![](bga.png)
We then applied it to extend to video BM3D and
NL-Bayes, two image denoising algorithms, to video.
We obtained a significant boost on the denoising per-
formance. This performance boost is only slightly
more costly than a local exhaustive search, including
the time spent building the tree thanks to an easy par-
allelization.
Latest contributions in video denoising advocate
for the use of 3D patches as a mechanism to im-
pose temporal consistency in the video (Protter and
Elad, 2009; Liu and Freeman, 2010a; Maggioni et al.,
2012a). Yet, in this work we showed that state-of-
the-art results can be obtained with 2D patches, using
global search. The results obtained are visually better
frame-by-frame, but can suffer from a flickering arti-
fact due to the lack of temporal consistency. This is
most noticeable for higher values of noise. Ongoing
work focuses on extending the current results to 3D
patches and video specific algorithms. One of the cur-
rent limiting factors associated to the global search is
that it increases the risk of matching the noise pattern
for patches with low SNR. We were able to alleviate
this problem in most cases by using large 2D patches,
but this causes problems with random, low-contrasted
textures which are better denoised with small patches.
3D patches can reduce the spatial patch size while still
keeping accurate distances (same dimension than the
2D patches), and therefore be more appropriate for
these types of textures.
The proposed heuristics for approximate global
patch search are not limited to the denoising appli-
cation, and could be useful for other applications re-
quiring a large number of nearest neighbors but not
requiring a dense or semi-dense NNF.
ACKNOWLEDGEMENTS
This work is supported by the ”IDI 2016” project
funded by the IDEX Paris-Saclay, ANR-11-IDEX-
0003-02. This Work is also partly founded by
BPIFrance and R
´
egion Ile de France in the FUI 18
Plein Phare project, the Office of Naval research
(ONR grant N00014-14-1-0023).
REFERENCES
Andoni, A. and Indyk, P. (2006). Near-optimal hashing al-
gorithms for approximate nearest neighbor in high di-
mensions. In Foundations of Computer Science, 2006.
FOCS’06. 47th Annual IEEE Symposium on, pages
459–468. IEEE.
Andoni, A., Indyk, P., Nguyen, H. L., and Razenshteyn, I.
(2014). Beyond locality-sensitive hashing. In Pro-
ceedings of the Twenty-Fifth Annual ACM-SIAM Sym-
posium on Discrete Algorithms, pages 1018–1028.
SIAM.
Arias, P. and Morel, J.-M. (2015). Towards a bayesian video
denoising method. In ACIVS, volume 9386 of Lecture
Notes in Computer Science, pages 107–117. Springer.
Barnes, C., Shechtman, E., Finkelstein, A., and Goldman,
D. (2009). Patchmatch: a randomized correspondence
algorithm for structural image editing. ACM Transac-
tions on Graphics-TOG, 28(3):24.
Barnes, C., Shechtman, E., Goldman, D. B., and Finkel-
stein, A. (2010). The generalized patchmatch cor-
respondence algorithm. In Computer Vision–ECCV
2010, pages 29–43. Springer.
Barnes, C., Zhang, F.-L., Lou, L., Wu, X., and Hu, S.-M.
(2015). Patchtable: Efficient patch queries for large
datasets and applications. In ACM Transactions on
Graphics (Proc. SIGGRAPH).
Bentley, J. L. (1975). Multidimensional binary search trees
used for associative searching. Communications of the
ACM, 18(9):509–517.
Buades, A., Lisani, J. L., and Miladinovi, M. (2016). Patch-
based video denoising with optical flow estimation.
IEEE Transactions on Image Processing, 25(6):2573–
2586.
Dabov, K., Foi, A., and Egiazarian, K. (2007a). Video de-
noising by sparse 3D transform-domain collaborative
filtering. In EUSIPCO, pages 145–149.
Dabov, K., Foi, A., and Egiazarian, K. (2007b). Video de-
noising by sparse 3D transform-domain collaborative
filtering. In Proc. 15th European Signal Processing
Conference, volume 1, page 7.
Dabov, K., Foi, A., Katkovnik, V., and Egiazarian, K.
(2007c). Image denoising by sparse 3-D transform-
domain collaborative filtering. Image Processing,
IEEE Transactions on, 16(8):2080–2095.
Dabov, K., Foi, A., Katkovnik, V., and Egiazarian, K.
(2007d). Image denoising by sparse 3d transform-
domain collaborative filtering. IEEE Trans. on IP,
16(8):2080–2095.
He, K. and Sun, J. (2012). Computing Nearest-Neighbor
Fields via Propagation-Assisted KD-trees. In Com-
puter Vision and Pattern Recognition (CVPR), 2012
IEEE Conference on, pages 111–118. IEEE.
Korman, S. and Avidan, S. (2011). Coherency sensitive
hashing. In Computer Vision (ICCV), 2011 IEEE In-
ternational Conference on, pages 1607–1614. IEEE.
Kumar, N., Zhang, L., and Nayar, S. (2008). What is a
good nearest neighbors algorithm for finding similar
patches in images? In Computer Vision–ECCV 2008,
pages 364–378. Springer.
Lebrun, M. (2012). An Analysis and Implementation of the
BM3D Image Denoising Method. Image Processing
On Line, 2:175–213.
Lebrun, M., Buades, A., and Morel, J.-M. (2013a). A Non-
local Bayesian Image Denoising Algorithm. SIAM
Journal on Imaging Sciences, 6(3):1665–1688.
Lebrun, M., Buades, A., and Morel, J.-M. (2013b). Imple-
mentation of the “Non-Local Bayes” (NL-Bayes) Im-
Global Patch Search Boosts Video Denoising
133