EE 583 Lecture06

EE-583: Digital Image Processing
Image Enhancement in the Spatial Domain

Example 1(PR3.1): Exponentials of the form e-ar2, a a positive constant, are
useful for constructing smooth gray-level transformation functions. Construct the
transformation functions having the general shapes shown in the following
figures. The constants shown are input parameters, and your proposed
transformations must include them in their specifications.
s=T(r)
s=T(r)
s=T(r)
A/2
B/2
L0
L0
(a)
(a) General form of the function: s

In Figure (a):
Ae
aL0
A/ 2
L0
(b)
T (r ) Ae
solving for a :
(c)
ar 2
aL0 ln(0.5) 0.693

2
a 0.693 / L20
Then:
s T (r ) Ae
0.693 2
r
L20
Prepared By: Dr. Hasan Demirel, PhD

s=T(r)
s=T(r)
Example 1(PR3.1):
D
B
B/2
C
L0
In Figure (b):
B(1 e
) B/2
s T (r ) B(1 e
0.693
2 r2
L0
(c) General form of the function:
L0
(c)
(b)
s T( r ) B Be
(b) General form of the function:

aL20
Then:
a r 2
B( 1 e
a r 2
aL0 ln(0.5) 0.693

2
a 0.693 / L20
)
s T (r ) ( D C )(1 e
0.693 2
r
L20
)C

Example 2(PR3.3): Propose a set of gray-level-slicing transformations
capable of producing all the individual bit planes of an 8-bit monochrome
image.
Consider mod(x,y) to be the modular division resulting the remainder of the integer
division x/y.
7
255
for
2
r
For Bit plane 7 (MSB): the following
T (r )
function can be written
otherwise
0
For Bit plane 6:
255
T (r )
0
for 26 mod( r ,27 )

otherwise
For Bit plane 5:
255
T (r )
0
for 25 mod( r ,26 )

otherwise

Example 2(PR3.3):
For Bit plane 4:
For Bit plane 3:

For Bit plane 2:
For Bit plane 1:
For Bit plane 0:
255
T (r )
0
255
T (r )
0
255
T (r )
0
255
T (r )
0
255
T (r )
0
for 24 mod( r ,25 )

otherwise
for 23 mod( r ,24 )
otherwise
for 22 mod( r ,23 )
otherwise
for 21 mod( r ,22 )
otherwise
for 20 mod( r ,21 )
otherwise

Example 2(PR3.3):
255
T (r )
0
Consider Bit plane 6:
for 26 mod(r , 27 )
otherwise
250
200
s=T(r)
Note that all the 1s corresponding to bit
6 are scaled to 255 and all the bits
corresponding to 0s are scaled down to
0.
150
s
100
50
50
31=00011111
100
150
200
111=01101111
190=10111110
250
250=11001110

Example 3(PR3.4): a) What effect would setting to zero the lower-order bit
planes have on the histogram of an image in general?
b) What would be the effect on the histogram if we set to zero the higherorder bit planes instead?
a) Removing the low order bit planes would mean the loss of some high frequency details.
Furthermore the image histogram will be more sparse as compared with the all 8-bit
plane case.
This is because, there will be no component representing intermediate pixel values
such as 1,2,3,4, 5, 6,7 and 9,10,11,12,13,14,15 etc. Instead there will be 0 and 8 and 16 etc.
This would cause the height some of the remaining histogram peaks to increase in general.
Typically, less variability in gray level values will reduce contrast.
2000
1000
1500
800
1000
600
500
400
200
0
0
0
0
50
100
150
200
250
All the bits included
50
100
150
200
2 of the LSBs removed
250

Example 3(PR3.4): b) What would be the effect on the histogram if we set
to zero the higher-order bit planes instead?
b) Removing the high order bit planes would mean the loss of some very
important DC components away from the image.
The meaning of this is that the image is much darker and a lot of
the low frequency components will be lost.
1200
1000
1000
800
800
600
600
400
400
200
200
0
0
50
100
150
200
250
All the bits included
50
100
150
200
250
MSB removed

Example 4(PR3.6): Suppose that a digital image is subjected to histogram
equalization . Show that a second pass of histogram equalization will produce
exactly the same result as the first pass?
Let n be the total number of pixels and let nrj be the number of pixels in the input image with
intensity value rj . Then, the histogram equalization transformation is:
k
sk T (rk )
j 0
nr j
1 k
nr j
n n j 0
Since every pixel (and no others) with value rk is mapped to value sk, it follows that nsk = nrk . A
second pass of histogram equalization would produce values vk according
to the transformation:
k
vk T ( sk )
ns j
, but ns j nr j , then
j 0 n
k n
r
vk T ( sk ) j sk
j 0 n
Which shows that a second pass of histogram equalization would yield the same result
as the first pass.

