EE 7730

Image Enhancement
Image Enhancement
◼ The objective of image enhancement is to process an
image so that the result is more suitable than the
original image for a specific application.
◼ There are two main approaches:
❑ Image enhancement in spatial domain: Direct manipulation
of pixels in an image
◼ Point processing: Change pixel intensities
◼ Spatial filtering
❑ Image enhancement in frequency domain: Modifying the
Fourier transform of an image

Bahadir K. Gunturk 2
Image Enhancement by Point Processing
◼ Intensity Transformation

Bahadir K. Gunturk 3
Image Enhancement by Point Processing
◼ Contrast Stretching

Bahadir K. Gunturk 4
Image Enhancement by Point Processing
◼ Contrast Stretching

T (r ) = c log(1 + r )
Bahadir K. Gunturk 5
Image Enhancement by Point Processing
◼ Intensity Transformation

Bahadir K. Gunturk 6
Image Enhancement by Point Processing
◼ Intensity Transformation

Bahadir K. Gunturk 7
Image Enhancement by Point Processing
◼ Intensity Transformation

Bahadir K. Gunturk 8
Image Enhancement by Point Processing
◼ Gray-Level Slicing

Bahadir K. Gunturk 9
Image Enhancement by Point Processing
◼ Histogram

 Number of pixels with intensity r 

p(r ) =  
 Total number of pixels 

p(r )

0 r 255
Bahadir K. Gunturk 12
Histogram Specification
◼ Intensity mapping
s = T (r )

◼ Assume
❑ T(r) is single-valued and monotonically increasing.
❑ 0  T (r )  1 and 0  r  1

◼ The original and transformed intensities can be

characterized by their probability density functions (PDFs)

pr (r )
ps ( s )

Bahadir K. Gunturk 13
Histogram Specification
◼ The relationship between the PDFs is
 dr 
ps ( s ) =  pr (r ) 
 p (s)ds =  p (r )dr = 1
s r
 ds  r =T −1 ( s )

◼ Consider the mapping

r
s = T (r ) = 
w= 0
pr ( w)dw Cumulative distribution function of r

r
ds d
= 
dr dr w=0
pr ( w)dw = pr (r )

 1 
ps ( s ) =  pr (r )  = 1, 0  s 1 Histogram equalization!
 pr (r )  r =T −1 ( s )

Bahadir K. Gunturk 14
Image Enhancement by Point Processing
◼ Histogram Equalization
 Number of pixels with intensity i  r 
T (r ) = round  255 
 Total number of pixels 
 r
Number of pixels with intensity i 
= round  255 
 i = 0 Total number of pixels 
 r

= round  255 p (i ) 
 i =0 

0  r  255

Bahadir K. Gunturk 15
Image Enhancement by Point Processing
◼ Histogram Equalization Example
Intensity 0 1 2 3 4 5 6 7
Number of pixels 10 20 12 8 0 0 0 0
p (0) = 10 / 50 = 0.2
p (1) = 20 / 50 = 0.4
p(2) = 12 / 50 = 0.24
p (3) = 8 / 50 = 0.16
p(r ) = 0 / 50 = 0, r = 4,5, 6, 7
 r 
T (r ) = round  7 p(i ) 
 i =0 
T (0) = round ( 7 * p(0) ) = round ( 7 *0.2 ) = 1
T (1) = round ( 7 * ( p (0) + p(1) ) ) = round ( 7 *0.6 ) = 4
T (2) = round ( 7 * ( p(0) + p(1) + p(2) ) ) = round ( 7 *0.84 ) = 6
T (3) = round ( 7 * ( p (0) + p(1) + p(2) + p(3) ) ) = 7
T (r ) = 7, r = 4,5, 6, 7

Intensity 0 1 2 3 4 5 6 7
Number of pixels 0 10 0 0 20 0 12 8

Bahadir K. Gunturk 16
Image Enhancement by Point Processing
◼ Histogram Equalization

Bahadir K. Gunturk 17
Histogram Specification
◼ Intensity mapping
s = T (r )

◼ Assume
❑ T(r) is single-valued and monotonically increasing.
❑ 0  T (r )  1 and 0  r  1

◼ The original and transformed intensities can be

characterized by their probability density functions (PDFs)

pr (r )
ps ( s )

Bahadir K. Gunturk 18
Histogram Specification
◼ The relationship between the PDFs is
 dr 
ps ( s ) =  pr (r ) 
 ds  r =T −1 ( s )

◼ Consider the mapping

r
s = T (r ) = 
w= 0
pr ( w)dw Cumulative distribution function of r

r
ds d
= 
dr dr w=0
pr ( w)dw = pr (r )

 1 
ps ( s ) =  pr (r )  = 1, 0  s 1 Histogram equalization!
 pr (r )  r =T −1 ( s )

Bahadir K. Gunturk 19
Histogram Specification
◼ Example
−2r + 2, for 0  r  1
Assume pr (r ) = 
 0, elsewhere

