Image Processing Reference
FIGURE 4 Difference to the reference for the test image “City” denoised using (a) the al-
sion of the SA-DCT.
5 Video Sequences
When comparing video sequences we observe that both Foi's and Zhang's method show tem-
poral flickering. This flickering appears to be lower for the proposed Bayer SA-DCT. This is
probably due to the lower spatial correlation of the remaining noise, which was shown in
Figure 3 (c). Low spatial correlation means fine-grain noise, which is less visible than coars-
er grain noise due to the frequency-dependent color sensitivity of the human visual system.
However, while the correlation is lower in the proposed SA-DCT compared to the other two
methods, we still observe some temporal flickering in our results. Therefore, we extend our
method using an additional temporal denoising step.
The flickering could be a consequence of the mean calculation, which dominates the result
for large homogeneous areas during the denoising. Large homogeneous areas are likely to
have maximum size , which results in a large threshold and hence the result is mostly
dominated by the mean and the DC-coefficient. As the DC-coefficient is not noise-free, tem-
poral flickering can be observed.
To improve the visual quality in video sequences, we therefore propose to add an additional
temporal denoising step. As motion estimation is difficult on Bayer data, we propose to use
a very simple criterion for motion detection: when the deviation from a pixel in one frame to
the same pixel in the next frame is large, we do not use it, because we assume that the pixel
changed due to motion. If the pixel is similar, we use it for the temporal denoising step. A
reasonable threshold depends on the noise variance.
To reduce the flickering we average over a certain number of frames N fr . So the temporal
smoothed output S x , f for every pixel x of the current frame f can be calculated from the input
images I x , z as follows: