nonzero at the onset and end of both
the ramp and the step; therefore, property (2) is satisfied for both derivatives. Fi-
nally, we see that property (3) is satisfied also for both derivatives because the
first derivative is nonzero and the second is zero along the ramp. Note that the
sign of the second derivative changes at the onset and end of a step or ramp. In
fact, we see in Fig. 3.36(c) that in a step transition a line joining these two values
crosses the horizontal axis midway between the two extremes. This zero crossing
property is quite useful for locating edges, as you will see in Chapter 10.
Edges in digital images often are ramp-like transitions in intensity, in which
case the first derivative of the image would result in thick edges because the de-
rivative is nonzero along a ramp. On the other hand, the second derivative would
produce a double edge one pixel thick, separated by zeros. From this, we con-
clude that the second derivative enhances fine detail much better than the first
derivative, a property that is ideally suited for sharpening images.Also, as you will
learn later in this section, second derivatives are much easier to implement than
first derivates, so we focus our attention initially on second derivatives.

3.6.2 Using the Second Derivative for Image

Sharpening—The Laplacian
In this section we consider the implementation of 2-D, second-order deriva-
tives and their use for image sharpening. We return to this derivative in
Chapter 10, where we use it extensively for image segmentation. The approach
basically consists of defining a discrete formulation of the second-order deriv-
ative and then constructing a filter mask based on that formulation. We are in-
terested in isotropic filters, whose response is independent of the direction of
the discontinuities in the image to which the filter is applied. In other words,
isotropic filters are rotation invariant, in the sense that rotating the image and
then applying the filter gives the same result as applying the filter to the image
first and then rotating the result.
It can be shown (Rosenfeld and Kak [1982]) that the simplest isotropic de-
rivative operator is the Laplacian, which, for a function (image) f(x, y) of two
variables, is defined as
0 2f 0 2f
§2f = + (3.6-3)
0x2 0y2
Because derivatives of any order are linear operations, the Laplacian is a lin-
ear operator. To express this equation in discrete form, we use the definition in
Eq. (3.6-2), keeping in mind that we have to carry a second variable. In the
x-direction, we have
0 2f
= f(x + 1, y) + f(x - 1, y) - 2f(x, y) (3.6-4)
0x2
and, similarly, in the y-direction we have
0 2f
= f(x, y + 1) + f(x, y - 1) - 2f(x, y) (3.6-5)
0y2

www.EBooksWorld.ir
 3.6 ■ Sharpening Spatial Filters 161

Therefore, it follows from the preceding three equations that the discrete
Laplacian of two variables is
§2f(x, y) = f(x + 1, y) + f(x - 1, y) + f(x, y + 1) + f(x, y - 1)
-4f(x, y) (3.6-6)
This equation can be implemented using the filter mask in Fig. 3.37(a), which
gives an isotropic result for rotations in increments of 90°. The mechanics of
implementation are as in Section 3.5.1 for linear smoothing filters. We simply
are using different coefficients here.
The diagonal directions can be incorporated in the definition of the digital
Laplacian by adding two more terms to Eq. (3.6-6), one for each of the two di-
agonal directions. The form of each new term is the same as either Eq. (3.6-4) or
(3.6-5), but the coordinates are along the diagonals. Because each diagonal term
also contains a -2f(x, y) term, the total subtracted from the difference terms
now would be -8f(x, y). Figure 3.37(b) shows the filter mask used to imple-
ment this new definition. This mask yields isotropic results in increments of 45°.
You are likely to see in practice the Laplacian masks in Figs. 3.37(c) and (d).
They are obtained from definitions of the second derivatives that are the nega-
tives of the ones we used in Eqs. (3.6-4) and (3.6-5). As such, they yield equiva-
lent results, but the difference in sign must be kept in mind when combining (by
addition or subtraction) a Laplacian-filtered image with another image.
Because the Laplacian is a derivative operator, its use highlights intensity
discontinuities in an image and deemphasizes regions with slowly varying in-
tensity levels. This will tend to produce images that have grayish edge lines and
other discontinuities, all superimposed on a dark, featureless background.
Background features can be “recovered” while still preserving the sharpening

a b
0 1 0 1 1 1 c d
FIGURE 3.37
(a) Filter mask used
1
4 1 1
8 1
to implement
Eq. (3.6-6).
(b) Mask used to
implement an
0 1 0 1 1 1 extension of this
equation that
includes the
diagonal terms.
0
1 0
1
1
1 (c) and (d) Two
other implementa-
tions of the
Laplacian found

1 4
1
1 8
1
frequently in
practice.

