Tampere University of Technology - Institute of Signal Processing - Transforms and Spectral Methods Group

Denoising of Video corrupted with Additive Gaussian Noise.

This page represents results of our work in the field of denoising of video signals corrupted with additive Gaussian noise. Recently, we proposed a novel 3D DCT-based video-denoising algorithm, 3D Sliding Window DCT (3D-SWDCT). The experimental results demonstrate that our algorithm provides competitive results with other state-of-art video denoising methods both in terms of PSNR and subjective quality.

Brief description of 3D-SWDCT algorithm

Video data is locally filtered in sliding 3D windows (arrays) consisting of highly correlated spatial layers taken from consecutive frames of video. Denoising in local windows is performed by a hard thresholding of 3D DCT coefficients of each 3D array.  Final estimates of reconstructed pixels are obtained by a weighted average of the local estimates from all overlapping windows. Please, see a sketch below.

Figure. 1. A general block-diagram of proposed video denoising algorithm

If you are interesting, more details can be found in the related article.

Simulations and denoising results

In our simulations we chose processing a 3D processing array to be of size of 8x8x8 due to existed and well developed software and hardware solutions for 8-point DCT. This 3D array was chosen to be sliding over a video data with the steps equal to 2 in both spatial directions (horizontal and vertical) and 1 in the temporal direction. Hard threshold was chosen to be equal to 2 deviation of noise. As a possible solution for selection of highly correlated 2D planes, we have used a fast block matching algorithm in the pixel domain (so called “logarithmic search”) with a minimal absolute error (MAD) as a cost function.

You can download and use in your experiments our software.

Download executables: Sliding Window 3D DCT Video Denoising (MS-DOS/MS Windows)

Table 1 and Figures 2, 3 present some results of denoising for several video sequences with a wide range of Gaussian noise deviations.

Table 1. Results of denoising in terms of average PSNR, dBs

Video

“Tennis”1

“Salesman” 2

“Flower” 1

Noise,

10

15

20

10

15

20

10

15

20

Noisy Video,

PSNR dBs

28.16

24.63

22.15

28.15

24.72

22.35

28.34

24.88

22.44

Denoised Video, PSNR, dBs

33.34

30.80

29.52

37.01

34.83

33.29

31.25

28.62

26.80

1. Average over 52 frames

2. Average over 44 frames

                                                            (a)                                                                                 (b)

Figure. 2. A fragment of the 30th frame of the “Salesman” video sequence. (a) Noisy (PSNR of fragment 22.29 dBs). (b) Denoised with the proposed algorithm (PSNR of fragment 33.06 dBs)

                                                            (a)                                                                                 (b)

Figure.3. A fragment of the 30th frame of the “Flower” video sequence. (a) Noisy (PSNR of fragment 22.41dBs). (b) Denoised with the proposed algorithm (PSNR of fragment 26.39 dBs)

If you are interesting, you can download processed video sequences (48MBs).


Dmytro Rusanovskyy, 8 March 2006 .