Venice

[DanK]

Ted: well, maybe.

The formula you give is based on the standard error of the mean, and in the paper you link to, it is used to quantify the signal:noise ratio when noise is on a continuous distribution. Whether this is germane, I think, depends on how the stacking algorithm works. Shot noise has a binary component, present or absent, with most values at zero. It seems to me that if the algorithm averages the images, the standard error of the mean (including the zeros) is germane. If it selects the median value for each pixel (like the median blend mode in photoshop), the standard error of the mean isn't relevant. Under most circumstances, assuming a reasonable number of images, the shot noise will disappear.

[Manfred M]

Searching the net for << "noise reduction" averaging factor >> gets us lots links to "solid work". For example:



From http://www.owlnet.rice.edu/~dodds/Fi...oise_notes.pdf

The generally accepted improvement in SNR is the square root of the number of averaged images. That is to say that the mean of e.g. 9 images will have an SNR 3x "better" than any one of those nine. The quote above mentions "complications". One complication would be the calculation of the SNR for the long exposure versus the SNR for the shorter exposure.

Anybody?



The wording of Manfred's post implies serious doubt as to any benefits of averaging multiple exposures.

This whole discussion suffers from the use of the single term "noise" without qualification and the lack of actual numbers or examples.

Pardon my pedantry ...

Thanks Ted - I will look these up.

I don't have a signal processing background so come from a disadvantage here.

In terms of my wording, I always look at two aspects of these discussion; one is the theoretical outcome and the other is the real world practical outcome. I spent a career working with pedantic engineers who loved arguing minutia and fortunately, these arguments were both in my area of knowledge and I generally had the authority to influence the outcomes of these discussions.

My approach was always the same; (1) quantity the impact this will have on the overall outcome, i.e. will this have a significant impact on the project and (2) what is this going to cost in terms of time and effort. People started to learn quite quickly that unless they could demonstrate the importance of their pedantry, I would not waste the company's time and money on it. If it made a useful difference, I was all ears and would support their contribution all the way.

In this case my question would be the same; will this methodology make a noticeable difference to final image? I'm not sure what the answer is but will probably do a side-by-side test at some point to figure this out.

[xpatUSA]

<>
In this case my question would be the same; will this methodology make a noticeable difference to final image? I'm not sure what the answer is but will probably do a side-by-side test at some point to figure this out.

A good question to which I don't have the answer either!

I have a dark corner in my living room which I shoot when I need a bit of "noise" (easy to come by with a Foveon sensor). Just need to puzzle out the actual exposure for the long shot and the actual exposure for each of the multiple shots. And for the multiple shots, are they summed or averaged? I'm guessing summed ...

[rpcrowe]

My days in Venice were so chocked full of activity, that after a fine dinner neither my wife nor I had a lot of energy left over to venture around. I know that was our loss but, that is just how it was. I wish that I had entered out at night at least once during the two nights we were in Venice...

[bambleweeney]

Ted, from what I remember from my 'experiment' with 10 shots all at the correct exposure. They were all placed in Photoshop and averaged out to 100% transparency between them. The first one being at 60% (I think, it was a long time ago) and the rest in increasing transparency to give a total of 100%. (i.e. 60, 20, 15, 5 % transparency - I forget the actual transparencies and even if the actual numbers have any meaning). All stacked up in Photoshop layers and - voila
Anyhow, good noise reduction in something like DXO Photolab is probably better and a damn lot easier. As an experiment it was interesting, practically ... a bit of a bind. For astrophotography it possibly still has some bearing.

[billtils]

Bill, I've taken lots of HDR images without a tripod ... some of my most notable failures

[DanK]

Ted, from what I remember from my 'experiment' with 10 shots all at the correct exposure. They were all placed in Photoshop and averaged out to 100% transparency between them. The first one being at 60% (I think, it was a long time ago) and the rest in increasing transparency to give a total of 100%. (i.e. 60, 20, 15, 5 % transparency - I forget the actual transparencies and even if the actual numbers have any meaning). All stacked up in Photoshop layers and - voila
Anyhow, good noise reduction in something like DXO Photolab is probably better and a damn lot easier. As an experiment it was interesting, practically ... a bit of a bind. For astrophotography it possibly still has some bearing.

Another option is to select all the layers, go to the layers menu, pick smart object, stack mode, and finally median. This will select for every pixel the median value in the stack. It will ignore exceptionally bright or dark instances--hence it will ignore noise.

I've tried this only once, but it is the same method you can use to remove unwanted moving objects, like people walking across a scene you have photographed.

[bambleweeney]

Dan yes I remember that method from an astrophotography turorial on YouTube, never thought about it for removing 'unwanted' people though. Must try that.

[maratstepanoff]

Nice photos

[xpatUSA]

Ted, from what I remember from my 'experiment' with 10 shots all at the correct exposure. They were all placed in Photoshop and averaged out to 100% transparency between them. The first one being at 60% (I think, it was a long time ago) and the rest in increasing transparency to give a total of 100%. (i.e. 60, 20, 15, 5 % transparency - I forget the actual transparencies and even if the actual numbers have any meaning). All stacked up in Photoshop layers and - voila
Anyhow, good noise reduction in something like DXO Photolab is probably better and a damn lot easier. As an experiment it was interesting, practically ... a bit of a bind. For astrophotography it possibly still has some bearing.

