Further Thoughts on Demosaicing Methods in Free Open Source Raw Processors
I suggested in Part 1 that I would also compare the performance of different demosaicing methods on noisy images. I now believe that this would be impractical as any simple comparison could give misleading results and any meaningful comparison would be very time-consuming.
The basic problem is that there are too many variables, and these variables are not well defined.
For example, in Part 1 we saw that with regard to “false colour” (the only criterion by which methods could be distinguished visually) igv gave excellent results with no “suppression steps” whilst amaze required 2 steps to be comparable. Should we infer from this that igv is “better” than amaze? I think not. We are comparing different algorithms, “suppression steps” may well have a different meaning in the context of igv to that of amaze; the suppression process, whatever it involves, may be inherent in and inseperable from the igv demosaicing algorithm.
It is therefore not likely to be possible to reach a simple conclusion of the form “Demosaicing Method A is best”.
A similar situation exists with regard to hot pixels (reported in Part 1). The images shown there implied that amaze was “best” as it did not spread the hot pixel’s effect to adjacent pixels. On the other hand the spreading of the effect and the consequent reduction in the brightness of the hot pixel might be considered beneficial as it represents a form of spatial low-pass filter. But this is of no practical significance if the hot/dead pixel filter is applied.
Perhaps it is the physicist in me that is reluctant to pursue further what I now see as soft science.
I shall, however, continue to use, evaluate and report on FLOSS applicable to photography.