Sunday, January 27, 2013

What Does it Mean When We Say That the D7000 Is An "ISO-less" Camera?



The concept of the D7000 being an "ISO-less" camera was raised soon after its introduction when photographers started noticing how the noise profile behaved throughout its ISO range. However, the concept is still an easy one to misunderstand, and at times can be quite counter-intuitive. Now that the camera is at the end of its retail life, it (and others using the Sony 16mp sensor) may very well be remembered for ushering the era of "ISO-lessness", as cameras that followed it such as the D800 and now the D5200 exhibit the same behaviour.

But what is the property of ISOlessness you might ask?


Exposure Theory


To step back, it helps to revisit the basics of the exposure triangle: shutter speed, aperture and ISO. These three variables control the amount of light that enters the photodiodes on the sensor. The classic example used is that that photodiodes are like buckets, light is like water coming out of a hose. Aperture is analogous to the size of the nozzle opening on the hose, and shutter speed is the equivalent of how long the water is turned on. Which leaves us ISO, which is like the pressure of the water coming out of the hose. Adjust any of the variables, nozzle size, water pressure or the amount of time the water is on and you can adjust how much water fills the bucks. Same goes for light; adjusting shutter speed, aperture and ISO controls how much light enters each individual photodiode. This is about as much as you need to know; most people don't ever develop a deeper understanding of exposure than this and are able to take terrific pictures nonetheless.

In electronic terms, ISO, the "pressure" of the light signal read off of the sensor, is really just amplification, but a key point that is overlooked on a lot of introductory tutorials is that the amplification is of the signal that is collected by the sensor, not the light itself. Many tutorials speak of amplification as if the sensor is being made more sensitive... after all, that was what the original meaning of the ISO rating was about back in the film days. ISO 200 film was more sensitive than ISO 100 film because the grains of silver halide were larger. However, this is not what is happening in a digital camera; the sensor is the same regardless of the ISO selected. The sensor is no more sensitive to light at ISO 6400 than it is at ISO 100.

Noise and Amplification


Before we talk more about amplification, we have to step back and talk about noise for a second. At it's heart, noise is nothing more than than a deviation from what the "true" value being measured is. Say your data set can be described in 8-bits, which means that there are 256 discreet values that can be assigned to the measured value. By convention, not recording a signal is also a value so the set of values can be described on a scale of 0 to 255. Say the true real-world value of what you are measuring is 100. In a situation with very low noise, repeated measurements will give you a tight grouping of values that are exactly on or very close to 100, say {100, 100, 99, 101...}. However, a noisy situation, so to speak, will produce measured values that have a wider scatter around 100, say {100, 95, 105, 110, 90...} In both cases, the average value is 100, but the variance in the noisier sample is greater. That is basically what noise is in a nutshell.

However, what the average person calls "noise" is actually a number of different processes. We'll break this down into three broad categories for simplicity (I'm lumping together a lot of different concepts here):

  • Amp Glow: This is caused by heat and interference from components adjacent to the sensor. Nikon D80 users were well versed with the dreaded purple amp glow problem on long exposures. Early models produced three distinct splotches at the bottom of the frame on exposures longer than a minute. Updated versions were improved so that "only" two splotches were present.... if you are wondering, no I did not consider this an "improvement".
  • Dark Current: I'm lumping together different ideas here, but this is the type of noise that tends to predominate with longer exposures, and which arises from within the sensor. You tend to see this predominantly as hot pixels. To be more precise, dark current isn't really noise, it's more correctly defined as an unwanted signal
  • Read Noise: This arises from the process of pulling the information collected by the sensor off the sensor chip and into the camera's processing unit. Think of the sensor data as a many-laned highway, and the process of reading the data off the chip as merging all of those lanes together to cross a bridge. There's going to be jostling involved.
  • Shot Noise: This is a fundamental property of light. Photons arrive at the sensor in a random, but predictable pattern. The shorter the exposure time is, the more variation in the signal there is because the sample size is smaller.
For the purposes of this discussion, we'll only concentrate on the latter two, as amp noise and dark current noise only become factors when shooting long exposures.

Read noise is a relatively fixed factor when considering the overall image noise. If it were not for sensor amplification (as described below), the amount of read noise would be a constant factor in every shot regardless of exposure time. Just after an exposure is achieved, the data that's on the sensor is physically packed together in side-by-side lines. Because the data is electrical in nature, it's close proximity creates interference with each other as it is read off of the sensor.

Here is an example of read noise from a D7000. This was taken with the cap on at ISO 800, 1/8000s, ambient temperature was -1 degrees Celsius. Yes, there really is a picture there...  Under these conditions, the camera should record a perfectly black image, as no light is entering the sensor and the ultra short exposure time prevents the build up of dark current noise. In case you are wondering about these settings, this was a bias frame I was using for a disappointing astrophotography outing (subject of a post for another day).

