A GPU-Accelerated Real-Time NLMeans Algorithm for Denoising Color Video Sequences - Advanced Concepts for Intelligent Vision Systems - page 55

ter estimates can be found for the denoised pixel intensities. Remarkable is also
the huge acceleration of the GPU compared to the CPU of a factor 200 to 400.
The main reason lies in the massive amount of parallellism in the NLMeans
algorithm, which can be fully exploited by the GPU but hardly by the CPU.
Especially this huge acceleration leads to a real-time denoising filter. We can de-
termine the optimal parameters for the algorithm by selecting a minimum frame
rate and by maximizing the output PSNR of the filter for this minimum frame
rate. For our results in Table 1, an optimal combination is a 7 × 7-search window
and D past =1, in order to attain a frame rate of 25 frames per second (fps).
4Conluon
In this paper, we have shown how the traditional NLMeans algorithm can be
eciently mapped onto a parrallel processing architecture such as the GPU. We
saw that a naive straightforward implementation inevitably leads to an inecient
algorithm with a huge number of parallel processing passes. We then analyzed
our NLMeans algorithmic acceleration techniques from previous work, and we
noted that these techniques can not be applied “as is”. Therefore, we adapted the
core ideas of these acceleration techniques (i.e. the moving averaging filter for
the fast computation of Euclidean distances and the exploitation of the weight
symmetry) to GPGPU programming methodology and we arrived at a GPU-
NLMeans algorithm that is two to three orders of magnitudes faster (depending
on the parameter choices) than the equivalent CPU algorithm. This technique
can process video sequences in real-time on a mid-range GPU.
References
1. Rudin, L., Osher, S.: Total variation based image restoration with free local con-
straints. In: IEEE Int. Conf. Image Proc (ICIP), vol. 1, pp. 31-35 (November 1994)
2. Portilla, J., Strela, V., Wainwright, M., Simoncelli, E.P.: Image denoising using
scale mixtures of gaussians in the wavelet domain. IEEE Trans. Image Process-
ing 12(11), 1338-1351 (2003)
3. Dabov, K., Foi, A., Katkovnik, V., Egiazarian, K.: Image denoising by sparse
3d transform-domain collaborative filtering. IEEE Trans. Image Processing 16(8),
2080-2095 (2007)
4. Goossens, B., Luong, H., Pižurica, A., Philips, W.: An improved Non-Local Means
Algorithm for Image Denoising. In: Int. Workshop on Local and Non-Local Approx.
in Image Processing (2008) (invited paper)
5. Goossens, B., Pižurica, A., Philips, W.: Removal of correlated noise by modeling
the signal of interest in the wavelet domain. IEEE Trans. Image Processing 18(6),
1153-1165 (2009)
6. Goossens, B., Pižurica, A., Philips, W.: Image Denoising UsingMixtures of Projected
Gaussian Scale Mixtures. IEEE Trans. Image Processing 18(8), 1689-1702 (2009)
7. Brailean, J.C., Kleihorst, R.P., Efstraditis, S., Katsaggeleos, K.A., Lagendijk, R.L.:
Noise reduction filters for dynamic image sequences: a review. Proc. IEEE 83(9),
1272-1292 (1995)