Wiener Filtering

1
Wiener Filtering

• Greg Kastan
• EE 133
• May 1, 2006

2
LPI System
TIME DOMAIN
FREQUENCY DOMAIN
3
Direct Inverse Filtering

• Ideally works for no additive noise
• Not usually realistic

Add noise and re-estimate F
What if H is very small??
4
Minimum Mean Square Error (MMSE) or Wiener

• Define error image
• Square error
• Take expected value
• Minimize expression

TURNS OUT
5

• Image power spectrum often unknown
• Noise power to image power ratio replaced with
  constant K
• K is chosen visually for best looks

We must know H!
6

• User moves slider to select K
• Press run for results

7
Motion Blurring
8
Add Gaussian Noise
9
A Physical ComparisonDirect Inverse and Wiener
10
Conclusions

• Wiener filtering is NOT perfect!
• Removal of blur is evident
• Noise statistics play a huge role
• Large s2 still corrupts image
• Poor performance of inverse filtering emphasized
• Noise CANT be neglected in accurate system
  models