D7000 Bias frame, ISO 800, 1/8000s
However, if you run this through post processing and slide the histogram out all the way to the left to brighten it as far as possible, you get this:

The above bias frame enhanced to show read noise.

What you are seeing is the the very minute trace of read noise that gets recorded when the data from the sensor is read out. The regular synthetic pattern is a direct result of the recording process. For every space that you see a pixel that is not black, the sensor recorded a value that was almost black (0,0,0) but which was not quite (0,1,0 etc.) For all intents and purposes, in normal viewing this is perfect black. The reason why you see red, blue and white specs is because I've severely bumped up the picture to highlight the noise.

However this is with a near state-of-the-art dSLR. What does read noise look like on a lesser camera? Unfortuantely, I don't have an older dSLR with me, but I do have my trusty Panasonic LX-5. The conditions aren't going to be analogous (the LX-5 only has a maximum shutter speed of 1/4000s and I'm not heading out into another bloody cold night unless I have something to show for it) but they do highlight the the difference in technology between the two. We'll use ISO 800 again as an arbitrary starting point.

Panasonic LX-5 Bias Frame, ISO 800, 1/4000s

First, to clear any misconceptions, the difference in sensor size doesn't matter (much). Since we are only recording absolute darkness, the difference in sensor size and light well capacity shouldn't matter. This is probably one of the few tests where it's fair to compare a dSLR against a compact camera. Here, you can see that the image is mostly black but if you look closely you can see some mottling, which is the read noise that has already been amplified because I selected ISO 800. If you enhance this frame the way that I did with the Nikon image you get this:

Panasonic LX-5 Bias Frame Enhanced

Big difference huh? Turns out most of the black wasn't completely black at all, just nearly black... but black enough for normal purposes. This is a really big difference, so the question is why? I can only guess, but the reasonable explanations hinge on manufacturing and build quality. There is a third factor that we can get out of the way, which is time. Read noise has decreased over time as technology has improved, but the LX-5 was introduced just a few months ahead of the D7000. If you look at build quality, it generally seems that dSLR sensors (and particularly full frame ones) are built to a higher specification than smaller sensors. Then there's manufacturing. Ignoring many other factors, CCD sensors are generally better at light-fill and shot noise than CMOS, but CMOS because of it's on-board processing circuitry is better with read noise. There's another factor at play here, which is that Panasonic has traditionally lagged behind Sony and Canon in sensor performance, though with the LX-5, it wasn't by much in practical terms.

Shot noise is an inherent property of the light gathering process. The less light that you gather, the more shot noise that is produced. This is not all that different from polling or consumer surveys. Say that the true population of Canon to Nikon users in the world is 60% to 40%. If you do a survey, the more responses that you can collect, the more likely it is that your measured results will be close to the real 60/40 split. However, if you only survey a handful of people, the results can vary quite a bit from this. Survey more people and this confidence internal narrows. In a round about way, what I am saying is that shot noise is fundamentally a statistical process. I don't have to provide too many examples of shot noise, because you've probably seen lots of examples already. Every time you see ISO samples of a test patch where the sample is a solid colour, the noise that you see as the ISO goes up. If you've been pixel peeping the new Nikon D5200, it has this property too. From ISO 100-800, the D5200, D800 and D7000 have very low read noise.

ISO-lessness


So the issue of noise is that it's generally undesirable, though the the pattern of the noise also matters. Early digital cameras had very ugly splotchy noise patterns as you went up the ISO range, which is a function of the read and dark current noise characteristics of the technology of the time. As cameras improved, the grain of high ISO noise started getting finer and tighter; sensors improved, the read noise started dropping, leaving only the shot noise which is an inherent property of the amount of light collected per exposure. Ideally, you want to maintain the lowest ISO possible for a given situation, because any noise that is present in the sensor signal will be amplified when it is read off the chip and into the image processor.

The question of ISOlessness is whether or not amplification is better when it is done in camera or whether it is done on a computer. This is exactly like the situation with the RAW demosaicing  procedure: you can do the RAW conversion on board the camera or you can do it post processing with your favourite software. However, the practice is a little bit reversed when it comes to noise control. When we are talking abut converting the RAW data to a usable image, it is better to do it on a computer because the computer either has more processing power or can use a more sophisticated conversion process. The computer also has an added advantage of being able to receive conversion updates.

However, if you are looking at it strictly in exposure control terms, it is usually better to control the signal amplification while on-board the camera. This is because the camera can do one thing that a computer cannot: analogue signal amplification. Analogue amplification is the equivalent of using vacuum tubes for audiophile equipment. When a signal is amplified with analogue means, the output has smooth curves to its profile. In audio terms, we would say that it has "warmth". Digital amplification can "damage data" by leaving it with spikes and dips. In fact, whenever you play with the levels/curves or brightness/contrast adjustments on a post processing program, you are essentially amplifying the camera's signal, it's just that you are doing it on your computer. If you over do, the histogram, which may have started off with a smooth looking curve, ends up looking spiky with gaps and peaks. The resulting image will show banding and colour blocking.

