Denoising camera data - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 4

The variance σ 2 ( x ) is approximated as a piecewise linear function depending on the signal
x , with the slope m ( x ) and the intercept t ( x ) based on the measurement data in Figure 1 (a).
Because of the dual-gain read-out the values for m ( x ) and t ( x ) are piecewise constant.
Based on the model for the camera noise in the raw data, we describe in the next section the
shape-adaptive DCT denoising method and how to integrate the noise model in the SA-DCT.
3 Adaptive raw data denoising
Our goal is to find an algorithm, which provides a high visual quality of the denoising results,
and which is additionally efficient in terms of hardware implementation costs. Regarding
the visual quality, a common problem of denoising algorithms is blurring of edges or ine
details in the image (oversmoothing). The shape-adaptive denoising algorithm [ 7 ] prevents
from oversmoothing by using a homogenous neighborhood for denoising. As proposed by Foi
we use the local polynomial approximation and intersection of confidence interval technique
(LPA-ICI) to find an adequate neighborhood for each pixel.
The method has been adapted for signal-dependent noise in Ref. [ 8 ] . However, it still cannot
directly be used on Bayer data, because the neighboring pixels do not have the same color due
to the Bayer mask.
To find a way of estimating the neighborhood based on the Bayer data, we apply a lumin-
ance transformation. In color image denoising, a color space transformation from RGB to a
luminance-chrominance color space (e.g. YCbCr) is usual. As the structural information in nat-
ural images is mostly contained in the luminance data, it is effective to perform the neighbor-
hood estimation on the luminance channel only and use this neighborhood for denoising all
three channels. In our case, we apply a similar strategy; we obtain an estimation of the lumin-
ance channel based on the Bayer data. We discuss this luminance transformation in the next
section.
3.1 Luminance Transformation of Bayer Data
To find a luminance estimation based on the Bayer data we tested different techniques: filter
ing with a fixed filter kernel, partial debayering, and a new method we call “virtual lumin-
ance.” Partial debayering means, we take the debayered green channel as luminance estima-
tion directly, as the green channel is most dominant for the luminance. We used the camera
debayering method (ADA), which can be applied by downloading the free “ARRIRAW Con-
verter (ARC)” tool [ 14 ] . Another low-cost luminance estimation applies filters directly to the
Bayer data. We used two different filters: a Gaussian filter kernel and a filter similar to the lu-
minance filter by Jeon and Dubois [ 13 ] . We additionally calculated a luminance directly on the
Bayer data by using the neighboring color values, which we call “virtual,” because the result
gives us luminance values which are located between the pixels.
The results on our test image “city” in Table 1 show that the difference in terms of peak
signal-to-noise ratio (PSNR) of the denoising result is marginal. The best value is reached by
the camera debayering and Gaussian filtering. We use the Gaussian filtering for our method,
as it shows one of the best results and additionally is a very simple and cost-efficient method.