Long exposure noise reduction

[sk66]

I was reading about combining multiple underexposed shots to get a good exposure in order to avoid sensor signal noise.
i.e. to get a proper exposure from 4 images yo would underexpose each by 2EV.

What I don't understand is, if each image is going to capture the same information (same scene all exposed -2EV), why not just combine one underexposed image 4 times? It would certainly eliminate the potential layers alignment issue. Has anyone done this? Can you explain what the difference might be?

[sybren]

There are two types of noise. There is sensor noise that is nearly the same from one photo to the next. Then there is the noise that is random in each shot. By combining the same photo taken multiple times, you keep what is the same in each shot and average out the colour differences between shots.

This process only reduces the random noise. The option "long exposure noise reduction" that your camera has works on the noise that's more-or-less the same between shots. It takes a photo, then closes the shutter and takes another "photo" of the same ISO and exposure time. The second photo only contains noise. By substracting it from the first photo you remove noise.

Both techniques work, but on different noise.

[sk66]

No, the technique is to combine several shorter/ underexposed images so that there is no noise to start with (or significantly less).

In camera noise reduction and standard "reference photo" subtraction have drawbacks...
This technique eliminates those issues and gives much better results.

But what I don't understand is why I need multiple images which are "the same" as opposed to combining one image multiple times...

[sybren]

It's not really polite to start a reply with "no" when someone tries to explain something to you. I've answered your question - it's all about random noise. When adding the same image four times you also add the noise four times. When adding four different photos of the same thing you have 4x the same thing + four different noise patterns. That's significantly different from 4x the same image + 4x the same noise pattern.

[sk66]

Originally Posted by sybren

It's not really polite to start a reply with "no" when someone tries to explain something to you. I've answered your question - it's all about random noise. When adding the same image four times you also add the noise four times. When adding four different photos of the same thing you have 4x the same thing + four different noise patterns. That's significantly different from 4x the same image + 4x the same noise pattern.

Ok, then I'm really not understanding. You were explaining two things: using a dark frame (either in camera or post), and image averaging. The question was not concerning using a dark frame, and I don't think this method is "Image Averaging".

As I understand this technique you take the multiple images underexposed so that the noise isn't there to start with. By underexposing you minimize/eliminate sensor heat buildup and therefore the sensor noise isn't ever developed/captured . It is sometimes called "image stacking" or "image summing". In the end, since each image is added at 100%, you end up with 400% for a 4 image stack. (that's why each is exposed at only 25% of the desired exposure for a 4 image stack)

Conversely, "image averaging" uses multiple exposures all properly exposed and all have developed random sensor noise. Therefore they need the averaging technique applied to get rid of the noise. In the end, by only adding a percentage of each layer in the stack the final is 100%.

Obviously I could be completely off in my understanding of the differences and that's why I'm asking. But if I'm not..why not stack one image 4 times?

(I do understand that summing a large number of layers has the benefit of minimizing errant light that might show up in only one image (such as a plane))

[sybren]

Originally Posted by sk66

You were explaining two things: using a dark frame (either in camera or post), and image averaging. The question was not concerning using a dark frame, and I don't think this method is "Image Averaging".

Indeed, I was explaining two very similar methods of noise reduction. The aim was to make it clear that there are different types of noise.

As I understand this technique you take the multiple images underexposed so that the noise isn't there to start with.

There is always noise.

By underexposing you minimize/eliminate sensor heat buildup and therefore the sensor noise isn't ever developed/captured.

This is indeed true. The noise that doesn't build up in such a case, is exactly the kind of noise you'd remove with dark frame subtraction. It's the kind of noise that's the same between frames. There is still the random noise, though, and that can be reduced by stacking frames.

In the end, since each image is added at 100%, you end up with 400% for a 4 image stack. (that's why each is exposed at only 25% of the desired exposure for a 4 image stack)

Here you see that it's actually averaging. Instead of using 100% of one image you use 25% of four images.

Conversely, "image averaging" uses multiple exposures all properly exposed and all have developed random sensor noise.

That's a technique you or anyone else has labelled "image averaging". That doesn't mean that this is the only application for the mathematical "averaging".

Obviously I could be completely off in my understanding of the differences and that's why I'm asking. But if I'm not..why not stack one image 4 times?

In four images the random noise is going to be different. Adding this noise to each other is going to look different than multiplying that same noise by 4.

