Tonal Range in Digital Cameras

the latest leaf camera back has 12 stops. in a recent test i read, can find it now though, the reviewer demonstrated that the d3 had 10 stops when shot in raw, 7-8 in jpg.

My question to the gurus who understand sensor technology is whether there are inherent benefits of larger sensors in these areas, particularly highlight areas?

In my quest to obtain higher and higher dynamic range images, in camera, is FX or DX likely to make a difference, or is it other design aspects of the sensor, camera firmware (EXSPEED, D-Lighting) etc that are more important?

I believe two things are important in answering those questions. First, all sensors work the same way. It's possible to imagine a sensor that is inherently nonlinear but we haven't seen one yet. Fuji's design is interesting but it's still a linear sensor; there's just two of them at the same time. Second, there's a technical limit to how much dynamic range modern DSLRs can deliver and it seems to be around 12 stops. Many of us thought that the transition to 14 bits would fix that but so far that hasn't been the case. It's possible that 12 stops can be exceeded but we're not seeing it in 35mm DSLRs because of other constraints, most likely minimum speed requirements. Regardless of the technical reasons, it appears that there's a ceiling on DNR so the goal becomes making sure we bump into it.

With those two technical limitations on the current state of the art, we can say that we need a photosite large enough to hit the performance ceiling at the ISO speed we need and still give us the exposure options to capture sunballs correctly. FX versus DX becomes a resolution argument at that point. Until manufacturers overcome the 12 stop limit, larger photosites and bigger sensors aren't going to help. I would happily accept a 1 fps limit on a hi-res body to get more DNR. That's what MF cameras do.

Finally, Eric's sunball RAW conversions demonstrate that raw converters have an effect on the quality of output we get from those types of shots. Ideally, we should overexpose nothing and capture the full range of the scene but that's nowhere near realistic. Until the time when we can, raw converters make a difference.

It's an interesting thread but I'm a bit confused. The common agreement seems to be that size matters and bigger pixels result in better dynamic range.

However, I've just checked the new review of a900 on dpreview and this camera seems to have a better DR than the a700 that have less pixels (and obviously having bigger photosites):

http://www.dpreview.com/reviews/sonydslra900/page24.asp

What's your opinion?

FX versus DX becomes a resolution argument at that point.

And/or a lens argument :)

It's an interesting thread but I'm a bit confused. The common agreement seems to be that size matters and bigger pixels result in better dynamic range.

However, I've just checked the new review of a900 on dpreview and this camera seems to have a better DR than the a700 that have less pixels (and obviously having bigger photosites):

http://www.dpreview.com/reviews/sonydslra900/page24.asp

 

What's your opinion?

The a700 has a cropped sensor, whereas the a900 is full frame, so in fact the pixels on the a700 are not necessarily larger or as much larger as you would expect. I'm not sure what the actual pixel sizes are on each camera, and can't find a reference to this in Sony's literature.

Obviously there may be other factors such as advances in manufacturing processes from the a700 to a900 that can affect pixel size for a given pixel number and sensor area, such as decreasing the space between pixels allowing more space on the sensor for either more or larger pixels - this is what Canon has apparently done with their new sensors as promoted for the 50D and there are other factors that can be improved like the A/D technology which is a feature of the new a900 sensor.

The basic rule still applies though when comparing sensors that are identical in every other aspect besides pixel size.

It's an interesting thread but I'm a bit confused. The common agreement seems to be that size matters and bigger pixels result in better dynamic range.

However, I've just checked the new review of a900 on dpreview and this camera seems to have a better DR than the a700 that have less pixels (and obviously having bigger photosites):

http://www.dpreview.com/reviews/sonydslra900/page24.asp

 

What's your opinion?

The pixels on an A700 are smaller than an A900.

Dpreview does most of its dynamic range testing using JPGs and that doesn't tell you much about a camera's theoretical capabilities. Regarding their A900 raw test, their result of 12.6 stops is theoretically impossible to achieve with a 12 bit camera. You would think that such a result would tip the reviewers off that something is wrong with their testing methodology. Don't trust dpreview's dynamic range testing.

The a700 has a cropped sensor, whereas the a900 is full frame, so in fact the pixels on the a700 are not necessarily larger or as much larger as you would expect. I'm not sure what the actual pixel sizes are on each camera, and can't find a reference to this in Sony's literature.

Obviously there may be other factors such as advances in manufacturing processes from the a700 to a900 that can affect pixel size for a given pixel number and sensor area, such as decreasing the space between pixels allowing more space on the sensor for either more or larger pixels - this is what Canon has apparently done with their new sensors as promoted for the 50D and there are other factors that can be improved like the A/D technology which is a feature of the new a900 sensor.

The basic rule still applies though when comparing sensors that are identical in every other aspect besides pixel size.

Sorry for my mistyping. I meant nikon d700 and not the a700. I hope my previous post makes sense in this way.

Don't trust dpreview's dynamic range testing.