0
1 0
1
1
1

www.EBooksWorld.ir
162 Chapter 3 ■ Intensity Transformations and Spatial Filtering

effect of the Laplacian simply by adding the Laplacian image to the original.
As noted in the previous paragraph, it is important to keep in mind which def-
inition of the Laplacian is used. If the definition used has a negative center co-
efficient, then we subtract, rather than add, the Laplacian image to obtain a
sharpened result. Thus, the basic way in which we use the Laplacian for image
sharpening is
g(x, y) = f(x, y) + c C §2f(x, y) D (3.6-7)
where f(x, y) and g(x, y) are the input and sharpened images, respectively.
The constant is c = -1 if the Laplacian filters in Fig. 3.37(a) or (b) are used,
and c = 1 if either of the other two filters is used.

EXAMPLE 3.15: ■ Figure 3.38(a) shows a slightly blurred image of the North Pole of the
Image sharpening moon. Figure 3.38(b) shows the result of filtering this image with the Lapla-
using the
cian mask in Fig. 3.37(a). Large sections of this image are black because the
Laplacian.
Laplacian contains both positive and negative values, and all negative values
are clipped at 0 by the display.
A typical way to scale a Laplacian image is to add to it its minimum value to
bring the new minimum to zero and then scale the result to the full [0, L - 1]
intensity range, as explained in Eqs. (2.6-10) and (2.6-11). The image in
Fig. 3.38(c) was scaled in this manner. Note that the dominant features of the
image are edges and sharp intensity discontinuities. The background, previously
black, is now gray due to scaling. This grayish appearance is typical of Laplacian
images that have been scaled properly. Figure 3.38(d) shows the result obtained
using Eq. (3.6-7) with c = -1. The detail in this image is unmistakably clearer
and sharper than in the original image. Adding the original image to the Lapla-
cian restored the overall intensity variations in the image, with the Laplacian in-
creasing the contrast at the locations of intensity discontinuities.The net result is
an image in which small details were enhanced and the background tonality was
reasonably preserved. Finally, Fig. 3.38(e) shows the result of repeating the pre-
ceding procedure with the filter in Fig. 3.37(b). Here, we note a significant im-
provement in sharpness over Fig. 3.38(d). This is not unexpected because using
the filter in Fig. 3.37(b) provides additional differentiation (sharpening) in the
diagonal directions. Results such as those in Figs. 3.38(d) and (e) have made the
Laplacian a tool of choice for sharpening digital images. ■

3.6.3 Unsharp Masking and Highboost Filtering

A process that has been used for many years by the printing and publishing in-
dustry to sharpen images consists of subtracting an unsharp (smoothed) ver-
sion of an image from the original image. This process, called 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.

www.EBooksWorld.ir
 3.6 ■ Sharpening Spatial Filters 163

a
b c
d e
FIGURE 3.38
(a) Blurred image
of the North Pole
of the moon.
(b) Laplacian
without scaling.
(c) Laplacian with
scaling. (d) Image
sharpened using
the mask in Fig.
3.37(a). (e) Result
of using the mask
in Fig. 3.37(b).
(Original image
courtesy of
NASA.)

-
Letting f (x, y) denote the blurred image, unsharp masking is expressed in
equation form as follows. First we obtain the mask:

gmask(x, y) = f(x, y) - f (x, y) (3.6-8)

Then we add a weighted portion of the mask back to the original image:

g(x, y) = f(x, y) + k * gmask(x, y) (3.6-9)

where we included a weight, k (k Ú 0), for generality. When k = 1, we have

unsharp masking, as defined above. When k 7 1, the process is referred to as

www.EBooksWorld.ir
164 Chapter 3 ■ Intensity Transformations and Spatial Filtering

a
b
c Original signal
d
FIGURE 3.39 1-D
illustration of the
mechanics of
unsharp masking. Blurred signal
(a) Original
signal. (b) Blurred
signal with
original shown
Unsharp mask
dashed for refere-
nce. (c) Unsharp
mask. (d) Sharp-
ened signal,
obtained by
adding (c) to (a).
Sharpened signal

highboost filtering. Choosing k 6 1 de-emphasizes the contribution of the un-

sharp mask.
Figure 3.39 explains how unsharp masking works. The intensity profile in
Fig. 3.39(a) can be interpreted as a horizontal scan line through a vertical edge
that transitions from a dark to a light region in an image. Figure 3.39(b) shows
the result of smoothing, superimposed on the original signal (shown dashed)
for reference. Figure 3.39(c) is the unsharp mask, obtained by subtracting the
blurred signal from the original. By comparing this result with the section of
Fig. 3.36(c) corresponding to the ramp in Fig. 3.36(a), we note that the unsharp
mask in Fig. 3.39(c) is very similar to what we would obtain using a second-
order derivative. Figure 3.39(d) is the final sharpened result, obtained by
adding the mask to the original signal. The points at which a change of slope in
the intensity occurs in the signal are now emphasized (sharpened). Observe
that negative values were added to the original. Thus, it is possible for the final
result to have negative intensities if the original image has any zero values or
if the value of k is chosen large enough to emphasize the peaks of the mask to
a level larger than the minimum value in the original. Negative values would
cause a dark halo around edges, which, if k is large enough, can produce objec-
tionable results.

