Denoising camera data - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 10

Table 5
Spatial Correlation of the Difference Images (Green Channel)

C Foi C Zhang C SA−DCT

The numbers correspond to the correlation of the noise at one pixel position to the noise at the neighboring pixel
position. The numbers are given for a 3 × 3 neighborhood.

4.2 Processing Real Camera Data
Applying image processing methods from literature to real camera data can lead to artifacts,
which are due to the different characteristic of the real camera data compared to the standard
test data. An investigation of the differences and resulting artifacts using debayering methods
is given in Ref. [ 10 ] . When we used the method denoised by Zhang [ 4 ] , we observed similar
artifacts like in Ref. [ 10 ] in the highlights, shown in Figure 4 (a). To solve this, we first conver-
ted the data using a logarithmic transformation into the 12-bit logarithmic format as described
in Ref. [ 10 ] . Figure 4 (b) shows the result with the logarithmic transformation: the artifacts
around the highlights are almost not visible anymore. However, there are strong color shifts
as it can be seen in the “U” sign and in the botom right corner in Figure 4 (b). The result of
the proposed method is shown in Figure 4 (c). We additionally calculated the PSNR of the real
camera sequences using a reference image calculated as the mean image over 240 frames of a
static sequence. The PSNR is clearly higher, as the proposed method does not show the men-
tioned color artifacts and effectively removes the noise in the image.