You suggested not to trust diwa-labs either.

Looking at the graphs they display for DR, it is hard to exactly draw the same conclusions that they do. The D700 also has the identical sensor to the D3 yet they show it to have less DR on their step wedge interpretation. I think the points made by Craig probably account for all of this and JPEG comparison of DR is probably invalid.

You suggested not to trust diwa-labs either.

Yes I did. For good reason, too. My comments there were valid. They are here, too.

Just because someone puts up data and pretty pictures doesn't mean they are right particularly in the internet age. In diwa labs's case, and in dpreview's in this instance, it's clear they have methodology problems. The first thing they teach incoming freshmen in Engineering 101 is how to do basic sanity checks. Both of these testers have failed at that.

The fact is that 12 bit sensors cannot deliver 12.6 stops of dynamic range. I realize it. The dpreview tester apparently does not. Choose to believe who you want.

Just because someone puts up data and pretty pictures doesn't mean they are right particularly in the internet age.

Spot on Craig. Unfortunately the internet is as good a place to spread misinformation as solid fact.

Having just been dealing with a large number of image files from differing cameras and different photographers but of similar subject matter, I can quite happily state that I can see differences (IMHO) in many aspects of the images. Whether these differences could be narrowed by adjustment of settings and careful post processing is another matter but to me there ARE differences and this to me is the reality of using different systems. Theoretical DR is one thing but actual may well be quite another.

Yes I did. For good reason, too. My comments there were valid. They are here, too.

 

Just because someone puts up data and pretty pictures doesn't mean they are right particularly in the internet age. In diwa labs's case, and in dpreview's in this instance, it's clear they have methodology problems. The first thing they teach incoming freshmen in Engineering 101 is how to do basic sanity checks. Both of these testers have failed at that.

 

The fact is that 12 bit sensors cannot deliver 12.6 stops of dynamic range. I realize it. The dpreview tester apparently does not. Choose to believe who you want.

Neither dpreview nor diwa. My only and last hope is Ken Rockwell :-)

Joke aside, I'm critical with these reviews but also with other statments written here. In the internet age you can find anything to prove that you are right. In this test, the nikon d700 is slightly better than the a900:

http://www.imaging-resource.com/PRODS/AA900/AA900IMATEST.HTM

Can I trust this? But it`s still not a huge difference. There are also lots of discussion on the forum of dpreview on this topic.

Why do you think, that 12 bit sensors has a 12 stop DR ceiling? Can you explain me!

I didn't see anything wrong with the Imaging Resource review. He doesn't know that 1/3 stop step charts don't produce 4 significant digits though. Once you realize that, seeing 12.1 stops out of a 12 bit camera isn't concerning, it's just an error of precision (and artifacts of software). The reviewer also expressed disbelief that the S3 could rate lower than other cameras. That suggests that he hasn't completely thought through what he is testing.

These days a "fact" can be established simply by stating it and having it recognized by someone else. We've even seen such tactics used as justification for a nation to go to war. I'm not equating the two, but it's very easy to publish something pretty, get a bunch of parasitic click-thru bloggers to link to it, then the next thing you know it's presented as fact on a Wikipedia page. I see wrong stuff posted on the internet every day now as bloggers don't feel any obligation to pass a sniff test. History is written and rewritten with unprecedented ease these days. Information is easy now so the burden is on us to be sharp.

I'll attempt to describe the 12 bit, 12 stop relationship:

Every stop represents a doubling or halving a light intensity. In the binary number system each digit (bit) represents a doubling or halving of the value of the digit next to it. The sequence 1,2,4,8 represents a doubling of the previous number and in binary that sequence is represented as 1,10,100,1000. As you can see, there's a direct correlation between photographic stops and bits in a digital value.

Let's assume that a digital camera has a 4 bit converter and that exposure is set perfectly to fully saturate highlights. In that case, a highlight pixel will have a value of 1111. Now let's stop down the exposure by one full stop. The resulting digital value will be half, or 0111. Doing it again results in 0011, a 3rd time results in 0001 and a 4th results in a totally black pixel. As you can see, the dynamic range of the camera is 4 stops and that's directly a result of the 4 bit converter. No linear digital system can have dynamic range greater than the number of bits in it's converter. It can have less.

In reality, digital cameras aren't quite perfect so a 12 bit system will have a bit less than 12 stops. If you read Clark he states that 11.6-11.8 is the realistic limit for 12 bits (figure 4 of that article). You would think that 14 bits would up the limit by two stops, and it does, but today's 14 bit converters don't have good enough noise performance at the speeds they run to realize the full gain in dynamic range.

There is also a difference in engineering dynamic range and photographic dynamic range and that confuses matters greatly. Since there is no standard for photographic dynamic range it's hard to compare in those terms, but photographic dynamic range is generally DNR - 3-5 stops. Photographers like to see shading in the lowest stop but they can't agree on how much.

