A GPU-Accelerated Real-Time NLMeans Algorithm for Denoising Color Video Sequences - Advanced Concepts for Intelligent Vision Systems - page 54

standard deviation 25 / 255 (input PSNR 20 . 17dB). We compare the processing
time of our proposed GPU implementation to the modestly optimized (single-
threaded) C++ CPU implementation from our previous work [4] (including all
acceleration techniques proposed in [4]), for different values of the parameters
A and D past . For these results, we use D future =0(resulting in a zero-delay
denoising filter, as explained in Section 2.5), B =4(corresponding to 9 × 9
patches) and we manually select h to optimize the PSNR ratio. In particular, we
use h =0 . 13 for A

≤ 3 and h =0 . 16 for A> 3 (note that the pixel intensities
are within the range 0 − 1).
Both the CPU and GPU version were run on the same computer, which
is equipped with a 2.4GHz Intel Core(2) processor with 2048 MB RAM and a
NVidia GeForce 9600GT GPU. This card has 64 parallel stream processing units
and is considered to be a mid-range GPU.The GPU algorithm is implemented
as a HLSL pixel shader in DirectX 9.1 (Windows XP) and makes use of 16-
bit floating point values. The main program containing the GPU host code, is
written in C# 3.0.
Processing time and output PSNR results (obtained after denoising) are re-
ported in Table 1. We only report PSNR results for the GPU denoising technique,
since both CPU and GPU algorithms essentially compute the same formula (i.e.
equation (2)). It can be seen that the PSNR values increase when using a larger
search window or a larger number of past frames. This is simply because more
similar candidate blocks become available for searching, and consequently bet-

Tabl e 1 . Experimental results for denoising a color video sequence, consisting of 99
frames of dimensions 720 × 480 and corrupted with additive stationary white Gaussian
noise with standard deviation 25 / 255 (PSNR=20.17dB)

Parameters GPU CPU GPU vs. CPU

A Search window D past FPS msec/frame PSNR [dB] msec/frame acceleration

2 5x5 0 100.00 10.00 33.09 4021 402 . 10 ×

2 5x5 1 69.57 14.37 34.42 N/A

2 5x5 2 50.79 19.69 34.88 N/A

2 5x5 3 40.34 24.79 35.04 N/A

3 7x7 0 52.46 19.06 35.40 7505 393 . 70 ×

3 7x7 1 30.38 32.92 36.19 N/A

3 7x7 2 21.43 46.67 36.26 N/A

3 7x7 3 16.58 60.31 36.33 N/A

5 11x11 0 18.46 54.17 36.34 18230 336 . 55 ×

5 11x11 1 10.22 97.81 37.11 N/A

5 11x11 2 7.07 141.35 37.26 N/A

5 11x11 3 5.42 184.48 37.23 N/A

10 21x21 0 4.32 231.56 36.79 50857 219 . 63 ×

10 21x21 1 2.36 424.27 37.20 N/A

10 21x21 2 1.62 615.73 37.24 N/A

10 21x21 3 1.24 805.83 37.17 N/A