EXAMPLE 3.16: ■ Figure 3.40(a) shows a slightly blurred image of white text on a dark gray
Image sharpening background. Figure 3.40(b) was obtained using a Gaussian smoothing filter
using unsharp
(see Section 3.4.4) of size 5 * 5 with s = 3. Figure 3.40(c) is the unsharp
masking.
mask, obtained using Eq. (3.6-8). Figure 3.40(d) was obtained using unsharp

www.EBooksWorld.ir
 3.6 ■ Sharpening Spatial Filters 165

a
b
c
d
e
FIGURE 3.40
(a) Original
image.
(b) Result of
blurring with a
Gaussian filter.
(c) Unsharp
mask. (d) Result
of using unsharp
masking.
(e) Result of
using highboost
filtering.

masking [Eq. (3.6-9) with k = 1]. This image is a slight improvement over the
original, but we can do better. Figure 3.40(e) shows the result of using Eq. (3.6-9)
with k = 4.5, the largest possible value we could use and still keep positive all the
values in the final result. The improvement in this image over the original is
significant. ■

3.6.4 Using First-Order Derivatives for (Nonlinear) Image

Sharpening—The Gradient
First derivatives in image processing are implemented using the magnitude of
the gradient. For a function f(x, y), the gradient of f at coordinates (x, y) is de-
fined as the two-dimensional column vector
We discuss the gradient
0f in detail in Section
gx 0x 10.2.5. Here, we are inter-
§f K grad( f ) K B R = D T (3.6-10) ested only in using the
gy 0f magnitude of the gradi-
ent for image sharpening.
0y
This vector has the important geometrical property that it points in the direc-
tion of the greatest rate of change of f at location (x, y).
The magnitude (length) of vector §f, denoted as M(x, y), where

M(x, y) = mag(§f ) = 2 g2x + g2y (3.6-11)

is the value at (x, y) of the rate of change in the direction of the gradient vec-
tor. 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 pixel locations in f. It is common practice
to refer to this image as the gradient image (or simply as the gradient when the
meaning is clear).

www.EBooksWorld.ir
166 Chapter 3 ■ Intensity Transformations and Spatial Filtering

Because the components of the gradient vector are derivatives, they are lin-
ear operators. However, the magnitude of this vector is not because of the
squaring and square root operations. On the other hand, the partial derivatives
in Eq. (3.6-10) are not rotation invariant (isotropic), but the magnitude of the
gradient vector is. In some implementations, it is more suitable computational-
ly to approximate the squares and square root operations by absolute values:

M(x, y) L ƒ gx ƒ + ƒ gy ƒ (3.6-12)

This expression still preserves the relative changes in intensity, but the isotropic
property is lost in general. However, as in the case of the Laplacian, the isotrop-
ic properties of the discrete gradient defined in the following paragraph are pre-
served only for a limited number of rotational increments that depend on the
filter masks used to approximate the derivatives. As it turns out, the most popu-
lar masks used to approximate the gradient are isotropic at multiples of 90°.
These results are independent of whether we use Eq. (3.6-11) or (3.6-12), so
nothing of significance is lost in using the latter equation if we choose to do so.
As in the case of the Laplacian, we now define discrete approximations to
the preceding equations and from there formulate the appropriate filter
masks. In order to simplify the discussion that follows, we will use the notation
in Fig. 3.41(a) to denote the intensities of image points in a 3 * 3 region. For

a
b c z1 z2 z3
d e
FIGURE 3.41
A 3 * 3 region of
an image (the zs z4 z5 z6
are intensity
values).
(b)–(c) Roberts z7 z8 z9
cross gradient
operators.
(d)–(e) Sobel
operators. All the
1 0 0
1
mask coefficients
sum to zero, as
expected of a
derivative 0 1 1 0
operator.

