Video Denoising Based on Spatiotemporal Gaussian Scale Mixture Model

 

              

Video Denoising, June 2007

              

Neighboring wavelet coefficients in still images exhibit strong correlations across space, scale and orientation. In video signals, co-located wavelet coefficients from adjacent frames are highly correlated as well. Following the idea of Gaussian Scale Mixture (GSM) modeling for still image denoising [Portilla'03], we employ GSM to model the statistics of 3D blocks of wavelet coefficients. Our noise robust motion estimation scheme further enhances the temporal correlations between wavelet coefficients. A Bayesian least square estimation approach is then applied for video denoising, resulting in competitive performance in terms of both PSNR and SSIM evaluations.

 

 

 

Publications

-          G. Varghese and Z. Wang, “Video denoising based on a spatiotemporal Gaussian scale mixture model,” IEEE Transactions on Circuit and Systems for Video Technology, vol. 20, no. 7, pp. 1032-1040, July 2010. [PDF, 571K]

-          G. Varghese and Z. Wang, “Video denoising using a spatiotemporal statistical model of wavelet coefficients,” IEEE International Conference on Acoustics, Speech, & Signal Processing, Las Vegas, Nevada, Mar. 30 – Apr. 4, 2008. [PDF, 334K]

Software Download

STGSM.zip

 

 

Results

Download Frame by frame PSNR&SSIM results

 

 

 

 

 

PSNR (dB) results of STGSM algorithm.

Noise(σ)

Foreman
(352x288x100)

Salesman
(352x288x100)

Miss America
(360x288x100)

Tennis
(352x240x100)

Garden
(352x240x100)

Football
(352x240x100)

10

36.74

38.04

40.57

34.05

31.48

32.11

15

34.98

36.03

39.40

31.97

29.08

29.87

20

33.72

34.61

38.50

30.59

27.49

28.36

50

27.80

26.87

31.62

25.85

22.23

23.58

100

21.54

20.87

22.55

20.40

16.78

19.66

 

SSIM results of STGSM algorithm.

Noise(σ)

Foreman
(352x288x100)

Salesman
(352x288x100)

Miss America
(360x288x100)

Tennis
(352x240x100)

Garden
(352x240x100)

Football
(352x240x100)

10

0.937

0.960

0.952

0.894

0.950

0.913

15

0.917

0.941

0.943

0.841

0.925

0.865

20

0.901

0.923

0.936

0.797

0.900

0.820

50

0.820

0.727

0.892

0.642

0.747

0.597

100

0.715

0.496

0.823

0.464

0.363

0.364

 

PSNR (dB) results STGSM algorithm on QCIF sequences .

Input PSNR(dB)

Salesman
(176x144x449)

Miss America
(176x144x108)

Suzie
(176x144x150)

Trevor
(176x144x90)

Foreman
(176x144x300)

28

37.93

41.43

38.36

37.17

36.85

24

35.17

 

36.21

34.68

34.37

 

 

 

Created: June 10, 2010

Last Updated: Oct. 5, 2010