Equalization mapping is then found to be

r r
s = T (r ) = 
w=0
pr ( w)dw = 
w=0
(−2r + 2)dw = −r 2 + 2r

The inverse mapping is

r = T −1 ( s ) = 1  1 − s = 1 − 1 − s

r must be between 0 and 1

Bahadir K. Gunturk 20
Histogram Specification
◼ Example

Check out the new PDF is

 dr 
ps ( s ) =  pr (r ) 


= (−2r + 2)
ds  r =T −1 ( s ) 
(d
ds
)

1− 1− s 
 r =1− 1− s
(

=  2 1− s

) 1 
 =1
2 1− s 

1
2 1− s

Bahadir K. Gunturk 21
Histogram Specification
◼ Assume we have a desired PDF pz ( z )
◼ Let the following be the equalization mappings
r
s = T (r ) = 
w= 0
pr ( w)dw

z
v = G( z) = 
w=0
pz ( w)dw

◼ Then, the desired mapping is

z = G −1 (T (r ) )

Bahadir K. Gunturk 22
Histogram Specification

Bahadir K. Gunturk 23
Histogram Specification

Bahadir K. Gunturk 24
Histogram Specification

Bahadir K. Gunturk 25
Histogram Specification

Bahadir K. Gunturk 26
Local Histogram Processing
◼ Histogram processing can be applied locally.

Bahadir K. Gunturk 27
Image Subtraction
g ( x , y ) = f ( x, y ) − h ( x, y )

The background is subtracted out, the arteries appear bright.

Bahadir K. Gunturk 28
Image Averaging
g ( x , y ) = f ( x, y ) + n ( x, y )

Corrupted Original Noise

image image

Assume n(x,y) a white noise with mean=0, and variance  2 @E n 2 ( x, y )  

If we have a set of noisy images gi ( x, y )
M
1
g
The noise variance in the average image ave ( x , y ) =
M
 g ( x, y )
i =1
i is


 1 
2



 i ( x, y) =
M M
1 1 2
E   ni ( x , y )   = 2
E n 2

 M
 i =1  
 M i =1 M

Bahadir K. Gunturk 29
Image Averaging

Bahadir K. Gunturk 30
Spatial Filtering

1 1 1
1 
1 1 1 A low-pass filter
9 
1 1 1

1 1 1
1 −8 1 A high-pass filter
 
1 1 1

Bahadir K. Gunturk 31
Spatial Filtering
◼ Median Filter
10 20 10 
 25 10 75 
  Sort: (10 10 10 20 25 75 85 90 100)
90 85 100 

◼ Example
Original signal: 100 100 100 100 10 10 10 10 10

Noisy signal: 100 103 100 100 10 9 10 11 10

Filter by [ 1 1 1]/3: 101 101 70 40 10 10 10

Filter by 1x3
100 100 100 10 10 10 10
median filter:

Bahadir K. Gunturk 32
Spatial Filtering
◼ Median filters are nonlinear.
◼ Median filtering reduces noise without blurring edges and
other sharp details.
◼ Median filtering is particularly effective when the noise
pattern consists of strong, spikelike components. (Salt-and-
pepper noise.)

Bahadir K. Gunturk 33
 Spatial Filtering
Original Salt&Pepper
noise added

3x3 3x3
averaging median
filter filter

Bahadir K. Gunturk 34
Spatial Filtering

Bahadir K. Gunturk 35
Spatial Filtering
◼ Gradient Operators
❑ Averaging of pixels over a region tends to blur detail in an
image.
❑ As averaging is analogous to integration, differentiation can
be expected to have the opposite effect and thus sharpen an
image.
❑ Gradient operators (first-order derivatives) are commonly
used in image processing applications.

Bahadir K. Gunturk 36
Spatial Filtering
◼ Gradient Operators

These are called the

Sobel operators

Bahadir K. Gunturk 37
Spatial Filtering
◼ Laplacian Operators
❑ Laplacian operators are second-order derivatives.

Bahadir K. Gunturk 38
Spatial Filtering

 −1 −1 −1
 −1 8 −1
 
 −1 −1 −1

Bahadir K. Gunturk 39
Spatial Filtering
◼ High-boost or high-frequency-emphasis filter
❑ Sharpens the image but does not remove the low-frequency
components unlike high-pass filtering

Bahadir K. Gunturk 40
Spatial Filtering
◼ High-boost or high-frequency-emphasis filter

❑ High pass = Original – Low pass

❑ High boost = (K)(Original) – Low pass

= (K-1)(Original) + Original – Low pass
= (K-1)(Original) + High pass

When K=1, High boost = High pass

When K>1, Part of the original is added back to the
highpass result.

Bahadir K. Gunturk 41
Spatial Filtering

 −1 −1 −1  −1 −1 −1
 −1 8 −1  −1 9 −1
   
 −1 −1 −1  −1 −1 −1

A high-pass filter A high-boost filter

Bahadir K. Gunturk 42
Spatial Filtering
◼ High-boost or high-frequency-emphasis filter

Bahadir K. Gunturk 43
Spatial Filtering

Bahadir K. Gunturk 44