1
2
1
1 0 1

0 0 0
2 0 2

1 2 1
1 0 1

www.EBooksWorld.ir
 3.6 ■ Sharpening Spatial Filters 167

example, the center point, z5, denotes f(x, y) at an arbitrary location, (x, y); z1
denotes f(x - 1, y - 1); and so on, using the notation introduced in Fig. 3.28.
As indicated in Section 3.6.1, the simplest approximations to a first-order de-
rivative that satisfy the conditions stated in that section are gx = (z8 - z5) and
gy = (z6 - z5). Two other definitions proposed by Roberts [1965] in the early
development of digital image processing use cross differences:

gx = (z9 - z5) and gy = (z8 - z6) (3.6-13)

If we use Eqs. (3.6-11) and (3.6-13), we compute the gradient image as

M(x, y) = C (z9 - z5)2 + (z8 - z6)2 D

1>2
(3.6-14)

If we use Eqs. (3.6-12) and (3.6-13), then

M(x, y) L ƒ z9 - z5 ƒ + ƒ z8 - z6 ƒ (3.6-15)

where it is understood that x and y vary over the dimensions of the image in
the manner described earlier. The partial derivative terms needed in equation
(3.6-13) can be implemented using the two linear filter masks in Figs. 3.41(b)
and (c). These masks are referred to as the Roberts cross-gradient operators.
Masks of even sizes are awkward to implement because they do not have a
center of symmetry. The smallest filter masks in which we are interested are of
size 3 * 3. Approximations to gx and gy using a 3 * 3 neighborhood centered
on z5 are as follows:
0f
gx = = (z7 + 2z8 + z9) - (z1 + 2z2 + z3) (3.6-16)
0x
and
0f
gy = = (z3 + 2z6 + z9) - (z1 + 2z4 + z7) (3.6-17)
0y
These equations can be implemented using the masks in Figs. 3.41(d) and (e).
The difference between the third and first rows of the 3 * 3 image region im-
plemented by the mask in Fig. 3.41(d) approximates the partial derivative in
the x-direction, and the difference between the third and first columns in the
other mask approximates the derivative in the y-direction. After computing
the partial derivatives with these masks, we obtain the magnitude of the gradi-
ent as before. For example, substituting gx and gy into Eq. (3.6-12) yields

M(x, y) L ƒ (z7 + 2z8 + z9) - (z1 + 2z2 + z3) ƒ

+ ƒ (z3 + 2z6 + z9) - (z1 + 2z4 + z7) ƒ (3.6-18)

The masks in Figs. 3.41(d) and (e) are called the Sobel operators. The idea be-
hind using a weight value of 2 in the center coefficient is to achieve some
smoothing by giving more importance to the center point (we discuss this in
more detail in Chapter 10). Note that the coefficients in all the masks shown in
Fig. 3.41 sum to 0, indicating that they would give a response of 0 in an area of
constant intensity, as is expected of a derivative operator.

www.EBooksWorld.ir
168 Chapter 3 ■ Intensity Transformations and Spatial Filtering

As mentioned earlier, the computations of gx and gy are linear opera-

tions because they involve derivatives and, therefore, can be implemented
as a sum of products using the spatial masks in Fig. 3.41. The nonlinear as-
pect of sharpening with the gradient is the computation of M(x, y) involving
squaring and square roots, or the use of absolute values, all of which are
nonlinear operations. These operations are performed after the linear
process that yields gx and gy.

EXAMPLE 3.17: ■ The gradient is used frequently in industrial inspection, either to aid hu-
Use of the mans in the detection of defects or, what is more common, as a preprocessing
gradient for edge
enhancement. step in automated inspection. We will have more to say about this in Chapters
10 and 11. However, it will be instructive at this point to consider a simple ex-
ample to show how the gradient can be used to enhance defects and eliminate
slowly changing background features. In this example, enhancement is used as
a preprocessing step for automated inspection, rather than for human analysis.
Figure 3.42(a) shows an optical image of a contact lens, illuminated by a
lighting arrangement designed to highlight imperfections, such as the two edge
defects in the lens boundary seen at 4 and 5 o’clock. Figure 3.42(b) shows the
gradient obtained using Eq. (3.6-12) with the two Sobel masks in Figs. 3.41(d)
and (e). The edge defects also are quite visible in this image, but with the
added advantage that constant or slowly varying shades of gray have been
eliminated, thus simplifying considerably the computational task required for
automated inspection. The gradient can be used also to highlight small specs
that may not be readily visible in a gray-scale image (specs like these can be
foreign matter, air pockets in a supporting solution, or miniscule imperfections
in the lens). The ability to enhance small discontinuities in an otherwise flat
gray field is another important feature of the gradient. ■

a b
FIGURE 3.42
(a) Optical image
of contact lens
(note defects on
the boundary at 4
and 5 o’clock).
(b) Sobel
gradient.
(Original image
courtesy of Pete
Sites, Perceptics
Corporation.)

www.EBooksWorld.ir