[inkista]

Originally Posted by sk66

What I don't understand is, if each image is going to capture the same information (same scene all exposed -2EV), why not just combine one underexposed image 4 times? It would certainly eliminate the potential layers alignment issue. Has anyone done this? Can you explain what the difference might be?

The difference is that you don't capture the same information each time. Remember, the iso noise (not the hot-pixel noise that dark frame removal deals with) will be random between all the different shots. This is why the more frames you have in the stack, the more effective this technique is. Chances are, that the noise pixel will only have a different value in one of the four shots, not all four on the same pixel. You average/sum/multiply those values together, and the noise will average out. You'll get one-fourth of the noise with the same exposure. You use 100 images, you get 100th the noise.

If you use the same image all four times, you'll just get the full amount of noise.

With stacking, they usually recommend at least 10 images, if not 100 or 1000. The stacking technique is often used with video camera frames.

[sk66]

OK, I get it.
But this part I disagree with.

Originally Posted by sybren

Here you see that it's actually averaging. Instead of using 100% of one image you use 25% of four images.

100% of 25% 4 times =100% which is summing. (4x25=100)

25% of 100% 4 times also = 100% which is averaging. ((100x4)/4 =400/4 =100)

And I think there is a significant potential difference depending on "how" you take the images for averaging.

It seems that the best possible results would be achieved with summing low ISO images underexposed AND subtracting a dark frame (I saw a particularly cool action for this as it gives better results than either in camera or simple layer subtraction).

I suppose this might be very similar in result if you used a low ISO for your averaging images. But then the exposure times would be even more painfully long and that can introduce additional issues. (tracking requirements, or unwanted trails) Probably why they suggest way more than 4 images.


I don't think I was far off at all, just thinking in less extremes.
Say I have a simple night scene which would require either a high ISO or a long exposure time (but nothing near astrophotography times) lets say 30 sec at ISO 3200. 30 seconds is not long enough (generally) to cause sensor noise, but ISO 3200 certainly can cause high ISO noise in many cameras.
If I take 4 images at ISO 400, 30 sec. and sum them I will get a low ISO noise and low sensor noise final image.
And even if I only take one image at ISO 400, 30 sec. and layer it 4 times I will still wind up with less total noise than a single image taken at ISO 3200. But it will probably be somewhat worse than the 4 separate images.

Now I'm thinking of how application of either averaging or summing might allow you to get a more typical image otherwise not possible.
Like a decent handheld shot where you otherwise couldn't (high ISO and averaging). Or get the absolute best image quality where you otherwise couldn't. (low ISO and summing)

[sybren]

By taking a 25% exposed shot, you have 25% of the signal. By dividing a properly exposed shot by 4, you also get 25% of the signal. Both ways can be called averaging, and indeed based on what you do the result will be different. The fact that the results are different are no reason to call one "averaging" and the other not. You're thinking about averaging 100% exposed images, I'm thinking about averaging signals.

And even if I only take one image at ISO 400, 30 sec. andlayer it 4 times I will still wind up with less total noise than a single image taken at ISO 3200.

It will still be under-exposed by a full stop (4x400 = 1600), unless I don't understand what you mean with "layer it 4 times".

If one could simply apply some algorithm to a ISO 400 image, and end up with the same image but with less noise, there would be no need for sensors that could do ISO 3200. Can you give an example image at ISO 3200 and one that you took at ISO 400 with the same shutter speed and multiplied?

[sk66]

Originally Posted by sybren

By taking a 25% exposed shot, you have 25% of the signal. By dividing a properly exposed shot by 4, you also get 25% of the signal. Both ways can be called averaging, and indeed based on what you do the result will be different. The fact that the results are different are no reason to call one "averaging" and the other not. You're thinking about averaging 100% exposed images, I'm thinking about averaging signals.

Interesting point. I hadn't thought of averaging the signals as any different...maybe it is.

Originally Posted by sybren

It will still be under-exposed by a full stop (4x400 = 1600), unless I don't understand what you mean with "layer it 4 times".

actually it's 400/800/1600/3200

Originally Posted by sybren

If one could simply apply some algorithm to a ISO 400 image, and end up with the same image but with less noise, there would be no need for sensors that could do ISO 3200. Can you give an example image at ISO 3200 and one that you took at ISO 400 with the same shutter speed and multiplied?

I don't have examples, but I'm planning on trying it....maybe today. I'll probably use ISO 800/6400 to get more noise as I'm using a D3.