Any chance for a mean blended stacking mode? The existing NR mode leaves behind lots of artefacts.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
Anonymous
-
2021-08-21
Well, the photon noise as well as thermal noise(s) in electrical components are more random phenomena than less, therefore, IMO, averaging is proper. Also way less susceptible to major outliers. And particularly the process to evaluate the mean value is severely slow and needs the full data-set to be available simultaneously.
However, user adjustable count of frames (like from 2 to 1024) for NR averaging would be a great improvement. In order to avoid round-off errors, averaging needs more calculation bit-depth than what the pixels holds, but summing 16-bit words in just IEEE32 allows up to 65536 NR frames. The process of averaging does not need the full data-set to be available simultaneously, just the sum and the count of frames.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
At its core, NR mode is doing a weighted mean, but with many other stages: Auto-alighment. Computing mean of images, but weighted to reduce ghosting effects. Noise reduction. Colour adjustment.
So yes, options could be supplied to turn any or all of these off. I don't know if you have any examples if input and resultant images showing problems?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
Anonymous
-
2021-08-31
I for one have always had good results with NR mode, but I do hope that someday there will be a new option: User adjustable count of frames, up to very many, like 128 or even 1024 frames perhaps, then the results will be even better, of course such shooting requires still phone and subject, but I believe still life photography is quite a common practice.
Tho', I'm baffled about the stages. May I ask in what way weighting is implemented in the process of computing the mean if images? And after stage there is nose reduction, is it done by the median operator? If so, I surely would like to trade hat stage, for much more initial frames.
I have been doing noise reduction with Open Camera in such way that I take 20, 50, 100 or 200 frames in Repeat-mode, then take the set to PC/Photoshop for the process of calculating the mean (synonym: average). In CS I only utilize frame-sets that are in powers of two, so 16, 32, 64, 128 frames, due to computational problem with round-off errors, the process averages the frames of the original set in pairs, then in that result set again frames in pairs, and so on until only the single resulting noise reduced frame remains (this way the round-off errors are kept low in the limited dByte depth that is available for layered computation in the old CS).
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Allowing more photos for NR is on my todo list (it needs some thought to avoid running out of memory, e.g., processing as it goes, or maybe just being careful to limit what devices it allows this).
For the weighting, roughly if the pixels are too different (comparing the first image to a later image), it assumes a ghosting (or misalignment) effect and so reduces the influence for that image and pixel. The currently implementation is rather simplistic, but the next version (1.49) will fix this to look at a nearby group of pixels.
The noise reduction at the end is a median filter, followed by smoothing on the colour components.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
Anonymous
-
2021-09-11
Yes, the storage space requirement needs consideration. Maybe there could be a dynamic maximum limit (calculated from the quantity of currently available free memory in the device) to what the user may specify as the frame count.
When calculation is bit-depth constrained, using the by-pair process will minimize the round-off errors, and not all the frames need to be collected to memory (RAM or flash) beforehand, because every time the count of frame reach the previously aggregated (power of two), the aggregating final image can be aggregated again.So the memory requirement is about FrameCount/2-times of the memory requirement of a single frame.
But if a wider calculation space like the IEEE-64 is available then only the amount of RAM memory is needed what a single frame takes in that format, as all the frames can be simply summed-up (non-scaled) into that, and finally the IEEE-64 image is scaled down to 8 (or 16) bit bit per color, by the FrameCount.
My guess is that multi-frame NR shots are always done with steady/firm camera and over frozen scene (still-life photography), so there never is ghosting. Even the physical disturbance that is generated by tapping the shutter release icon is avoided by a small value in the shutter release delay timer.
Creating a NR final photo from a larger set of frames makes the median filter absolutely unnecessary.
In most of the scenes I see not much improvement with over 16 frames. Scenes that have large, severely dark area(s) will benefit from more frames. Well lit non-demanding scenes need no more than 8 frames. A seriously good quality result can be had with 256 frames.These have been so with all my phones, over the years I've have had many kinds of them.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Any chance for a mean blended stacking mode? The existing NR mode leaves behind lots of artefacts.
