Comparison Study on the Performance of Critically Sampled Pyramid vs. Over-Complete Pyramid

11
Sangeun Han and Yi Liang Dec. 4, 2000 Comparison Study on the Performance of Critically Sampled Pyramid vs. Over-Complete Pyramid

description

Comparison Study on the Performance of Critically Sampled Pyramid vs. Over-Complete Pyramid. Overview. Introduction of multiresolution and pyramid coding Critically sampled and overcomplete pyramids Bit allocation and quantization Comparison and results. Multiresolution and Pyramid Coding. - PowerPoint PPT Presentation

Transcript of Comparison Study on the Performance of Critically Sampled Pyramid vs. Over-Complete Pyramid

Page 1: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

Sangeun Han and Yi Liang

Dec. 4, 2000

Comparison Study on the Performance of Critically Sampled Pyramid vs. Over-Complete Pyramid

Page 2: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

Overview Introduction of multiresolution and

pyramid coding Critically sampled and

overcomplete pyramids Bit allocation and quantization Comparison and results

Page 3: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

Multiresolution and Pyramid Coding

r=0.09, 0.23, 0.60, 1.00 bpp

Multiresolution offers excellent coding efficiency.

Multiresolution supports scalability.

Page 4: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

Critically Sampled PyramidH1

H0

2

2H1

H0

2

2

QQ

Q

2

2

G1

G02+

2 G1

G0+Input Recon.

ChannelOctave subband coding

Original Image Decomp. Subbands

Page 5: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

Oversampled Pyramid

Page 6: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

Choosing the FiltersChoosing filters

10

15

20

25

30

35

40

45

0 1 2 3 4 5Rate [bpp]

PSN

R [d

B]

Gaussian

Decimate/Interp.

Avg./Bilinear

Page 7: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

Bit Allocation Analytical form

Critically sampled, open-loop oversampled

The “greedy” algorithm

Mnkkkk

lllll kng

ngMRr /2

2

2 )/(/

log21

Close-loop oversampled

,0

0,log21

1

0

211

2

2

lrnR

lgg

r

L

lll

lll

lll

,

11

1

1

1

1

,1

0,

lll

ll

ll

ll

lll

nnnnnnn

nnnn

where

.constdrndD

ll

l

Page 8: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

Comparison of Rate-distortion Performance

R-D performance comparison

10152025303540455055

0 1 2 3 4 5Rate [bpp]

PSN

R [d

B]

Oversampled open-loop, greedy bit alloc.

Oversampled open-loop, analytical bitalloc.Oversampled close-loop, L=5

Critially sampled,analytical bit alloc.

Critically sampled,greedy bit alloc.

In general, critically sampled outperforms oversampled in terms of rate-distortion. Reason: oversampling.

Page 9: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

ScalabilityCritically sampled Overcomplete

PSNR = 28.4 dB

PSNR = 25.8 dB

PSNR = 28.4 dB

PSNR = 22.8 dB

PSNR = 27.2 dB PSNR = 27.2 dB

PSNR = 23.8 dB PSNR = 20.9 dB

Page 10: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

Number of LayersR-D performance by using different number of levels,

open-loop oversampled

10

15

20

25

30

35

40

45

0 1 2 3 4 5Rate [bpp]

PSN

R [d

B]

L=2L=3L=4L=5L=6

R-D performance by using different number of levels, critically sampled

10

15

20

25

30

35

40

45

50

55

0 1 2 3 4 5

Rate [bpp]

PSN

R [d

B] L=2

L=3

L=4

L=5

L=6

Critically sampled Overcomplete

Page 11: Comparison Study on the Performance of  Critically Sampled Pyramid vs.  Over-Complete Pyramid

EE 368B Project Sangeun Han and Yi Liang

Conclusions Critically sampled pyramid outperforms

the over-complete pyramid in terms of rate-distortion.

Coarse images from overcomplete pyramids show better subjective quality.

Other features of overcomplete pyramid: lower complexity, freedom in choosing filters, control of quantization noise.