8 LinearProcessingFiltering PDF

Linear Image Processing and Filtering
n Math of 2-d linear systems

n Separability
n Shift-invariant linear systems, 2-d convolution
n Integral image for fast box-filtering
n 2-d frequency response
!
n Blurring and sharpening 1
n Zoneplate 0.8
0.6
n Image sampling/aliasing 0.4
0.2
Wiener filtering
0
n 2
3
2
0 1
wx 0
wx
n Nonlinear noise reduction/sharpening -2
-3
-2
-1
Digital Image Processing: Bernd Girod, © 2013-2018 Stanford University -- Linear Image Processing and Filtering 1
Linear image processing
n Image processing system S(.) is linear, iff superposition principle holds:
( ) ( )
S α ⋅ f #$ x, y %& + β ⋅ g #$ x, y %& = α ⋅ S f #$ x, y %& + β ⋅ S g #$ x, y %& for all α , β ∈! ( )
n Any linear image processing system can be written as
 
g = Hf Note: matrix H need not be square.
by sorting pixels into a column vector

( )
T
f = f ⎡⎣0,0 ⎤⎦ f ⎡⎣1,0 ⎤⎦  f ⎡⎣ N − 1,0 ⎤⎦ f ⎡⎣0,1⎤⎦  f ⎡⎣ N − 1,1⎤⎦   f ⎡⎣0, L − 1⎤⎦  f ⎡⎣ N − 1, L − 1⎤⎦
Impulse response
n Another way to represent any linear image processing scheme
N −1 L−1
g [α , β ] = ∑ ∑ f [ x, y ] ⋅ h [ x, α , y, β ]
x=0 y=0
impulse response
n Input: unit impulse at pixel [a,b]
$& 1 x = a ∧ y = b
f [ x, y ] = δ [ x − a, y − b ] = %
'& 0 else
n Output: impulse response
N −1 L−1
g [α , β ] = ∑ ∑ δ [ x − a, y − b ] ⋅ h [ x, α , y, β ] = h [ a, α ,b, β ]
x=0 y=0
Interpretation #1: superposition of impulse responses
δ [ x − a, y − b ] f [ a,b ] g [α , β ] = h [ a, α ,b, β ] f [ a,b ]
a
x +
α
b x
y β
a +
x
α
b x
y β +
Interpretation #2: linear combination of input values
N −1 L−1
f [ x, y ] ⋅ h [ x, α , y, β ] g [α , β ] = ∑ ∑ f [ x, y ] ⋅ h [ x, α , y, β ]
x=0 y=0
x α
å x
y β
x
α
å
x
y β
Relationship of H and h[x,α,y,β ]
y®
x® x® x®
•• •
¯a ¯a ¯a
x® x® x®
•• •
¯a ¯a ¯a
• • •
b¯ •
•
•
•
•
•
x® x® x®
•• •
¯a ¯a ¯a
Separable linear image processing
n Impulse response is separable in [x,α] and [y,β], i.e., can be written as
N −1 L−1
g [α , β ] = ∑ ∑ f [ x, y ] ⋅ hx [ x, α ] hy [ y, β ]
x=0 y=0
n Processing can be carried out
l row by row, then column by column
L−1 N −1
g [α , β ] = ∑ hy [ y, β ] ∑ f [ x, y ] ⋅ hx [ x, α ]
y=0 x=0
l column by column, then row by row

N −1 L−1
g [α , β ] = ∑ hx [ x, α ] ∑ f [ x, y ] ⋅ hy [ y, β ]
x=0 y=0
Separable linear image processing (cont.)
n If the digital input and output images are written as a matrices f and g, we can
conveniently write
g = H ⋅ f ⋅ Hx T
y
⎡ h 0,0 ⎤ ⎡ h 0,0 ⎤
y[ ] hy [ 0,1]  hy ⎡⎣ 0, Lg − 1⎤⎦ ⎢ x[ ] hx [ 0,1]  hx ⎡⎣ 0, N g − 1⎤⎦
⎥
⎢ ⎥
⎢ h 1,0 ⎥ ⎢ h 1,0 ⎥
y[ ] hy [1,1]  hy ⎡⎣1, Lg − 1⎤⎦ x[ ] hx [1,1]  hx ⎡⎣1, N g − 1⎤⎦
Hy = ⎢ ⎥ Hx = ⎢ ⎥
⎢ ⎥ ⎢    ⎥
   ⎢ ⎥
⎢ ⎥
⎢ hy [ L − 1,0 ] hy [ L − 1,1]  hy ⎡⎣ L − 1, Lg − 1⎤⎦ ⎥ ⎢ hx [ N − 1,0 ] hx [ N − 1,1]  hx ⎡⎣ N − 1, N g − 1⎤⎦ ⎥
⎣ ⎦ ⎣ ⎦
n Output image g has size Lg x Ng

