Image Enhancement

(C) 2002-2012 by Yu Hen Hu 1

 Image Enhancement

• Goal: to modify an image • Methods:

so that its utilization on a – Spatial domain
particular application is • Pixel processing
enhanced. – Gray level transformation:
Data independent
• A set of ad hoc tools – histogram processing:
applicable based on Data-dependent
– Arithmetic ops
viewer’s specific needs. • Spatial filtering
• No general theory on – Frequency domain filtering
image enhancement
exists.

(C) 2002-2012 by Yu Hen Hu 2

 Image Enhancement Effects

• Resizing, cropping
• Contrasts enhancement: sharpening & softening
• Edge enhancement
• Brightness adjustment, equalization
• Color adjustment, gamma correction
• Noise reduction/unwanted object removal
• Geometric adjustment, lens error correction
• Special enhancement techniques:
– Red eye removal
– Hand motion compensation, motion blur reduction

(C) 2002-2012 by Yu Hen Hu 3

 Image Enhancement:
Pixel Gray Scale Processing
 Gray-level Transforms
• Operated on individual pixel’s intensity values: s =
T(r). r: original intensity, s: new intensity
• Data independent pixel-based enhancement
method.
• Approaches
– Image negatives
– Log transform
– Power law transform
– Piece-wise linear transform

(C) 2002-2012 by Yu Hen Hu 5

 Image Negatives

• s = T(r) = L-1-r
• Similar to photo
negatives.
• Suitable for
enhancing white or
gray details in dark
background.

(C) 2002-2012 by Yu Hen Hu 6

 Log Gray-level Transform

• s = T(r) = c log(1+r) • expand dark value to enhance

details of dark area

(C) 2002-2012 by Yu Hen Hu 7

 Power Law Gray-level Transform

• s = T(r) = c rg • Gamma correction: to compensate

the built-in power law compression
due to display characteristics.

(C) 2002-2012 by Yu Hen Hu 8

Piece-wise Linear Gray-level Transform

• Allow more control on the complexity of T(r).

– Contrast stretching
– Gray-level slicing
– Bit-plane slicing

(C) 2002-2012 by Yu Hen Hu 9

Contrast Stretching

(C) 2002-2012 by Yu Hen Hu 10

Gray-level Slicing

(C) 2002-2012 by Yu Hen Hu 11

Bit-Slicing

(C) 2002-2012 by Yu Hen Hu 12

 Histogram Processing

• Data-dependent pixel-
based image enhancement
method.
• Histogram = PDF of image
pixels.
– Assumption: each image
pixel is drawn from the
same PDF independently
(i.i.d.)
– Several effects of
histograms are shown at the
right side.

(C) 2002-2012 by Yu Hen Hu 13

 Histogram Equalization

• A gray-level transformation
method that forces the
transformed gray level to
spread over the entire
intensity range.
– Fully automatic,
– Data dependent,
– (usually) Contrast enhanced
• Usually, the discrete-valued
histogram equalization
algorithm does not yield exact
uniform distribution of r
histogram.
• In practice, one may prefer
s = T (r ) = ∫ p (w)dw
r

“histogram specification” w=0

⇒ ps ( s ) = 1, 0 ≤ s ≤ 1.

(C) 2002-2012 by Yu Hen Hu 14

 Functions of Random Variables
Lemma 1. Let FR (r ) and FS ( s ) be the cdf of original
and transformed images respectively. Then for each
s T (r ), 0 ≤ r , s ≤ 1, FS ( s=
= ) Pr .{S ≤ s=} Pr .{R ≤ r=} FR ( r )
In orther words, fraction of pixels whose value R ≤ r
and fraction of pixels of transfored image whose values
S≤s=T ( r ) are the same.

(C) 2002-2012 by Yu Hen Hu 15

 Histogram Equalization
An equalized histogram ⇒ ps ( s )= 1, 0 ≤ s ≤ 1.
0 s < 0;

Equivalently, FS =
( s )  s 0 ≤ s ≤ 1;
1 s > 1.

( r ) FS=
In other words, FR= ( s ) T=
(r ) s
r
=
Thus, ( r ) FR=
s T= (r ) ∫ pR ( w)dw
w=0

(C) 2002-2012 by Yu Hen Hu 16

 Practical Considerations

• r, s ∈ {0, 1, …, L-1} instead of [0, 1].

