Why the new DSLR have so little noise with high iso values?

[Lauri Kettunen]

Is somebody able to explain why the images created by the new Nikon and Canon DSLR bodies have amazingly low amount of noise with high ISO values. What's the explanation in precise terms?

[David Rasberry]

Is somebody able to explain why the images created by the new Nikon and Canon DSLR bodies have amazingly low amount of noise with high ISO values. What's the explanation in precise terms?

Big reductions in noise in the processor ASICs. Should see the same improvements in the next gen RED's too.

[Lauri Kettunen]

Big reductions in noise in the processor ASICs.

Sure, that's obvious, but which way the noise is reduced?

[Kenneth Elkington]

Sure, that's obvious, but which way the noise is reduced?

I'm sure it's thoroughly proprietary. If everybody knew how they were doing it i'm sure a kodak easyshare would have some pretty good high iso shots too.

[Tim Whitcomb]

Big reductions in noise in the processor ASICs. Should see the same improvements in the next gen RED's too.

I know little about this stuff but is
interesting, Jim said in his last post regarding state of economy that they were experiencing some ASIC challenges...

[victors]

The new 5D takes great stills with high ISO, but the low light video performance still has a ways to go.

[Josh Negrin]

The new 5D takes great stills with high ISO, but the low light video performance still has a ways to go.

I disagree. Having shot a movie with the RED and a fun video with the 5d, the 5D's low light performance far outperformed the RED's. We shot the RED with Nikon super speeds (F1.4 I believe) and the 5D had a f4 zoom lens. Even still, 5d was less noisy in low light and looked better.

[OptiTek]

Sure, that's obvious, but which way the noise is reduced?
Although I don't know the specifics(obviously) Canon's newest DSP-DIGIC IV
as the name implies is the fourth generation of their signal processing chip and offers major improvements in signal to noise ratio. It's the signal to noise ratio optimized for low light that results in this outstanding performance.
Canon, as a company, has probably the most DSP experience and being so big employs an army of engineers constantly improving their products.
Having four generations of DSP experience and hundreds if not thousands of engineers on payroll really makes the difference.... and it shows.
I can only imagine what's sitting on their executive desks now waiting to be released...
I went to CES this year and visited Canon, Nikon, Sony, Panasonic and Casio booths just to see what they will offer in near future. Talked to their sales reps and, however they all were pretty tight lipped found out some pretty exciting news coming this year. Next time I'll take videos of it- some converations were pretty hillarious, some very educating..
My two pennies,
Jacek Zakowicz, OptiTek.org

[ericyoung]

RED's electronics also run hotter than DSLRs, so thermal noise is going to be bigger.

[GlennChan]

I'm sure it's thoroughly proprietary. If everybody knew how they were doing it i'm sure a kodak easyshare would have some pretty good high iso shots too.
My notes:

1- The noise reduction performed by a lot of consumer cameras is... not very good IMO. Compare the results against shooting RAW and using NR plugins like Noise Ninja, Noiseware, neat image, etc. Big difference.

2- The sensors in the newer cameras seem to have (very) low noise to begin with.

3-

Sure, that's obvious, but which way the noise is reduced?
I don't believe anybody divulges their algorithm, except for the academic community (but none of them seem to be doing very high quality NR; maybe except for lepennec's bandlet stuff).

Looking at commercial plugins out there, the underlying algorithms seem to be fundamentally different. You can make an image with a box in it, add noise, and then let NR plugins loose on it.
Noise Ninja will produce vertical lines. This seems to suggest that they are doing NR horizontally (1-D), then NR vertically. *I don't know what their exact algorithm is.
Noiseware reduces the amount of NR as it approaches edges. Its sharpening algorithm will also produce some artifacts unique to it.
I have my own NR plugin and it produces yet different artifacts. (Ringing on edges.)

4- A lot of RAW processors will do some amount of NR.

5- IMO, the best approach would be to shoot (compressed) RAW and do NR later on a desktop computer. This way you don't damage the image and you can get the high quality NR by doing it on a desktop computer (because they can run any algorithm, and you can update that algorithm; and speed doesn't matter).
I'm not sure how good the NR is on an ASIC, but you'd be able to do the same thing on a desktop computer. (In some cases, specialized hardware will do a better job because there is no desktop software with equivalent algorithms... e.g. format and standards conversion. But that shouldn't be the case here.)