n If the operator does not change image size, Hx and Hy are square matrices
Example: Separable Haar transform
Original Bike 256x256

256x256 Haar transform
Hx=Hy=Hr256
Haar transform
n Haar transform matrix for sizes N=2,4,8

æ 0 ö
1 æ 1 1 ö ç
1 1 2 0 2 0 0
÷
Hr2 =
2 çè 1 -1 ÷ø
ç 1 1 2 0 -2 0 0 0 ÷
ç ÷
ç 1 1 - 2 0 0 2 0 0 ÷
æ 1 1 2 0 ö 1 ç 1 1 - 2 0 0 -2 0 0 ÷
Hr8 = ç ÷
ç ÷ 8ç 1 -1 0 2 0 0 2 0 ÷
1 ç 1 1 - 2 0 ÷ ç ÷
Hr4 = ç 1 -1 0 2 0 0 -2 0 ÷
4ç 1 -1 0 2 ÷
÷ ç 1 -1 - 2 0 0 0 2 ÷
ç ç
0
÷
çè çè 0 0 0 -2 ÷ø
1 -1 0 - 2 ÷ø 1 -1 0 - 2
n Can be computed by taking sums and differences

n Fast algorithms by recursively applying Hr2
Example: Subsampling
n Image subsampling 2:1 horizontally and vertically
n Small input image of size 8x8, output image size 4x4
é 1 0 0 0 ù
ê ú
ê 0 0 0 0 ú
ê 0 1 0 0 ú
ê 0 0 0 0 ú
Hx = Hy = ê ú
ê 0 0 1 0 ú
ê 0 0 0 0 ú
ê 0 0 0 1 ú
ê ú
ë 0 0 0 0 û
Example: Subsampling (cont.)
n A somewhat better technique for 2:1 image size reduction
é 0.5 0 0 0 ù
ê ú
ê 0.5 0 0 0 ú
ê 0 0.5 0 0 ú
ê 0 0.5 0 0 ú
Hx = H y = ê ú
ê 0 0 0.5 0 ú
ê 0 0 0.5 0 ú
ê 0 0 0 0.5 ú
ê ú
ë 0 0 0 0.5 û
Side by side comparison
Hx = Hy = Hx = Hy =
! 1 0 0 0 $ ! 0.5 0 0 0 $
# & # &
# 0 0 0 0 & # 0.5 0 0 0 &
# 0 1 0 0 & # 0 0.5 0 0 &
# 0 0 0 0 & # 0 0.5 0 0 &
# & # &
# 0 0 1 0 & # 0 0 0.5 0 &
# 0 0 0 0 & # 0 0 0.5 0 &
# 0 0 0 1 & # 0 0 0 0.5 &
# & # &
" 0 0 0 0 % " 0 0 0 0.5 %
Another example: filtering
n Each pixel is replaced by the average of two horizontally (vertically) neighboring
pixels
⎡ 0.5 0  0 ⎤
⎢ ⎥
⎢ 0.5 0.5 ⎥
⎢ 0 0.5 0.5 ⎥
⎢ 0.5 0.5   ⎥
Hx = Hy = ⎢ ⎥
⎢ 0.5 0.5 ⎥
⎢   0.5 0.5 ⎥
⎢ 0.5 0.5 0 ⎥
⎢ ⎥
⎣ 0  0 0.5 0.5 ⎦
n Shift-invariant operation (except for image boundary)
Shift-invariant systems and Toeplitz matrices
n For a separable, shift-invariant, linear system
hx [x, a ] = hsiv/x [a - x ] and hy [y, b ] = hsiv/y [b - y ]
n Matrices Hx and Hy are square, and Toeplitz matrices, e.g.,