How easily your image data can be damaged by post processing also depends on how noisy it is. Here is where we get back to the concept of read noise. Read noise is essentially the floor of the signal. Raising the ISO or brightening the image in post processing does the same thing; both will amplify the usable signal as well as the noise. The difference is that the camera will amplify the read noise smoothly because the process of analogue amplification smooths out the signal.  The post processing program will amplify it "as is", so that the more you brighten the image, the more the banding becomes apparent.

With the current generation of sensors, a side benefit of the ultra low read noise is that you can bend the signal digitally on your computer to a greater degree than ever before. By lowering the read noise, you increase the leeway with which you can digitally amplify the signal before the amplification of the read noise becomes objectionable. In practice, cameras use a mixture of on-board digital and analogue amplification. The Sony sensor in the Nikon D7000 uses analogue amplification until about ISO 800 and amplifies the signal digitally after that. For a very detailed read, go to this thread. I highly recommend that you do, as the detail that Marianne Oelund has produced is stellar, and is not the everyday sort of information that you will find in your cameras manual:

"For the user, the implication of this is that ISO settings above 800 provide no more true image data than ISO 800 itself does; you only lose highlight range in your raw files when you select those settings. Although those high settings have practical value in producing usable JPEG images out of camera, for Raw-only shooting, we are better off staying at ISO 800 and using scaling in post, where we can use tone curve customization to preserve highlights that would have otherwise been clipped."

The implications of this took people a few moments to grasp two years ago and is still counter-intuitive today. There is almost no difference between applying a gain in ISO on board the camera, or brightening the image afterwards in post processing. For all intents and purposes, the two were now essentially the same.  And as Ms. Oelund demonstrated, this is doubly true above ISO 800 on the D7000. Hence the term "ISOless"... the ISO value on a modern camera is no longer a physical property of the image, but it is akin to meta data that is tagged onto the image and which can be altered with impunity. For examples I refer you to this thread.

Practical Considerations


However, this does not mean that you don't have to worry about achieving correct exposure anymore. ISOlessness is not a free ride, and whether or not you amplify the signal in-camera or in post-processing, the dynamic range will decrease in both cases as you amplify. Even if read noise is now a lesser part of the equation, shot noise is still the predominant form of noise that you see in a picture. You still want to minimize the amount of shot noise that is present in the image, and that means maximizing the amount of light that you can capture, because the more light that you can capture, the stronger the signal is over the noise.

Few people actually put the concept of ISOlessness in practice. When you are shooting, there are many other factors to take into consideration, so it's easier to think of exposure in the classic exposure triangle terms. However, an understanding of "ISOlessness"can lead to better picture taking, as you are no longer mindful of the ISO setting but rather, you are thinking of of the total amount of photons collected. A classic example of this is night shooting. Many beginning shooters automatically equate "dark" with "needs more ISO". Say that you need to take a shot at ISO 3200 and 1/32s if you are hand holding the camera. The equivalent exposure at ISO 100 would require 1s to achieve the same exposure. If you have a tripod, ISO 100 and 1 second will produce virtually clean (ignoring hot pixels and dark current noise) where as the ISO 3200 at 1/32s will be grainy. Both are the equivalent exposure, but they are not physically the equivalent process. Exposing for 1 second over 1/32 of second means that the sensor has captured more light. More light means more signal and less shot noise.

Therein lies the heart of understanding: It is not the high ISO that produces the noisier image, but it is the necessity of using high ISO that is the root cause of noise. By shifting your thinking from ISO-terms to light gathering terms, you can be more mindful of creating quality pictures.

7 comments:

  1. Brillant explanation ! I have read many pages on that topics on dpreview and I must say that your explanation is astoundingly clear and right to the point. Very much appreciated !

    ReplyDelete
    Replies
    1. Thank you; glad that you found it helpful.

      Delete
  2. It's rare to find any good articles without brandrelated thinking or myths about new sensors. This article is one of the few, worth reading. Thank you.

    ReplyDelete
  3. What about other sensors (d7100/d7200) are they iso-less to a degree too?

    ReplyDelete
    Replies
    1. It's case by case. The D7100.. not really, it will show banding/ pattern noise in the very deep deep shadows, and if you start pushing your exposure past 3EV or so. The D7200 is a bit better, but it's not as clean. The D750, sure absolutely. The 5DMark IIII? definitely not. ISOlessness is a nice fall-back if you need to do heavy editing, but you will still get the best results nailing the exposure at the time of shooting, regardless of camera.

      Delete