Thanks for the information, Paul. All at the same exposure then averaged does make sense. Did you induce 'test' noise with a high ISO setting? I can imagine that a properly exposed scene (sensor-wise) might not end up with much noise to reduce.

I don't use Adobe products nor DxO, but I do have ImageJ which can combine images mathematically so I'll try that.

[DanK]

the median stack approach ought to reduce noise more than the averaging approach, if I understand what the underlying math is, but I have never tried this comparison. Might be interesting. I'll have to create a series of really noisy images...

EDIT: I was puzzled by the opacity percentages in Paul's post and realized that I had seen a bunch of different ramps suggested, but I had never seen an explanatinon. I had never thought this through before, but I think if you want an unweighted mean (all images count equally), the opacity for each level should be 1/N, where N is position, starting from the bottom. E.g., with 4 images, the first would be 1/1=100%, the second would be 1/2=50%, etc. I had seen this ramp suggested but never explained. Here's what I think is the explanation:

Start with the first two layers. To get an unweighted mean, you would want the second to have a weight of 1/2. Set the opacity to 50%, and it has a weight of1/2. That forces the first to have a weight of 1/2. then assign the third a weight of 1/3, which forces the first two combined to have a weight of 2/3 (since they have to total 100%). Etc.

[xpatUSA]

the median stack approach ought to reduce noise more than the averaging approach, if I understand what the underlying math is <>

Didn't know what that was so I looked it up:

https://petapixel.com/2013/05/29/a-l...dian-blending/

Apparently, it's also good for eliminating moving objects!

[xpatUSA]

Didn't know what [median blending] was so I looked it up:

https://petapixel.com/2013/05/29/a-l...dian-blending/

Apparently, it's also good for eliminating moving objects!

ImageMagick can do it too:

Pat David wrote: The fast way is with Imagemagick. Once you’ve aligned all your images (or they are spot on already), put them into a directory. Then in that directory: convert *.jpg -evaluate-sequence median OUT.jpg

https://patdavid.net/2013/05/noise-removal-in-photos-with-median_6.html
.

[Cantab]

Dan and Ted, thank you both for your posts earlier today. Despite the lack of a science background, I understood what you were saying. I was previously aware of the very general process but, like some others, had thought of it only in the context of astrophotography.

[Manfred M]

I've tried this only once, but it is the same method you can use to remove unwanted moving objects, like people walking across a scene you have photographed.

That's exactly the technique I used to get this image. I had 5 or 6 handheld shots that I imported into Photoshop, aligned and then used Statistics / Median functionality to get rid of all the people in the foreground. Being a popular tourist spot, there are generally at least a few dozen people in the shot.

The "trick" is to try and shoot enough images that allow the people to move into different parts of the shot and the software will remove them. If there is too much overlap or the people aren't moving then more traditional techniques like cloning are required for the final cleanup.

I didn't remove everyone, but the few people that were too stationary area all very close to the building and can't be seen without pixel peeping.

[GrahamS]

Very nice set of images, very different to the usual Venice fare.

[Cantab]

That's exactly the technique I used to get this image. I had 5 or 6 handheld shots that I imported into Photoshop, aligned and then used Statistics / Median functionality to get rid of all the people in the foreground. Being a popular tourist spot, there are generally at least a few dozen people in the shot.

The "trick" is to try and shoot enough images that allow the people to move into different parts of the shot and the software will remove them. If there is too much overlap or the people aren't moving then more traditional techniques like cloning are required for the final cleanup.

I didn't remove everyone, but the few people that were too stationary area all very close to the building and can't be seen without pixel peeping.
...

Manfred, so I'm clear: did you use the manual technique as described by Dan in post #31?

Start with the first two layers. To get an unweighted mean, you would want the second to have a weight of 1/2. Set the opacity to 50%, and it has a weight of1/2. That forces the first to have a weight of 1/2. then assign the third a weight of 1/3, which forces the first two combined to have a weight of 2/3 (since they have to total 100%). Etc.

[Manfred M]

Manfred, so I'm clear: did you use the manual technique as described by Dan in post #31?

Not quite - I used a technique where I let Photoshop do the heavy lifting using the tools that Dan described using a median stack, but all the layers in the stack had the same opacity. I followed this up by cleaning up any areas where Photoshop did not clean things up to the level that I wanted.

I was not reducing noise, but rather getting the tools to eliminate any parts of the captures that were not identical. Same tools but ultimately using them for a different purpose.

[xpatUSA]

ImageMagick can do it too:

https://patdavid.net/2013/05/noise-removal-in-photos-with-median_6.html
.
.

I was intrigued by the median merge method discussed earlier here. Foveon is notorious for green/magenta color blotching at low exposure levels. I wondered if there was a tendency to patternation dependent on scene luminance; because patternation by definition is not 'random', 'Gaussian', 'Poisson', etc. etc.

So I bit the bullet with 7 shots of my dark corner at 3200 ISO and -1 EC. Then I converted the raws with RawDigger which applies no NR in it's RGB conversion. Then did a median merge in ImageMagick and also an average out of interest:



At bottom left is one of the 7 shots rendered by RawDigger. When I mentioned Foveon noise earlier, I was not joking! At top right, the median output from ImageMagick ... surprisingly noisy. At top left, the average output from ImageMagick showing a fair amount of noise reduction. And at bottom right, the embedded JPEG extracted from the raw file showing what most punters see OOC in Sigma's raw converter.

I conclude sadly that median merging is no good for the Foveon color blotching.

[Wavelength]

Excellent images...