⎡ h [0] hsiv/x [1]  hsiv/x [ N − 1] ⎤
⎢ siv/x
⎥
⎢ hsiv/x [ −1] hsiv/x [ 0 ]  hsiv/x [ N − 2 ] ⎥
Hx = ⎢ ⎥
⎢    ⎥
⎢ hsiv/x [1− N ] hsiv/x [ 2 − N ]  hsiv/x [ 0 ] ⎥
⎣ ⎦
n Operation is a 2-d separable convolution („filtering“)
N -1 L-1
g [a , b ] = å å f [x, y ]× hsiv/x [a - x ]hsiv/y [b - y ]
x=0 y=0
Non-separable 2-d convolution
n Convolution kernel of linear shift-invariant system („filter“) can also be
non-separable
N -1 L-1
g [a , b ] = å å f [x, y ]× hsiv [a - x, b - y ]
x=0 y=0
n Viewed as a matrix operation . . .

 
g = Hf
. . . H is a block Toeplitz matrix
Structure of H for non-separable convolution
y®
x® x® x®
•• •
¯a ¯a ¯a
x® x® x®
•• •
¯a ¯a ¯a
• • •
b¯ •
•
•
•
•
•
x® x® x®
•• •
¯a ¯a ¯a
Convolution: superposition of impulse responses
f [ a,b ]δ [ x − a, y − b ] g [ x, y ] = hsiv [ x − a, y − b ] f [ a,b ]
a +
x x
b x 1 x 5
2 3 4
y
+
y a +
x x
b x 1 x 5
2 3 4
y
+
y +
Convolution: linear combination of neighboring pixel values
N -1 L-1
f [ x, y ] ⋅ hsiv [α − x, β − y ] g [a , b ] = å å f [x, y ]× hsiv [a - x, b - y ]

x=0 y=0
α
x x
4 3 2
5 x 1 å xx β
y
+
y
x x
4 3 2
5 x 1 å x β
+
y y α
Convolution examples
Original Bike blurred by convolution

Bike Impulse response „box filter“
⎛ 1 1 1 1 1 ⎞
⎜ 1 1 1 1 1 ⎟
1 ⎜ ⎟
⎜
25 ⎜
1 1 [1] 1 1 ⎟
⎜ 1 1 1 1 1 ⎟⎟
⎝ 1 1 1 1 1 ⎠
Original Bike blurred horizontally

Bike Filter impulse response
1
5 (
1 1 [1] 1 1 )
Original Bike blurred vertically

Bike Filter impulse response
⎛ 1 ⎞
⎜ 1 ⎟
1⎜ ⎟
⎜
5⎜
[1] ⎟
⎜ 1 ⎟⎟
⎝ 1 ⎠
Integral image
Σ j [x, y ] = j [x, y - 1]+ f [x, y ]

i [x, y ] = i [x - 1, y ]+ j [x, y ]
f [x, y ]
x
i [x, y ]= å å f [u,v ]
u=0 v=0
Box filtering with integral image
Σ
B D
C A
Sum = A + B – C – D
Gaussian filtering by repeated box filtering
N times
≈ ∗ … ∗
Variance σ2 σ 12 / N
n Approximate convolution with Gaussian kernel by

convolution with box kernel, repeated N times
n Convolution with box kernel can be computed efficiently
using integral image
Gaussian filtering by repeated box filtering
Direct implementation:
6561 multiplications and
6560 additions per pixel
x-y separable filtering:

162 multiplications and
160 additions per pixel
Original image Gaussian filtered

500x500 σ = 20, 81x81 kernel
20 additions or subtractions per pixel
Box filtered Box filtered Box filtered Box filtered

after N = 1 after N = 2 after N = 3 after N = 4
1-d discrete-time Fourier transform
n Given a 1-d sequence s[k ] , k ∈Z = {…,−1,0,1,2, 3,…}
n Fourier transform
∞
S e( )jω
= ∑ ⎣ ⎦
s ⎡
k=−∞
k ⎤ e − jω k
ω ∈ℜ
n Fourier transform is periodic with 2π

