I have just finished a series of exercises building
output RAW DNG files (Bayer format) enhanced in some way with respect to their source input RAW files:
EXERCISE 1: LINEAR LIGHT DECOMPOSITION
(
https://www.overfitting.net/2021/05/descomposicion-lineal-de-fuentes-de.html)
Goal: thanks to sensor linearity we want to isolate some artificial light source by subtracting the environment RAW data from the (artificial+environment) RAW data. Interesting thoughts about photon noise in subtraction operations.
EXERCISE 2: MEAN STACKING TO EMULATE A 4-STOPS ND FILTER/ISO6
(
https://www.overfitting.net/2021/05/apilado-por-media-simulando-iso-ultra.html)
Goal: by averaging RAW data from 16 shots, we want to build a new RAW file with 2 stops of extra DR. The same technique is implemented in some cameras to mimic ND filters/ultra low ISO.
EXERCISE 3: MEDIAN STACKING TO REMOVE MOVING SUBJECTS
(
https://www.overfitting.net/2021/05/apilado-por-mediana-para-eliminar.html)
Goal: classical exercise where by calculating the median over several shots (5 in this case) in a highway, we try to remove moving cars. Even if the shots were not correctly aligned, to my surprise the median was incredibly robust in not producing Bayer false colours while preserving sharpness. The mean (right image) didn't work for this purpose.
EXERCISE 4: HDR RAW
(
https://www.overfitting.net/2021/05/raw-hdr.html)
Goal: ultra optimum (cherry picked at photodetector level) HDR blend of a 3 shots bracketing {0EV, +3EV, +6EV}. The output 16-bit DNG RAW file can be heavily processed without noise appearing.
This article is the starting point with links to the 4 exercises:
https://www.overfitting.net/2021/04/generando-un-raw-en-formato-dng-partir.html(Google translation available)
Hope you find them interesting.
Regards!