Difference between revisions of "Raw preprocessing"

Revision as of 17:54, 18 January 2016

1 Row noise correction using black reference columns

Using octave/matlab-like pseudocode.

1.1 For previous frame, compute:

For each row:

med_left(y)  = median(im(y,1:8))                   % median for first 8 columns
med_right(y) = median(im(y,end-7:end))             % median of last 8 columns

For the entire image:

offset_odd_left   = median(med_left(1:2:end))      % median of medians from odd left rows
offset_odd_right  = median(med_right(1:2:end))     % median of medians from odd right rows
offset_even_left  = median(med_left(2:2:end))      % median of medians from even left rows
offset_even_right = median(med_right(2:2:end))     % median of medians from even right rows

1.2 For current frame:

1. Subtract a dark frame (constant image)

im -= dark_frame

2. Fix black level offsets:

these are expected to change very slowly over time, not from frame to frame, so we are reusing them from previous frame to avoid multiple passes
subtract an offset for odd rows and another offset for even rows
the offset varies linearly from left to right
we will attempt to keep a constant black level: let's say target_black_level = 128
subtract these values from each row:

offset_odd(x)  = offset_odd_left + (offset_odd_right - offset_odd_left) * x / width - target_black_level;
offset_even(x) = offset_even_left + (offset_even_right - offset_even_left) * x / width - target_black_level;
im(1:2:end,x) -= offset_odd(x)
im(2:2:end,x) -= offset_even(x)

3. Fix dynamic row noise

this changes with every frame, and is not correlated from frame to frame
we have a noisy estimation of this row noise in the black reference columns
average black columns (16 values for each row):

mb(y) = mean(black_columns)(y) = mean(im(y, [1:8 end-7:end])

from each row, we will subtract a scalar value: k * mb(y), where k is 0.6 for gain x1:

im(y,:) -= k * mb(y)

proof that 0.6 is optimal: https://github.com/apertus-open-source-cinema/misc-tools-utilities/commit/48de47b2a544dc32bbd5a8fd7701bb44a31ea850#diff-624053a553f49c0036b4d31282e58b2fR301

Optional steps (to be tested):

apply a gain frame:

im = (im - target_black_level) .* gain_frame + target_black_level

apply a clip frame in overexposed highlights:

im(im > clip_thr) -= clip_frame(im > clip_thr)  % clip_thr is about 2500 for gain x1; smooth transition might be also needed

That's it.

Doing the math on 16 bits is preferred, and when getting back to 12 bits, add one bit of random noise (better results). See pack12/unpack12 on raw2dng for details on the implementation, and http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p3.html for the rationale behind it.

@@ Line 27: / Line 27: @@
   im(1:2:end,x) -= offset_odd(x)
   im(2:2:end,x) -= offset_even(x)
-'''3. Fix dynamic row noise''' (which changes from every frame):
+'''3. Fix dynamic row noise'''
+* this changes with every frame, and is not correlated from frame to frame
 * we have a noisy estimation of this row noise in the black reference columns
 * average black columns (16 values for each row):
   mb(y) = mean(black_columns)(y) = mean(im(y, [1:8 end-7:end])
-* from each row, we will subtract k * mb(y), where k is 0.6 for gain x1:
+* from each row, we will subtract a scalar value: k * mb(y), where k is 0.6 for gain x1:
   im(y,:) -= k * mb(y)
 * proof that 0.6 is optimal: https://github.com/apertus-open-source-cinema/misc-tools-utilities/commit/48de47b2a544dc32bbd5a8fd7701bb44a31ea850#diff-624053a553f49c0036b4d31282e58b2fR301