n Inverse Fourier transform
1
( )
π
s ⎡⎣ k ⎤⎦ =
2π ∫−π
S e jω e jω k dω
2-d discrete-space Fourier transform
n Given a 2-d array of image samples
s [m, n ] , m, n ÎZ 2
n Fourier transform
( )
∞ ∞
∑ ∑ s ⎡⎣ m,n ⎤⎦ e
jω y − jω x m− jω y n
S e jω x ,e = ω x ,ω y ∈ℜ 2
m=−∞ n=−∞
n Fourier transform is 2π-periodic both in ω x and ω y

n Inverse Fourier transform
s ⎡⎣ m,n ⎤⎦ =
1
4π 2
π
∫ ∫
−π −π
π
(
S e jω x
,e
jω y
)e jω x m+ jω y n
d ω x dω y
5x5 box filter revisited
Original Bike blurred by convolution

Bike Impulse response „box filter“
⎛ 1 1 1 1 1 ⎞
⎜ 1 1 1 1 1 ⎟
1 ⎜ ⎟
⎜
25 ⎜
1 1 [1] 1 1 ⎟
⎜ 1 1 1 1 1 ⎟⎟
⎝ 1 1 1 1 1 ⎠
Frequency response of 5x5 lowpass filter
( )
∞ ∞
∑ ∑ h ⎡⎣ m,n ⎤⎦ e
jω x jω y − jω x m− jω y n
H e ,e =
m−∞ n=−∞
1 2 2 − jω x m− jω y n 1 2 − jω x m 2 − jω y n
= ∑ ∑e
25 m=−2 n=−2
= ∑e ∑e
25 m=−2 n=−2
!
( )
1
1
0.8
=
25
(1+ 2cosω x + 2cos(2ω x )) 1+ 2cosω y + 2cos(2ω y )
0.6
0.4
0.2
Separable filter:
1-d frequency responses
0 are multiplied
2
3
2
0 1
wx -2 -2
-1
0
wx
-3
Horizontal lowpass filter
(
H e jω x ,e
jω y
) =
1
5
(1+ 2cosω x + 2cos(2ω x ))
! 1
0.8
0.6
0.4
0.2
Bike blurred horizontally 2

3
2
Filter impulse response 0 1
wx -1
wx
0
( )
-2 -2
1
[1]
-3
1 1 1 1
5
Vertical lowpass filter
(
H e jω x ,e
jω y
) (
=
1
5
1+ 2cos ω y + 2cos(2ω y ) )
1
0.8
0.6
0.4
0.2
Bike blurred vertically 2

3
2
Filter impulse response wy
0 1
wx
0
-2 -1
⎛ 1 ⎞ -3
-2
⎜ ⎟
1
1⎜ ⎟
⎜ ⎡1⎤ ⎟
5⎜ ⎣ ⎦ ⎟
⎜ 1 ⎟
⎜⎝ 1 ⎟⎠
Sharpening filter
Original Bike sharpened

Bike Filter impulse
response
æ 0 -1 0 ö
1ç ÷
-1 [8 ] -1
4 çç ÷
è 0 -1 0 ÷ø
Frequency response of sharpening filter
( )
∞ ∞
∑ ∑ h ⎡⎣ m,n ⎤⎦ e
3 H e ,e =
! m−∞ n=−∞
( )
2.5
1 − jω jω − jω jω
= 8− e x − e x − e y − e y
2 4
1 1
1.5
= 2 − cos ω x − cos ω y
2 2
1
4
2 4
2
wy
0
0
-2
-4 -4
-2
wx
More aggressive sharpening
(
H e jω x ,e
jω y
) = 5 − 2cosω x
− 2cos ω y
!
10
0
Bike sharpened 4
2 4
Filter impulse response 0 2
wx -2 -2
wx
0
æ 0 -1 0 ö -4 -4
ç -1 5 -1 ÷
ç [] ÷
ç 0 -1 0 ÷
è ø
Zoneplate pattern to visualize frequency plane
Equation to generate pattern:
s(x, y) = ŝ cos(ax x 2 + a y y 2 ) + s0
Local frequency at ( x, y )
∂
∂x
( )
ax x 2 + a y y 2 = 2ax x
∂
∂y
( )
ax x 2 + a y y 2 = 2a y y
x
⎯ ⎯⎯⎯⎯
interpolation
s ⎡⎣ x, y ⎤⎦ ←⎯⎯⎯ → s(x, y)
sampling
Sampling interpretation of 2-d discrete-space Fourier transform
n How is the Fourier transform of s[m,n] related to the Fourier transform of the
continuous signal
∞ ∞
s ( x, y ) = ∑ ∑ s !" m,n #$ δ ( x − m, y − n)2
m=−∞ n=−∞
„continuous“
continuous delta function
discrete
n Continuous-space 2-d Fourier transform