BTW, I have great respect for the work required to perform meaningful testing of today's camera equipment and I believe that sites like dpreview do good work. It's still necessary to understand what they are testing, how to interpret it, and occasionally identify errors. Usually when you see something huge, like the diwa labs result, what is interesting is why the abnormal result occurred, not what the result suggests.

I'll attempt to describe the 12 bit, 12 stop relationship:

Doh, I think I've got it.

Craig,

Thanks for the very lucid explanation.

Craig,

Many thanks for the explanation!

Marcell

I was under the impression that the bit depth did not actually correlate with the breadth of the dynamic range, but rather the number of gradations within the dynamic range.

In other words one could still have a very broad dynamic range with any given bit depth - i.e. deep black to pure white, the difference with higher bit depth would be the finer density of steps in between and thus yield a smoother ramp from black to white.

I was under the impression that the bit depth did not actually correlate with the breadth of the dynamic range, but rather the number of gradations within the dynamic range.

In other words one could still have a very broad dynamic range with any given bit depth - i.e. deep black to pure white, the difference with higher bit depth would be the finer density of steps in between and thus yield a smoother ramp from black to white.

Yes, that's a popular claim but it's not true. The number of "gradations" is, itself, dynamic range. If you start with a saturated sensor, then keep halving the exposure until you end up with either pure black or pure noise, you will have measured the dynamic range and adding bits to that will either increase range or increase noise, it will not increase "gradations" because there is no way add bits to a fixed amount of dynamic range. One bit equals one stop forever with linear sensors. Remember, a stop is a doubling/halving of light and a bit is a doubling/halving of value. That correlation cannot be changed; it's a mathematical truth not an engineering design decision. There is no possible flexibility in how that works.

In photographic terms there is a concept of zones. There are a number of zones separated by a single stop but the darkest zone has gradations. If we add bits we are adding gradations to the darkest zone but not necessarily adding zones (since the number of zones is an arbitrary choice) so in that sense you can add bits to improve gradations. The problem with that is there's a variable amount of dynamic range in the darkest zone, it's simply a failure to adequately define dynamic range quantitatively. Zones are related to dynamic range but they aren't the same.

Keep in mind that all this relates to the scene captured by the lens and recorded by the camera, the so-called "scene referred data". It not about the output TIFF or JPG which is (usually) not linearly encoded and is tone mapped. These files are known as "output referred data" and can have any amount of scene-referred dynamic range in them regardless of how the data is encoded. That's another common source of confusion. Let's say you have a JPG that's half white and half black. How much dynamic range does it have? The answer is that the question makes no sense.

Let's assume that a digital camera has a 4 bit converter and that exposure is set perfectly to fully saturate highlights. In that case, a highlight pixel will have a value of 1111. Now let's stop down the exposure by one full stop. The resulting digital value will be half, or 0111. Doing it again results in 0011, a 3rd time results in 0001 and a 4th results in a totally black pixel. As you can see, the dynamic range of the camera is 4 stops and that's directly a result of the 4 bit converter. No linear digital system can have dynamic range greater than the number of bits in it's converter. It can have less.

I am little bit confused. If you have 4 bits, than you have

0000 = 0 pure black

0001 = 1 |

0010 = 2 |

0100 = 4 |

1000 = 8 v

1111 = 15 pure white

Isn't that almost 5 stops from pure black to pure white?

regards

rachid

OK; now I'm confused again. To use the photon in a bucket analogy again. Let's say a large D3 pixel (bucket) vs a smaller Olympus pixel. Take a photograph of a very high contrast scene at a set exposure. Some of the pixels will be full, some in the middle and some almost empty.

The larger pixel will collect more photons from the deep shadows before noise is evident than the smaller pixel, and the larger pixel will be able to collect more photons from the very bright highlights before it overflows and blows so to speak. This will be the case irrespective of bit depth at the moment of exposure. The difference that bit depth will make is the increments within the bucket that can be differentiated. It would seem to me therefore that bit depth is more important to facilitating the larger pixels recording more gradations or shades of grey (each representing a zone or EV value or stop), than it is to the extremes of dynamic range that it can record. In other words a large pixel sensor with low bit processing would still be able to record deep shadows with low noise, and extreme highlights without blowing, but might exhibit more contrast and or banding let's just say with a 4-bit processor vs a 14-bit. And of course the speed of processing / frame rate is improved with higher bit depth.

Or, to continue with this analogy will the 4-bit processor simply position itself toward the mid greys (partially filled pixels) recorded by the sensor and simply ignore the very bright and very dark areas' and then as you add bits more and more shades of grey at either end are added.

I am little bit confused. If you have 4 bits, than you have

 

0000 = 0 pure black

0001 = 1 |

0010 = 2 |

0100 = 4 |

1000 = 8 v

1111 = 15 pure white

 

Isn't that almost 5 stops from pure black to pure white?

 

regards

rachid

To answer that, does the 1 bit case have 1 stop of range or almost 2?