Denoising camera data - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 7

(7)
We showed how to apply neighborhood estimation on Bayer data and discussed how this
neighborhood can be used for denoising with SA-DCT hard thresholding. In the following we
evaluate our method and compare it to other state-of-the-art methods.
4 Experiments: Image quality vs system performance
We compare our method to the state-of-the-art denoising method, BM3D [ 5 ] , which was spe-
ciically adapted for raw data [ 6 ] , and the PCA-based method from Ref. [ 4 ] .
While we also tested our method on real camera data, we compare our method quantitat-
ively using simulated camera video sequences, as this provides us a realistic reference. This
test method was described in Ref. [ 2 ] and we use it for our data similarly: we rendered the
high-resolution image data and applied the camera simulation including the optical low-pass
of the camera optics. We obtain a simulated reference image and including the camera noise
added to the sensor data, we obtain realistic noisy images. These images are then denoised
and compared to the reference. We tested the method using two different debayering meth-
ods: the ARRI debayering method, which is implemented in the camera processing tool freely
available in the internet [ 14 ] , and a method called “demosaicing with directional filtering and
a posteriori decision” (DDFAPD) [ 15 ] .
We perform a test with different noise levels. In a digital camera the noise is signal-depend-
ent corresponding to the characteristic described in Section 2 . To obtain different noise levels,
we change the simulated brightness of the image and subsequently process the images with a
higher ASA level to obtain comparable results. This leads to a higher noise level, as the ASA
operates as a gain: The higher the ASA value, the higher the amplification and thus the lower
the signal-to-noise-ratio. Three different noise levels were simulated: ASA 800, which means a
low noise level, ASA1600, and ASA 3200, which corresponds to a quite high noise level.
4.1 Visual Quality of Denoising Results
We first compare the method quantitatively, thus we calculate quality metrics enabling us to
do so. We calculate the PSNR, as it is a very usual metric, and we use three additional ob-
jective quality metrics that according to Ref. [ 2 ] correlate beter with the human perception of
visual quality: a PSNR adapted to the human visual system (PSNR-HVS) [ 16 ] , the structural
similarity index (SSIM) [ 17 ] and the visual information fidelity (VIF) [ 18 ] . Tables 2 - 4 show the
results for 800 ASA, 1600 ASA and 3200 ASA respectively. While the Bayer SA-DCT reaches
the highest VIF for the 800 ASA sequences, for higher ASA values, thus higher noise levels, the
ClipFoi [ 6 ] method reaches the highest metric results, while SA-DCT and PCA [ 4 ] show ap-
proximately the same results. We conclude that our method achieves competitive results with
respect to usual quality metrics.