6- NR does and doesn't work.
Statistically and visually speaking, it can improve the image for real world scenes.

But if the noise is extreme, it will obscure fine textures. If you get rid of all the noise, then you'll have a 'plastic' image without any texture. And if you leave some noise in (recommended), the image will have noise textures in place of actual texture of the object. That won't look right.

You need the sensor to have low noise in the first place.

[Stephen Effendy]

Correct me if i'm wrong but from physical pixel count :

RED One Mysterium sensor : Total 11480800pixel (11.5 megapixel)
5D M2 Sensor : 21 megapixel

My theory :
-The 5D churn out a 1080p downrezing from a full 21 megapixel sensor. Even if there is noise, the downres will decrease the size of the noise to virtually gone.
-The 5D in camera NR is much improved than the 1Ds Mark III for instance, thats why with this SLR ISO3200-6400 is still usable.
-The NR sacrifice sharpness, but since the image then down res to 1/6th of its original size (from 21K-2K), then the noise are gone and sharpness adds up.
-The H264 compression actually soften the noise a bit.

The NR artifact is visible on low-no light and dark shadows.

Our company has been using RED one (2K 120fps) and 5D M2 (30p) as second camera, for quick shoot and running gun.
Online on autodesk Flame/smoke, low latitude on the 5D but twice more flexible than HDV.

We did a tabletop reel on RED one with ISO 100 - 250, high shutter speed. even at 2K it is very clean on the shadows blue channel.
(but baked by 20.000 watts of various tungsten next the camera + 1/2 CTB)

[Lauri Kettunen]

Thank you all for your comments. Still, have not been able to make much progress in building better understanding based on the feedback.

What I'm really after is this: As far as I can see the low noise level is indeed about signal processing rather than some special properties of the sensor. If so, and if the noise reduction is made after the AD-conversion, then, in fact, isn't it that the best result were obtained if one did the noise reduction in post? I mean, if there is signal processing within the ASIC, then it has to be an automatized system. And now, it does not sound likely that an automatic system tackled all cases in an optimal way. Instead, I assume manual adjustments and finetuning gave even better results.

Next, if the question is of signal processing, then in principle we should be able to improve R3D-files shot in low light conditions precisely he same way. If we only knew how noise is reduced. This is why would like to get a grip on the basics. Thereafter, I'm sure, this community will be able to work out the details.

[ericyoung]

Better optimised noise reduction algorithms are useful, but not as useful as improving signal to noise ratio at the front end in the first place. Cooler running and less noisy sensor designs and A-D asics is where the main improvements have probably been made in DSLRs, and where there is most scope for improvement in future RED cameras.

[Daniel Browning]

Is somebody able to explain why the images created by the new Nikon and Canon DSLR bodies have amazingly low amount of noise with high ISO values. What's the explanation in precise terms?

There can be many reasons:


Unequal sensor sizes.
Unequal technology.
Unequal spatial frequencies.
Unequal processing.
Unequal expectations.

Unequal sensor sizes are one mistake made in comparison. FF35 sensors (36x24mm) will always gather more light for a given f-number than S35 (24.4x13.7mm). It's no surprise that it gathers more light, given its tremendous size. Of course, FF35 cinema lenses don't yet exist.

Unequal technology is another mistake. So many things that matter for motion images can be ignored for still cameras. FPS requirements, skew, codec, etc. can make comparisons nonsensical. For example, it may be that reducing skew and increasing FPS takes up space in the pixel, reducing the room available for the photodiode (fill factor). That would reduce sensitivity (QE), full well capacity (FWC), and increase noise relative to a sensor that didn't have to bother with 60 FPS and skew.

Unequal spatial frequencies is when you compare the noise of one system under high magnification against the noise of another system at low magnification. If the low magnification appears to have less noise, you might wrongly conclude that the final displayed image will have less noise as well, but that's wrong. The most common mistake of this nature is to compare both at "100% crop", which is to magnify each image by an amount that corresponds with its resolution, rather than at any sort of equal spatial frequency or magnification.

Unequal processing is another huge reason that has already been discussed in this thread. So many things happen after the readout on the sensor. For example, Nikon clips blacks in the raw data above the mean read noise value, sacrificing valuable data on the altar of covering up noise. Recent Nikon JPEG styles (e.g. D3) have been to desaturate the shadows heavily, so that there is much less color than a standard RED ONE raw conversion. This creates a big difference in noise visibility.

Unequal expectations are when you apply a different standard to one camera than another. For example, if you measure the noise of a standard definition DVD from an old Sony 150 camera, then measure the noise of a 4K master from RED ONE, but expect them both to be the same, you are applying unequal expectations. Equal expectations would be to measure both at the size of a DVD.

[Antti Pirskanen]

My theory :
-The 5D churn out a 1080p downrezing from a full 21 megapixel sensor. Even if there is noise, the downres will decrease the size of the noise to virtually gone.
-The 5D in camera NR is much improved than the 1Ds Mark III for instance, thats why with this SLR ISO3200-6400 is still usable.
-The NR sacrifice sharpness, but since the image then down res to 1/6th of its original size (from 21K-2K), then the noise are gone and sharpness adds up.
-The H264 compression actually soften the noise a bit.


It has been suggested that the 5d MkII simply skips every 2 out of 3 sensor lines while downsizing the image. There is very visible diagonal line aliasing on videos shot with 5d MkII.

I've been using 5d MkII with Leica-R objectives. I have a gut feeling that most of the improved high ISO performance is due to a better sensor.

My theory is that there is a maximum cap of ISO 3200 in video mode because the Digic 4 simply can't simultaniously perform both processor-intensive NR and H.264 compression at 30 fps in real time. Perhaps the NR gets increasingly more processor-intensive as the ISO goes up.

[Antti Pirskanen]

If so, and if the noise reduction is made after the AD-conversion, then, in fact, isn't it that the best result were obtained if one did the noise reduction in post?

It is my understanding that as you decrease the noise prior to recording the signal, you are also preparing the data stream for higher compressibility. With h.264 it is particularly important, but with Redcode it is not as much of an issue. It seems obvious that you could do better noise reduction in post rather than in camera. Noise varies with brightness and color and you could analyze each shot in the post to create an optimized NR profile for that particular shot.

As a side note, I have noted that the H.264 implementation of 5d MkII really breaks down while shooting temporally complex subjects like falling snow. It is, however, due to inter-frame issues rather than intra-frame issues. It is also one of the reasons that Redcode is far superior in professional applications compared to H.264 or XDCAM. That is, of course, if bandwidth and file sizes are not an issue.

[Matthew Bennett]

That is, of course, if bandwidth and file sizes are not an issue.

You just said a mouthful!

[Josh Negrin]

we've gotten some pretty technical answers here from guys who really know what they're talking about. Let's get back to the fundamentals though, just like with film cameras, the sensors got better and better every year. 35mm film became the standard and instead of increasing resolution and making that the standard (70mm etc) didn't the enthusiast get the new film slrs every year because of their upgraded sensors?

I guess what I'm saying is, this is cyclical. The cameras in two years will be better than the cameras now. It should go without saying. It's like looking at a 100gb harddrive from three years ago and questioning how they are now able to fit 1.5tb into the same amount of space, and if we can some how hack our 100gb hd's to fit the new 1.5tb.

To answer your question though, the R3D's have their own algorithms (and for what they are, they are amazing), and while in theory it seems that we should be able to some how recreate the noise reduction found in the mark II, it's just not possible. The red one is an amazing camera, just don't shoot it in low light. Get a mark II for those shots, or wait for a scarlet, because we all know that will blow away the 5d mark II. :)

[Charles Angus]

I think the answer is ten years + of intensive sensor design.

From what I've seen of the 5d MkII video, the blacks are crushed to hell and that's why there is very little noise. Crush the blacks a couple stops on RED and I'm sure you'll see a big decrease in noise.

Or run really good DNR software on the footage.

[GlennChan]

Epic is specced to have something like 2 stops less of noise. (That is a lot.)