Denoising camera data - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 13

Table 6
SA-DCT Denoising Results without the Additional Temporal Denoising Step

Sequence ASA value PSNR PSNRHVS VIF MSSIM

Color test chart 800 43.72 44.31 0.4912 0.9963

Night Odeonsplaz 800 41.18 39.33 0.7175 0.9949

Siegestor night 1600 39.89 39.22 0.6280 0.9911

Table 7
SA-DCT Denoising Results with the Additional Temporal Denoising Step

Sequence ASA value PSNR PSNRHVS VIF MSSIM

Color test chart 800 44.63 44.88 0.5008 0.9976

Night Odeonsplaz 800 41.77 39.41 0.7308 0.9958

Siegestor night 1600 40.15 38.87 0.6492 0.9924

The metrics are calculated and averaged over 10 frames.

The metrics are calculated and averaged over 10 frames.

5.1 Implementation Aspects
Denoising algorithms like BM3D usually assume that the complete noisy image is available at
each pixel position. In stream-based image processing, as it is usual in embedded systems and
communication, only one pixel at a time is available. In these applications an additional buffer
must be implemented to provide the neighborhood for the denoising step. For algorithms like
BM3D therefore the memory cost is quite high. Our method requires only a local neighbor-
hood for the denoising step and is therefore beter suited for stream-based processing. Addi-
tionally it operates on Bayer data and the raw Bayer data has only one value per pixel whereas
the processed RGB data has three values per pixel. Thus, a three times lower complexity can
be expected compared to algorithms that require the fully processed image data.
While the temporal denoising step implementation requires additional frame buffers, the
computational cost of the calculation is kept very low.

6 Conclusion
We proposed a method for real camera Bayer data denoising based on a neighborhood estim-
ation combined with a shape-adaptive DCT. While the method has been proposed for RGB for
grayscale image data, it could not be applied to Bayer data directly. To perform the SA-DCT
on Bayer data we apply LPA-ICI-based neighborhood estimation to the luminance data. As the
luminance data is not available in the Bayer data, we estimate the luminance efficiently using
diferent methods. The best tradeoff between computational cost and quality was found using
Gaussian filtering. Based on the neighborhood estimation a hard thresholding is performed