Example 4(PR3.6): Given an image, the following histograms and CDF (T(r) graphs can
be obtained.
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
T(r)=CDF of the first pass
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
T(r)=CDF of the 2nd pass

(~identity transformation)
0.5
Note that the Identity Transformation does not

change the histogram of the input image
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
6000
Histogram of the 2nd pass

(~uniformly distributed)
6000
5000
Histogram of the first pass
5000
4000
4000
3000
3000
2000
2000
1000
1000
50
100
150
200
250
50
100
150
200
250

Example 6(PR3.22): The three images shown below were blurred using the square
averaging masks of sizes n=23, 25 and 45 respectively. The vertical bars on the lower part
of (a) and (c) are blurred, but clear separation exists between them. However, the bars
have merged in in image (b), in spite of the fact that the mask that produced the image is
significantly smaller than the mask produced image (c) .Explain this.
(Note that the vertical bars are 5 pixels wide and 20 pixels apart.)
Original image
(a)
(b)
(c)

Example 6(PR3.22):
Note that the vertical bars are 5 pixels wide and 20 pixels apart.
(a)
(b)
(c)
The reason why the mask with size n=25 producing merged uniform region around the bars is
because of the sizes of the bars and the separation of the bars in the horizontal direction. The
width of each bar is 5 pixels and each bar is separated by 20 pixels.
In such an environment as the mask moves in the horizontal direction there will be 5 black and
20 light gray pixels in each row at a time. This will provide the same average value for each pixel
in the region producing a merged uniform gray level.
However this will not be the same in 23 and 45 pixel masks!!

Example 7(PR3.24): In a given application an averaging mask is applied to input
images to reduce noise, and then a Laplacian mask is applied to enhance small details.
Would the result be the same if the order of these operations were reversed?
Laplacian operation and averaging can be expressed by the following 3x3 masks:
Laplacian g ( x, y) f ( x 1, y) f ( x 1, y) f ( x, y 1) f ( x, y 1) 4 f ( x, y)
1 9
Averaging h( x, y ) fi
9 i 1
1/9
1/9
1/9
1/9
1/9
1/9
-4
1/9
1/9
1/9
Averaging Mask
Laplacian Mask
Both of the two operators are multiplying the pixels in a 3x3 neighborhood with constant
numbers and perform the addition. Therefore, these operations are linear operations.
The order of two linear operations does not matter. The result would be same in any order.
Image Enhancement in the Frequency Domain

Example 8(PR3.25) :Show that the Laplacian operation is isotropic (invariant to
rotation). You will need the following equations relating coordinates after axis rotation
by an angle q.
x x cos q y sin q
y x sin q y cos q
Laplacian operator is defined as:
For the rotated Laplacian operator:

Given that:
If we show that the right sides of the first 2 equations are equal than the Laplacian operation
is rotation invariant.
We start with,
Image Enhancement in the Frequency Domain

Example 8(PR3.25 ):
Taking the partial derivative of this expression again with respect to x yields:
Repeat the same operation for y :
Taking the partial derivative of this expression again with respect to y yields:
Adding the 2 expressions for the second derivatives:
Both sides are equal, Hence Laplacian is rotational invariant.

Example 9(PR4.7): What is the source of nearly periodic bright points in the
horizontal axis of the spectrum in the following figure.
The nearly periodic bright points in the frequency spectrum corresponds to the periodic bars as well as
the repeating boxes, letters and circles in the horizontal direction.