Well, the photon noise as well as thermal noise(s) in electrical components are more random phenomena than less, therefore, IMO, averaging is proper. Also way less susceptible to major outliers. And particularly the process to evaluate the mean value is severely slow and needs the full data-set to be available simultaneously.
However, user adjustable count of frames (like from 2 to 1024) for NR averaging would be a great improvement. In order to avoid round-off errors, averaging needs more calculation bit-depth than what the pixels holds, but summing 16-bit words in just IEEE32 allows up to 65536 NR frames. The process of averaging does not need the full data-set to be available simultaneously, just the sum and the count of frames.
At its core, NR mode is doing a weighted mean, but with many other stages:
Auto-alighment.
Computing mean of images, but weighted to reduce ghosting effects.
Noise reduction.
Colour adjustment.
So yes, options could be supplied to turn any or all of these off. I don't know if you have any examples if input and resultant images showing problems?
I for one have always had good results with NR mode, but I do hope that someday there will be a new option: User adjustable count of frames, up to very many, like 128 or even 1024 frames perhaps, then the results will be even better, of course such shooting requires still phone and subject, but I believe still life photography is quite a common practice.
Tho', I'm baffled about the stages. May I ask in what way weighting is implemented in the process of computing the mean if images? And after stage there is nose reduction, is it done by the median operator? If so, I surely would like to trade hat stage, for much more initial frames.
I have been doing noise reduction with Open Camera in such way that I take 20, 50, 100 or 200 frames in Repeat-mode, then take the set to PC/Photoshop for the process of calculating the mean (synonym: average). In CS I only utilize frame-sets that are in powers of two, so 16, 32, 64, 128 frames, due to computational problem with round-off errors, the process averages the frames of the original set in pairs, then in that result set again frames in pairs, and so on until only the single resulting noise reduced frame remains (this way the round-off errors are kept low in the limited dByte depth that is available for layered computation in the old CS).
Allowing more photos for NR is on my todo list (it needs some thought to avoid running out of memory, e.g., processing as it goes, or maybe just being careful to limit what devices it allows this).
For the weighting, roughly if the pixels are too different (comparing the first image to a later image), it assumes a ghosting (or misalignment) effect and so reduces the influence for that image and pixel. The currently implementation is rather simplistic, but the next version (1.49) will fix this to look at a nearby group of pixels.
The noise reduction at the end is a median filter, followed by smoothing on the colour components.
Yes, the storage space requirement needs consideration. Maybe there could be a dynamic maximum limit (calculated from the quantity of currently available free memory in the device) to what the user may specify as the frame count.
When calculation is bit-depth constrained, using the by-pair process will minimize the round-off errors, and not all the frames need to be collected to memory (RAM or flash) beforehand, because every time the count of frame reach the previously aggregated (power of two), the aggregating final image can be aggregated again.So the memory requirement is about FrameCount/2-times of the memory requirement of a single frame.
But if a wider calculation space like the IEEE-64 is available then only the amount of RAM memory is needed what a single frame takes in that format, as all the frames can be simply summed-up (non-scaled) into that, and finally the IEEE-64 image is scaled down to 8 (or 16) bit bit per color, by the FrameCount.
My guess is that multi-frame NR shots are always done with steady/firm camera and over frozen scene (still-life photography), so there never is ghosting. Even the physical disturbance that is generated by tapping the shutter release icon is avoided by a small value in the shutter release delay timer.
Creating a NR final photo from a larger set of frames makes the median filter absolutely unnecessary.
In most of the scenes I see not much improvement with over 16 frames. Scenes that have large, severely dark area(s) will benefit from more frames. Well lit non-demanding scenes need no more than 8 frames. A seriously good quality result can be had with 256 frames.These have been so with all my phones, over the years I've have had many kinds of them.