⇒ integration is replaced by summation
• Assume # pixels in the image is N. # of pixels whose
gray scale value is nr. Then the mapping becomes
r
T (r ) = ∫ pR ( w)dw
w=0

 L −1 r 
⇒ s=  ∑
 N w=0 
nw = ( L − 1) ⋅ cdf ( r ) 

the bracket indicates rounding to nearest integer.

(C) 2002-2012 by Yu Hen Hu 17

 Histogram Equalization Example

• Consider a 5 × 5 image with L = 4. 0 0 1 1 2

1 2 3 0 1 

r 0 1 2 3
3 3 2 2 0
p(r) 6/25 7/25 7/25 5/25  
Cdf(r) 6/25 13/25 20/25 25/25 2 3 1 0 0
s 1 2 2 3 1 1 3 2 2 

• Since original image already has an 1 1 2 2 2

equalized histogram, the effect is 2 2 3 1 2 

not clear in this example. 3 3 2 2 1
 
2 3 2 1 0
 2 2 3 2 2 
(C) 2002-2012 by Yu Hen Hu 18
 More Practical Considerations

• The number of non-zero bins in the transformed

histogram is no larger than that of the original image.
• As such, the equalization process will
– Move some bins to other locations
– Combine two or more bins into one at perhaps a different
location

(C) 2002-2012 by Yu Hen Hu 19

 Histogram Modification

• One may want to convert the histogram to a target

histogram that is not uniformly distributed. Rather with
a new pdf g(s).

( r ) FS=
In this case, one has FR= ( s ) T=
(r ) g (s)
 r

Thus, T ( r )= s= g ( FR ( r ) )= g  ∫ pR ( w)dw 
−1 −1

 w=0 
Assuming g −1 ( ) exists over [ 0 1].

(C) 2002-2012 by Yu Hen Hu 20

 Histogram Matching

• Transform pdf of r to a desired pdf ps(s).

• A generalization of histogram equalization.
• Basic idea: Given pr(r) and desired pdf pz(z), find a
transform z = T(r), such that P(Z ≤ z) = P(R ≤ r).

(C) 2002-2012 by Yu Hen Hu 21

 Indirect Method

• Indirect approach:
– First equalize the
histogram using
transform s = T(r).
– Equalize the desired
histogram v = G(z).
– Set v = s to obtain the
composite transform

z = G −1 (T (r ))

Fig. 3.19

(C) 2002-2012 by Yu Hen Hu 22

Histogram Matching Example

(C) 2002-2012 by Yu Hen Hu 23

Histogram for Local Enhancement

(C) 2002-2012 by Yu Hen Hu 24

Image subtraction

Mask mode radiography

(C) 2002-2012 by Yu Hen Hu 25

Image Averaging

• Same signal, but different

noise realization.
• Averaging of many such images
will enhance SNR.

(C) 2002-2012 by Yu Hen Hu 26

 Image Enhancement:
Spatial Domain Filtering
 Spatial Filtering

I J
g (m, n) = ∑ ∑ w(i, j ) f (m − i, n − j )
i =− I j =− J

• 2D FIR filtering
– Mask filtering:
convolution of the image
with a 2D mask
– Applications to image
enhancement: • Data-dependent
• Smoothing: low pass nonlinear filters
• Sharpening: high pass – Local histogram
– Order statistic filters
• Medium filter

(c) 2002-
2006 by Yu
 Smoothing Linear Filters

I J

∑ ∑ w(i, j ) f (m − i, n − j )
i =− I j =− J
g (m, n) = I J

∑ ∑ w(i, j )
i =− I j =− J

Normalization of coefficient to ensure 0

≤ g(m,n) ≤ L-1

(c) 2002-
2006 by Yu
 Sharpening Linear Filters

• High boosting filter: • Laplacian operator:

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

– A≥1
• Derivative filter:
– Use derivatives to approximate
high pass filters. Usually 2nd
derivatives are preferred. The
most common one is the
Laplacian operator.