( )
∞ ∞
∑ ∑ s ⎡⎣ m,n ⎤⎦ δ 2 ( x − m, y − n) e
− jω x x− jω y y
S ω x ,ω y = ∫ ∫ dx dy
x y m=−∞ n=−∞
( )
∞ ∞
∑ ∑ s ⎡⎣ m,n ⎤⎦ e
− jω x m− jω y n jω y
= = S e jω x ,e
m=−∞ n=−∞
Lowpass filtering
Original Zoneplate (512x512) Lowpass filtered with 5x5 box filter
Zoneplate demonstration: 1-d sampling of 2-d signals
Horizontal/vertical 2:1 subsampling without prefiltering
Horizontal/vertical 2:1 subsampling with prefiltering
2d impulse response Frequency response
æ 1 1 1 ö
ç 16 8 16 ÷ 1
ç ÷
ç 1 1 1 ÷
ç 8 4 8 ÷ 0.5
|H|
ç ÷
ç 1 1 1 ÷
çè 16 8 16 ÷ø 0
2
0 2
0
-2 -2
wy wx
Image magnification (10%)
Nearest neighbor interpolation Bilinear interpolation
Optical anti-aliasing filter removal
http://www.mosaicengineering.com/products/vaf/60d/remove.html
Optical anti-aliasing filter removal
without anti-aliasing with anti-aliasing
maxmax.com
Temporal aliasing (“wagon-wheel effect”)
youtube.com/watch?v=jHS9JGkEOmA
Blurry images
Point source on focal plane maps to point
focal plane
Blurry images
Away from focal plane: out-of-focus blur
blurred point
focal plane
Blurry images
Linear + shift-invariant è convolution!
focal plane
Blurry images
Motion blur: convolution as well.
focal plane
Image deconvolution
n Given an image f [x,y] that is a blurred version of original image s[x,y], recover the
original image.
n
jω
Assume linear shift-invariant blur, transfer function B e jω x ,e y ( )
s [x, y ] (
B e jω x ,e
jω y
) f [x, y ] (
H e jω x ,e
jω y
) g [x, y ]
blur recovery
filter
n Naive solution: inverse filter

H e ( jω x
,e
jω y
) = B(e 1
jω x
,e
jω y
)
n ( )
Problem: B e jω x ,e jω y might be zero, noise amplification
Naïve deconvolution
Original (
Blurred w/ B e jω x ,e
jω y
) (
Deblurred w/ B −1 e jω x ,e
jω y
)
+ white noise rms = 2.5
Wiener filtering
n Model
s [x, y ]
filtering
& noise f [x, y ]
(
H e jω x ,e
jω y
) g [x, y ]
recovery
filter
n Minimize mean squared estimation error
{( ) } ⎯⎯⎯⎯⎯→ min.
H ⎛⎜ e
jω y ⎞
{ }
jω x
2 ,e ⎟⎠
⎝
E e ⎡⎣ x, y ⎤⎦ = E g ⎡⎣ x, y ⎤⎦ − s ⎡⎣ x, y ⎤⎦
2
n Power spectral density of estimation error

(
Φ ee e jω x ,e
jω y
) (
= Φ gg e jω x ,e
jω y
) ( ) (
− Φ gs e jω x ,e
jω y
) ( )
− Φ sg e jω x ,e
jω y
+ Φ ss e jω x ,e
jω y
= Φ ff (e jω x
,e
jω y
) H (e ,e ) H (e ,e ) − Φ (e ,e ) H (e ,e )
jω x jω y ∗ jω x jω y
fs
jω x jω y jω x jω y
− Φ ( e ,e ) H ( e ,e ) + Φ ( e ,e )
jω x jω y ∗ jω x jω y jω x jω y
sf ss
Review: Power spectrum and cross spectrum
n 2-d discrete-space cross correlation function for ergodic, stationary signals
{
ϕ fg ⎡⎣ m,n ⎤⎦ = E f ⎡⎣ x + m, y + n ⎤⎦ g ∗ ⎡⎣ x, y ⎤⎦ }
n Special case: autocorrelation function
{
ϕ ff ⎡⎣ m,n ⎤⎦ = E f ⎡⎣ x + m, y + n ⎤⎦ f ∗ ⎡⎣ x, y ⎤⎦ }
n Cross spectral density
( )
∞ ∞
∑ ∑ ϕ fg ⎡⎣ m,n ⎤⎦ e
Φ fg e ,e =
m=−∞ n=−∞
n Power spectral density

( )= ∑ ∑ϕ
∞ ∞
Φ ff e ,e ff
⎡⎣ m,n ⎤⎦ e
m=−∞ n=−∞
Wiener filtering (cont.)
n Power spectrum Φee is minimized separately at each frequency ω x , ω y if
H e ( jω x
,e
jω y
) =
Φ∗fs e ( jω x
,e
jω y
)= (
Φ sf e
jω x
,e
jω y
)
Φ ff (e jω x
,e
jω y
) Φ (e
ff
jω x
,e
jω y
)
n Can be shown to be global minimum by considering filter
H e( jω x
,e
jω y
) =
(
Φ∗fs e jω x ,e
jω y
) + ΔH (e jω x jω y
)
(e )
jω y
,e
jω x
Φ ff ,e
Wiener filter for linear distortion and additive noise
statistically
independent n [x, y ]
s [x, y ] (
B e jω x ,e
jω y
) å
f [x, y ] (
H e jω x ,e
jω y
) g [x, y ]
(
Φ sf e jω x ,e
jω y
) (
= Φ ss e jω x ,e
jω y
) ( B∗ e jω x ,e
jω y
)
Φ (e ) = Φ (e ) B(e ) ( )
2
jω x jω y jω x jω y jω x jω y jω y
ff
,e ss
,e ,e + Φ nn e jω x ,e
H e( jω x
,e
jω y
)= (
Φ ss e jω x ,e ) ( jω y
B∗ e jω x ,e
jω y
)
( ) B (e ,e ) ( )
2
jω x jω y jω x jω y jω y
Φ ss e ,e + Φ nn e jω x ,e
Wiener filter for linear distortion and additive noise
( )
Wiener Filter
jω x jω y Deblurred
Blurred w/ B e ,e ( ) B (e ,e )
( )
jω x jω y ∗ jω x jω y
Φ ss e ,e
jω x jω y
Add. white noise, rms = 2.5 B −1 e ,e (
Φ ss e jω x ,e
jω y
) B (e ,e ) + Φ (e
jω x jω y
2
nn
jω x
,e
jω y
)
Wiener filter for additive noise
statistically
independent n [x, y ]
s [x, y ]
å
f [x, y ] (
H e jω x ,e
jω y
) g [x, y ]
H e( jω x
,e
jω y
) = Φ (e (
Φ ss e jω x ,e
jω y
)
ss
jω x
,e
jω y
) (
+ Φ nn e jω x ,e
jω y
)
Wiener filter for additive noise
Original Additive
white noise
rms = 17.7
Noise reduction
by Wiener filtering
rms = 8.8
Nonlinear noise reduction/sharpening
n Noise reduction: smooth the image, lowpass filtering
n Deblurring: sharpen edges, highpass filtering
n How can both be achieved simultaneously?
n Key insight: large amplitude of highpass filtered image indicates presence of
edge
α ( f "# x, y $% ) h
HPF
(
Z e jwx , e
jw y
) fh éë x, y ùû
- fh éë x, y ùû
f [x, y] + g [x, y ]
å å
n Can be extended to multiple HPFs
Nonlinear noise reduction/sharpening (cont.)
HPF Soft coring
(
α f i "# x, y $% )
(
Z1 e
jω x
,e
jω y
) - f1 éë x, y ùû
!(.)
f [x, y ] + g [x, y ]
å å
fi éë x, y ùû
-
m
pe
HPF Soft coring
S lo
Z2 e ( jω x
,e
jω y
) f2 éë x, y ùû !(.)
n Flat areas: α ≈ 0 → H e jω x ,e ( jω y
) ) ( )
≈ 1− Z1 e jω x ,e( jω y
− Z 2 e jω x ,e
jω y
n f1 small; f2 large: H e jω x ,e ( jω y
) ≈ 1− Z (e ,e ) + ( m − 1) Z (e ,e )
1
jω x jω y
2
jω x jω y
H (e ) ≈ 1+ ( m − 1) Z (e ,e ) − Z (e ,e )
jω x jω y jω x jω y jω x jω y
n f1 large; f2 small: ,e 1 2
n Both large: (
H e jω x ,e
jω y
) ⎣
jω
( jω
) (
≈ 1+ ( m − 1) ⎡⎢ Z1 e jω x ,e y + Z 2 e jω x ,e y ⎤⎥
⎦ )
Nonlinear noise reduction/sharpening example
blurred, noisy image noise-reduced

and sharpened
Highpass filtered images
log magnitude of log magnitude of log magnitude of log magnitude of

image filtered with image filtered with image filtered with image filtered with
æ 0 0 0 ö æ 0 - 0.5 0 ö æ - 0.5 0 0 ö æ 0 0 - 0.5 ö
ç ÷ ç ÷ ç ÷ ç ÷
ç - 0.5 [1] - 0.5 ÷ ç 0 [1] 0 ÷ ç 0 [1] 0 ÷ ç 0 [1] 0 ÷
ç 0 0 0 ÷ø ç 0 - 0.5 0 ÷ ç 0 0 - 0.5 ÷ø ç - 0.5 0 0 ÷ø
è è ø è è
Soft coring function
⎛ f i ⎡⎣ x,y ⎤⎦ ⎞
γ
( )
−
α f i ⎡⎣ x, y ⎤⎦ = m ⋅ f i ⎡⎣ x, y ⎤⎦ ⋅ ⎜ 1− e ⎟
τ
⎜ ⎟
⎝ ⎠
Example:
m=3
g =2
τ = 10
(
α f i ⎡⎣ x, y ⎤⎦ )
f i ⎡⎣ x, y ⎤⎦
Soft coring of highpass filtered images
Linear vs. nonlinear noise reduction/sharpening
Noise reduction by Sharpening by highpass Combined noise reduction and

lowpass filter (linear) filter (linear) sharpening (nonlinear)

8 LinearProcessingFiltering PDF

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

8 LinearProcessingFiltering PDF

Загружено:

Авторское право:

Доступные форматы

Linear Image Processing and Filtering

n Math of 2-d linear systems

n Image sampling/aliasing 0.4

l column by column, then row by row

n Output image g has size Lg x Ng

Original Bike 256x256

n Haar transform matrix for sizes N=2,4,8

n Can be computed by taking sums and differences

n Shift-invariant operation (except for image boundary)

n Matrices Hx and Hy are square, and Toeplitz matrices, e.g.,

n Viewed as a matrix operation . . .

f [ x, y ] ⋅ hsiv [α − x, β − y ] g [a , b ] = å å f [x, y ]× hsiv [a - x, b - y ]

Original Bike blurred by convolution

Original Bike blurred horizontally

Original Bike blurred vertically

Σ j [x, y ] = j [x, y - 1]+ f [x, y ]

n Approximate convolution with Gaussian kernel by

x-y separable filtering:

Original image Gaussian filtered

20 additions or subtractions per pixel

Box filtered Box filtered Box filtered Box filtered

n Fourier transform is periodic with 2π

n Fourier transform is 2π-periodic both in ω x and ω y

Original Bike blurred by convolution

Bike blurred horizontally 2

Bike blurred vertically 2

Original Bike sharpened

n Continuous-space 2-d Fourier transform

Original Zoneplate (512x512) Lowpass filtered with 5x5 box filter

Nearest neighbor interpolation Bilinear interpolation

without anti-aliasing with anti-aliasing

Point source on focal plane maps to point

Away from focal plane: out-of-focus blur

Linear + shift-invariant è convolution!

Motion blur: convolution as well.

n Naive solution: inverse filter

+ white noise rms = 2.5

n Power spectral density of estimation error

n Power spectral density

n Can be extended to multiple HPFs

blurred, noisy image noise-reduced

log magnitude of log magnitude of log magnitude of log magnitude of

Noise reduction by Sharpening by highpass Combined noise reduction and

Вам также может понравиться