Example 10(PR4.6) -a): Prove the validity of the following Equation.
[ f ( x, y)(1) x y ] F (u M / 2, v N / 2) , [.] denotes the Fourier Transform
1
F( u,v )
MN
M 1 N 1
1
F( u M / 2,v N / 2 )
MN
M 1 N 1
MN
f ( x, y )e
j 2 ( ux / M vy / N )
x 0 y 0
f ( x, y )e
j 2 (( u M / 2 )x / M ( v N / 2 ) y / N )
x 0 y 0
M 1 N 1
f ( x, y )e
j 2 (( ux / M 1 / 2 x ) ( vy / N 1 / 2 y ))
x 0 y 0
e j 2 (( ux/ M 1/ 2 x )(vy / N 1/ 2 y )) e j ( x y ) e j 2 (ux/ M vy / N ) (1) x y e j 2 (ux/ M vy / N )

e j( x y ) ( 1 )x y e j( x y ) cos( ( x y )) j sin( ( x y ))
always zero
1 or -1 depending on the addition of x+y. If (x+y) is even then 1, otherwise -1.
MN
M 1 N 1
f ( x, y )( 1 )
x 0 y 0
x y
e j 2( ux / M vy / N )

Example 10(PR4.6) -b): Prove the validity of the following 2 Equations.
f ( x, y)e j 2 (u0 x / M v0 y / N ) F (u u0 , v v0 )
f ( x x0 , y y0 ) F (u, v)e j 2 (ux0 / M vy0 / N )
1
[ f ( x, y )e j 2 (u0 x / M v0 y / N ) ]
MN
1
MN
f ( x, y)e
M 1 N 1
j 2 ( u0 x / M v0 y / N )
j 2 ( ux / M vy / N )
x 0 y 0
M 1 N 1
f ( x, y )e j 2 ([ u u0 ] x / M [ v v0 ] y / N )
x 0 y 0
F (u u0 , v v0 )
Similarly, it can be shown that :
[ F (u, v)e j 2 (ux0 / M vy0 / N ) ] f ( x x0 , y y0 )
This is the Translation property of the 2D Fourier transform:
When x0=u0=M/2 and x0=u0=N/2, then
f ( x, y)(1) x y F (u M / 2, v N / 2)
f ( x M / 2, y N / 2) F (u, v)(1)u v

Example 11(PR4.9): Consider the images shown below. The image on the right is
obtained by (a) multiplying the image on the left by (-1)x+y ; (b) computing the DFT; (c)
taking the complex conjugate of the transform; (d) computing the inverse DFT; and
(e) multiplying the real part of the result by (-1)x+y . Explain (mathematically) why the
image on the right appears as it does.
The complex conjugate simply changes j to j in the inverse transform, so
the image on the right is given by:
M 1 N 1
[ F (u, v)*] F (u, v)e j 2 (ux / M vy / N )

u 0 v 0
M 1 N 1
F (u, v)e j 2 (u ( x ) / M v ( y ) / N )
u 0 v 0
f ( x, y)
Which simply mirrors f(x,y) about the origin, thus producing the image on
the right

Example 12(PR4.14): Suppose that you form a low pass filter that averages the
four immediate neighbors of a point (x,y), but excludes the point itself.
(a) Find the equivalent filter H(u,v) in the frequency domain.
(b) Show that H(u,v) is a lowpass filter.
a) The spatial average is given by:
Then, using the following property:
f ( x x0 , y y0 ) F (u, v)e j 2 (ux0 / M vy0 / N )
Where the H(u,v) is the filter function. We get the following transfer function:

Example 12(PR4.14): Suppose that you form a low pass filter that averages the
four immediate neighbors of a point (x,y), but excludes the point itself.
(a) Find the equivalent filter H(u,v) in the frequency domain.
(b) Show that H(u,v) is a lowpass filter.
a) The H(u,v) Filter function can be centered by:
b) Consider one variable for convenience. As u ranges from 0 to M, the value of cos(2[u-M/2]/M)
starts at -1, peaks at 1 when u = M/2 (the center of the filter) and then decreases to -1 again when u = M.
Thus, we see that the amplitude of the filter decreases as a function of distance from the origin of the
centered filter, which is the characteristic of a lowpass filter.
A similar argument is easily carried out when considering both variables simultaneously.