(c) 2002-
2006 by Yu
 Laplacian Filter for Image Enhancement

 f ( x, y ) − ∇ 2 f ( x, y ) center of mask < 0;

g ( x, y ) = 
 f ( x, y ) + ∇ f ( x, y ) center of mask > 0.
2

(c) 2002-
2006 by Yu
Gradient filters

Roberts cross-gradient operator

Sobel operator

(c) 2002-
2006 by Yu
 Local Statistic Filters

• Calculate a local statistics and • Medium filter

then replace the center pixel – Useful in removing impulsive
value with the calculated noise (salt-and-pepper noise)
statistics. without smoothing the rest of
the image.

(c) 2002-
2006 by Yu
 Image Enhancement:
Frequency Domain Processing
Image and Its Fourier Spectrum

(C) 2002-2012 by Yu Hen Hu 35

Filtering in Frequency Domain: Basic Steps

Basic Steps
1. Multiply pixel f(x,y) of the
input image by (-1)x+y.
2. Compute F(u,v), using DFT
3. G(u,v)=F(u,v)H(u,v)
4. g1(x,y)=F-1{G(u,v)}
5. g(x,y) = g1(x,y)*(-1)x+y

(C) 2002-2012 by Yu Hen Hu 36

 Notch Filter

0 u = v = 0
H (u , v) = 
1 otherwise.
• The frequency response
H(u,v) has a notch at
origin (u = v = 0).
• Effect: reduce mean
value.
• After post-processing
where gray level is
scaled, the mean value
of the displayed image
is no longer 0.

(C) 2002-2012 by Yu Hen Hu 37

Low-pass & High-pass Filtering

(C) 2002-2012 by Yu Hen Hu 38

 Gaussian Filters

• Fourier Transform pair of

Gaussian function
 H (u ) = Ae − u / 2σ
2 2


h( x) = 2π ⋅ σ ⋅ Ae − 2π σ x
2 2 2

• Depicted in figures are

low-pass and high-pass
Gaussian filters, and their
spatial response, as well as
FIR masking filter
approximation.
• High pass Gaussian filter
can be constructed from
the difference of two
Gaussian low pass filters.

(C) 2002-2012 by Yu Hen Hu 39

Gaussian Low Pass Filters

 (D(u , v) )2 
H (u , v) = exp − 
2σ 2 
 
D(u,v): distance from the origin
of Fourier transform

(C) 2002-2012 by Yu Hen Hu 40

 Ideal Low Pass Filters

• The cut-off frequency Do

determines % power are
filtered out.

• Image power as a function of distance

from the origin of DFT (5, 15, 30, 80,
230)
(C) 2002-2012 by Yu Hen Hu 41
Effects of Ideal Low Pass Filters

• Blurring can be modeled as the

convolution of a high resolution
(original) image with a low pass
filter.

(C) 2002-2012 by Yu Hen Hu 42

Ringing and Blurring

(C) 2002-2012 by Yu Hen Hu 43

Butterworth Low Pass Filters

1
H (u , v) =
1 + [D(u , v) / Do ]
2n

(C) 2002-2012 by Yu Hen Hu 44

High Pass Filters

• Ideal high pass filter

0 if D(u , v) ≤ Do
H (u , v) = 
1 otherwise.

• Butterworth high pass

filter
1
H (u , v) =
1 + [D0 / D(u , v)]
2n

• Gaussian high pass filter

 (D(u , v) )2 
H (u , v) = 1 − exp− 2 
 2 D 0 

(C) 2002-2012 by Yu Hen Hu 45

Applications of HPFs

• Ideal HPF
– Do = 15, 30, 80

• Butterworth HPF
– n = 2,
– Do = 15, 30, 80

• Gaussian HPF
– Do = 15, 30, 80

(C) 2002-2012 by Yu Hen Hu 46

Laplacian HPF

• 3D plots of the Laplacian

operator,
• its 2D images,
• spatial domain response
with center magnified,
and
• Compared to the FIR
mask approximation
∇ 2 f ( x, y ) ⇔
[ ]
− (u − M / 2 ) + (v − N / 2 ) ⋅ F (u , v)
2 2

(C) 2002-2012 by Yu Hen Hu 47

Properties of 2D DFT

(C) 2002-2012 by Yu Hen Hu 48

Properties of 2D DFT

(C) 2002-2012 by Yu Hen Hu 49