Page 218 -

Section 6.4 Image Denoising 186

FIGURE 6.17: Denoising images artificially corrupted with additive Gaussian noise. Left:
noisy images. Right: restored ones using LSSC. Note that the algorithm reproduces the
original brick texture in the house image (σ = 15) and the hair texture for the man image
(σ = 50), both hardly visible in the noisy images. Reprinted from “Non-local Sparse Models
for Image Restoration,” by J. Mairal, F. Bach, J. Ponce, G. Sapiro, and A. Zisserman,
Proc. International Conference on Computer Vision, (2009). c 2009, IEEE.

6.4.4 Results
The three methods discussed in this section all give very good results, with a slight
edge to BM3D and learned simultaneous sparse coding (or LSSC), according to the
experiments of Dabov, Foi, Katkovnik, and Egiazarian (2007) and Mairal, Bach,
Ponce, Sapiro, and Zisserman (2009), which use standard images with added Gaus-
sian noise (Figure 6.17) for quantitative evaluation. Qualitative results on a photo-
graph corrupted by real noise are shown in Figure 6.18. The image was taken by a
Canon Powershot G9 digital camera at 1,600 ISO with a short time exposure. At
this setting, pictures are typically quite noisy. This time, we compare the original
JPEG output of the camera, and results from Adobe Camera Raw 5.0, the DxO
Optics Pro 5.3 package for professional photographers, and LSSC.