Example 8-PR4.19: Derive the frequency domain filter that corresponds to the
Laplacian operator in the spatial domain.
Consider theLaplacian mask given. Then,
Where
The H(u,v) Filter function can be centered by:

Example 8-PR4.22: The two fourier spectra shown are of the same image. The spectrum
of the left corresponds to the original image, and the spectrum on the right was obtained after the
image was padded with zeros.
(a) Explain the difference of the overall contrast.
(b) Explain the significant increase in signal strength along the vertical and the horizontal axes of
the spectrum shown on the right.
(a) Padding with zero increases the size but reduces the average
gray level of image. The average gray level of the padded image is
less than the original image. F(0,0) in the padded image is less than
F(0,0) of the original image. All the others away from the origin are
less in the padded image than the original image. This produces a
narrower range of values hence a lower contrast spectrum in the
padded image.
(b) Padding with 0s introduces significant discontinuities at the borders of the original image. This
process introduces strong horizontal and vertical edges where the image ends abruptly and then continues
with o values. These sharp transitions correspond to the strength of the spectrum along the horizontal and
vertical axes.
Image Restoration
Restoration in the presence of Noise: Only-Spatial Filtering
Example 1-PR5.10: Given the two subimages below. The sub image on the
left is the result of using arithmetic mean filter of size 3x3. The other
subimage is the result of using the geometric mean filter of the same size.
a) Why the subimage obtained with geometric filtering is less blurred. Hint you
can start your analysis with 1-D step edge profile of the image.
b) Explain why the black components on the right image are thicker.
a) Lets consider the mathematical expressions of the arithmetic and geometric

mean filters:
1
1
f ( x, y )
g ( s, t )
mn s ,tS xy
mn
f ( x, y ) g ( s, t )
s ,tS xy
Image Restoration
Example 1-PR5.10 :
a) If we take a rough estimate of 1-D
Step function profile before the filtering:
Arithmetic mean filtering:
Before filtering:
6666006666
After filtering:
6664224666
f ( x, y ) 1
g ( s, t )
mn s ,tS xy
Geometric mean filtering:
Before filtering:
6666006666
After filtering:
6660000666
f ( x, y )
g ( s, t )
s ,tS xy
1
mn
Image Restoration
Example 1-PR5.10 :
a) The resulting 1-D profiles clearly indicates that the arithmetic mean filter
produces smoother/blurred transition and
b) The geometric filtering increases the thickness of the black components.
Arithmetic mean filtering:
Before filtering:
6666006666
After filtering:
6664224666
Geometric mean filtering:

Before filtering:
6666006666
After filtering:
6660000666

EE 583 Lecture06

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

EE 583 Lecture06

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

EE 583 Lecture06

Uploaded by

Copyright:

Available Formats

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

(a) General form of the function: s

aL0 ln(0.5) 0.693

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

(c) General form of the function:

(b) General form of the function:

aL0 ln(0.5) 0.693

Prepared By: Dr. Hasan Demirel, PhD

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

For Bit plane 6:

for 26 mod( r ,27 )

For Bit plane 5:

for 25 mod( r ,26 )

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

For Bit plane 3:

For Bit plane 1:

For Bit plane 0:

for 24 mod( r ,25 )

Prepared By: Dr. Hasan Demirel, PhD

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

Consider Bit plane 6:

Prepared By: Dr. Hasan Demirel, PhD

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

All the bits included

2 of the LSBs removed

Prepared By: Dr. Hasan Demirel, PhD

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

All the bits included

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

T(r)=CDF of the first pass

T(r)=CDF of the 2nd pass

Note that the Identity Transformation does not

Histogram of the 2nd pass

Histogram of the first pass

Prepared By: Dr. Hasan Demirel, PhD

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

Prepared By: Dr. Hasan Demirel, PhD

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

Prepared By: Dr. Hasan Demirel, PhD

EE-583: Digital Image Processing

Image Enhancement in the Spatial Domain

Prepared By: Dr. Hasan Demirel, PhD

EE-583: Digital Image Processing

Image Enhancement in the Frequency Domain

Laplacian operator is defined as:

For the rotated Laplacian operator: