The Roles of Fourier Spectrum and Phase

Angle
Shape features are evident

Phase angle is 0
DC term is dominant
Basics of Filtering in Frequency Domain
Frequency Domain Filtering Fundamentals
Low-Pass and High-Pass Filters
Correspondence to the Spatial Domain Filter

The FT of a Gaussian function is still a Gaussian function

An Example (Sobel Mask)
Image Smoothing Using Frequency Domain
Filters – Ideal Lowpass Filter
Applying the ILPF – Blurring and Ringing

Original ILPF, cutoff 10,

Energy 87%

ILPF, cutoff 60
ILPF, cutoff 30 Energy 95.7%
Energy 93.1%

ILPF, cutoff 160 ILPF, cutoff 460

Energy 97.8% Energy 99.2%
Gaussian Lowpass Filters

 D 2 (u,v) / 2 D02
H (u, v)  e
Applying the GLPF
 Image Degradation and Restoration
Image degradation due to
• noise in transmission
• imperfect image acquisition
• environmental
condition
• quality of sensor

http://fireoracleproductions.com/services samp
les

Google image
Image Degradation and Restoration
Restoration in the Presence of Noise Only –
Spatial Filtering
 Image Restoration with Additive Noise

𝑔 𝑥, 𝑦 = 𝑓 𝑥, 𝑦 + 𝜂 𝑥, 𝑦

𝐺(𝑢, 𝑣) = 𝐹(𝑢, 𝑣) + 𝑁(𝑢, 𝑣)

Noise models:
• Impulse noise: pepper and salt
• Continuous noise model:
• Gaussian, Rayleigh, Gamma, Exponential, Uniform
Properties of Noise

• Spatial properties
• Spatially periodic noise
• Spatially independent noise
• Frequency properties
• White noise – noise containing all frequencies within
a bandwidth
 Some Important Noise Model

( z z ) 2
1 
p(z)  e 2 2

2
http://www.gergltd.com/cse486/project2/
• Due to electronic circuit
• Due to the image sensor
• poor illumination
• high temperature
 Some Important Noise Model

2  ( za)
2
 a b z b1 az
p(z)   b (z  a)e za
 z0
b
e
 za p(z)   (b 1)!
0
 0 z0
Rayleigh noise
• range imaging Gamma noise
•Background model for Magnetic • laser imaging
Resonance Imaging (MRI)
images
 Some Important Noise Model

Pa for z  a
P for z  b
aeaz z0 P(z)   b
P(z)   
 0 z0
 0 Otherwise
Impulse noise
Uniform noise • salt and pepper noise
Exponential noise
• laser imaging
An Example

What is its histogram?

An Example (cont.)
An Example (cont.)
Estimation of Noise Parameters
 Degradation VS Restoration

Note: a linear, position invariant degradation system with additive

noise can be modeled as the convolution of the degradation function
with the image plus the additive noise.

Degradation (convolution)
f (x, g(x, y)
Restoration (deconvolution)
Estimate the Degradation Function

• Observation
• Experimentation
• Mathematical modeling
 Estimate Degradation Function - Observation

Assumptions:

• The degradation function is linear and position-invariant

• No other knowledge about the degradation function
Estimation by image observation:

• Extract a subimage with strong signal Higher signal-to-noise

ratio
• Perform restoration on the subimage

degradation Gs (u, v) Observed subimage

function in the H s (u, v) 
subimage F̂s (u, v) Restored subimage
H (u,v)
Application: restoring old pictures
Estimate Degradation Function - Experimentation

Assumptions:
•A similar equipment is available

•Change the system setting can achieve similar degraded images

G(u, v) Observed image
H (u, v) 
A Impulse signal
 Estimation by Modeling – Motion Blur

Constant velocity along x and y direction:

𝑎𝑡 𝑏𝑡
𝑥0 (𝑡) = 𝑦0 (𝑡) =
𝑇 𝑇
 Estimation by Modeling – Cont.

An example of motion blur

T
g(x, y)   f [x  x0 (t), y  y0 (t)]dt
0

Motion in both x and y direction during acquisition

 Estimation by Modeling – Cont.

An example of motion blur

T
H (u, v)   e  j 2 [ux0 (t )vy 0 (t )] dt
0
 Estimation by Modeling – Example

Constant velocity along x and y direction:

x0 (t)  y0 (t)  bt / T
What is 𝐻(𝑢, 𝑣)?

sin[ (ua  vb)]  j (uavb )

H (u, v)  T e
 (ua  vb)
Image Restoration

Given the degradation system H and the input image G,

recover the original image F

• Inverse filtering
• Wiener filtering
 Inverse Filtering

Ideally: In practice:
G(u, v)  H (u, v)F(u, v) G(u, v)  H (u, v)F(u, v)  N (u, v)

G(u, v) N (u, v)
F̂(u, v)  F̂(u, v)  F (u, v) 
H (u, v) H (u, v)

Limiting the analysis to

Low pass filtering frequencies near the origin
(0,0)
 An Example of Inverse Filtering

Original image

G(u, v)
H (u, v)
Degraded image
 An Example of Inverse Filtering (Cont.)
Full filter Cutoff radius = 40
Original image

Degraded image G(u, v)

H (u, v)

Cutoff radius = 70 Cutoff radius = 85

Minimum Mean Square Error (Wiener) Filtering

Assumptions:
• Noise and image are uncorrelated
• Noise has zero mean

Original Inverse filtering Wiener filtering

with cutoff = 70
 The Formulation

Minimize mean squared error: e2  E{( f  fˆ)2}

 1 H (u, v)
2

F̂ (u,v)   G(u,v)
 H (u, v) H (u,v)  | N (u, v) | / | F(u,v) | 
2 2 2

Least square error filter

2
N (u, v) is the power spectrum of noise
2
F (u, v) is the power spectrum of undegraded image
 The Formulation (Cont.)

Signal to noise ratio: the metric to evaluate the restoration performance

M 1 N 1
 F(u,v) 2

Frequency domain: SNR  u0 v0

M 1 N 1
 N (u,v) 2

u0 v0

M 1 N 1
 fˆ(x, y) 2

Spatial domain: SNR  M 1 N 1

u0 v0

[ fˆ (x, y)  f (x, y)] 2

u0 v0
 The Formulation

 1 H (u, v) 2 
F̂ (u,v)   G(u,v)
 H (u, v) H (u,v)  | N (u, v) | / | F (u,v) | 
2 2 2

An approximation
 1 H (u, v)
2

F̂(u, v)   G(u, v)
 H (u, v) H (u, v)  K 
2
Example 2 – Motion Blur + Additive Noise
Inverse Wiener
filtering filtering
Noise level