Title: Pyramid Coder with Nonlinear Prediction
1Pyramid Coder with Nonlinear Prediction
Panu Chaichanavong
- Burt/Adelson pyramid coder
- A nonlinear prediction
- Aliasing effect
- A switching method
- Conclusion
2Burt/Adelson Pyramid Coder
Introduced by Burt and Adelson (1983)
Gaussian
Laplacian
Quantized
and transmitted
Original image
Reconstructed image
3A Nonlinear Prediction
By Florencio and Schafer (1994)
Filter Just subsample!
Interpolator
- Replication
- Weighted median of 6 neighbor pixels
- Median of 4 neighbor pixels
1 2 1 2 3 2 1 2 1 2 3 2 1 2 1
Lower resolution image
Predicted image
4A Nonlinear Prediction (2)
Burt/Adelson coder
with nonlinear prediction
5Aliasing Effect
Test image
6Aliasing Effect (2)
Gaussian filter
Only subsampling
7A Switching Method
Lets look at the following 4x4 filters
.032 .058 .058 .032 .058 .102 .102
.058 .058 .102 .102 .058 .032 .058 .058 .032
0 0 0 0 0 0 0 0 0 0 1 0 0 0
0 0
0 0 0 0 0 1 2 0 0 3 4 0 0 0
0 0
Filter3 Linear low-pass with weight shown above
Filter2 Median of average
Filter1 Pick one !
8A Switching Method (2)
Input1
Input2
Input1 may be an edge, use filter1 Input2
high frequency, use filter3
How can we distinguish these inputs?
9A Switching Method (3)
Let D sum of difference of adjacent pixel
values sd standard deviation of 16 pixel
values p D/sd
Decision criterion p lt 18, use filter1 18 lt p lt
22, use filter2 p gt 22, use filter3
10A Switching Method (4)
Filtered subsampled image
Filter used
Filter1
Filter2
Filter3
11A Switching Method (5)
Method PSNR (dB) Bit-rate (bit/pixel)
B/A 32.77 0.8278
Nonlinear 34.00 0.6544
Switch 33.58 0.6501
Reconstructed image
12A Switching Method (6)
Images in the Gaussian pyramid
Filter used
Filter1
Filter2
Filter3
13Conclusion
- Low-pass filter reduces aliasing effect but gives
blurred image - Some nonlinear prediction preserves edges and
details but may introduces annoying aliasing - A decision criterion is presented to switch among
various filters to select an appropriate one for
a particular input