A note on CMOS and working at high ISO values.
The dynamics of the capabilities of sensors built using CMOS technology at high ISO values (threshold values are shown at which a particular sensor is still capable of producing high image quality). The whole point of the note in one slide. Enlarge Image... Full frame cameras are shown from left to right from old to new. The vertical scale indicates the ISO limit that meets the quality criteria DXO mark.
I took the data from DXO mark (they have a fairly extensive base of sensors). Not everyone likes their test data, but they still give some insight into the parameters of a particular sensor. How they do their tests is described here... Briefly about the capabilities of the sensor at high ISO values, then the signal-to-noise ratio of 30 dB is taken as a basis, which implies good image quality. The histogram shows the maximum ISO of a particular camera, which allows it to hold a signal-to-noise ratio of 30 dB and a dynamic range of 9 EV and a color depth of 18 bits. It's good that the methodology is the same for all cameras.
- 13 Years of Evolution of Full Frame Sensors Since Launch Nikon D3 in 2007 and up to the Nikon Z7 II in 2020, only increased the ISO threshold by a factor of 1.24. Now just think about how, for example, smartphones have changed during this time.
- any dramatic increase in performance at high ISO did not happen. For example, by comparison, the same Nikon company had a significant increase in high ISO capabilities when moving from CCD and LBCAST technology to CMOS in their Nikon DX series cropped cameras (for example, Nikon D100 -> D200 -> D300/D300s -> Nikon D500).
- I took full-frame Nikon cameras just as an example, but this also applies to many other manufacturers
The dynamics for the Nikon DX are more joyful. Cameras are shown from left to right from old to new
You cannot jump above your head and you cannot squeeze something much more out of the CMOS. BSI and Stacked CMOS technologies also did not affect the work at high ISO in any way. Of course, there is a difference in the subtleties, but in general CMOS has been marking time practically in place for decades.
What does the future hold for us? And do not forget to write about your attitude towards DXOmark in the comments. Peace. Work. CMOS!
Comments on this post do not require registration. Anyone can leave a comment. Many different photographic equipment can be found on AliExpress.
Material prepared Arkady Shapoval. Training/Consultations | Youtube | Facebook | Instagram | Twitter | Telegram
But COLOR!
So manufacturers are constantly increasing the number of pixels, even 6 years ago 36 pixels were considered a lot for FF, now 45 megapixels are the standard for flagships, so light sensitivity is marking time. If someone with current technologies released FF with 16 MP, there would be working ISO times or even three times higher!
I think you are mistaken when trying to draw a direct relationship between the number of megapixels on the sensor and its light sensitivity.
There is dependence. The smaller the photosensitive element, the more difficult it is to make it with a good signal-to-noise ratio.
So the graphics are lacking in comparisons of cameras with the same pixel density
You cannot tell from the diagram that this dependence exists.
You can see, for example, DRL parrots for d5, 1dx mark2 (which have 20MP each) and forty-five megapixel models. Do you see a difference of 2-3 times in terms of indicators? I do not see.
The size of the pixel affects the averaging of the information received by it from single photons over the area of their capture. Averaging is possible over the area and over time, over time - exposure, over the area - pixel. According to the Liouville - Arnold theorem, these averages are equal :) so the larger the pixel, the better the light sensitivity. On patricks of increased density, it would be more correct to compare the photosensitivity after converting the images to the same resolution, and here, most likely, multi-pixel cameras in post-processing with comparable parameters, the sensitivity will give a significant increase in the quality of the picture, because the processing of a group of pixels can be with more sophisticated mathematics than simple complication in area. A correct algorithm, more complex than a simple sum, will simply be able to use the excess of information in a group of pixels to restore a better picture.
That's for sure. Those who are worried about the lack of large-pixel cameras simply forget about this moment.
After all, if you process a noisy 45MP image with the correct noise reduction without fanaticism, and then reduce the size to 16MP, the final picture will outperform the base one with 16MP both in detail and noise.
I raised this question here https://radojuva.com/2019/07/how-to-make-plumbus/ no one gave clear manipulations on how this can be done.
I repeat the questions:
The task is very simple: to reduce the amount of noise by using downsize (i.e. reducing the size of the image).
Related issues:
Is it possible to reduce a photo by 100 megapixels to 10 megapixels (10 times) and still get a 10 times cleaner, noise-free picture?
Is it possible to leave only useful pixels-information in the picture when downsizing?
Is there specialized software for such manipulations?
How to make noise reduction during downsizing while maintaining an acceptable amount of detail using classic tools in Adobe Photoshop and other common editors?
Is it possible to shoot at a high ISO, halve the image and get a noise level corresponding to ISO / 2?
I think this picture will tell a lot clearly.
The freshest large-pixel sony bzk versus the freshest and very small-pixel sony bzk.
The noise, as we can see, is comparable (this is without noise reduction), but the large-pixel one has lost the details, and lost them irrevocably, while on the a7r, the details are all in place. By suppressing small noise, we will get a clean picture on it, and on the A7s we will get a blurry mess.
Yes, the flip side of the coin is huge raves, a lot of processor time and disk space, but how else?
Of course, I exaggerated about the “mess” and “clean”, but the fact that the final picture on the screen will come out more detailed and cleaner is for sure.
I know many bloggers who shoot videos on her and she is so good that the video at night turns out much better than from other cameras. I got the impression that Esca was conceived not only and not so much for an ordinary photo, where sharp details are in high esteem.
Victor is absolutely right, there is more information on the right image, and if both pictures are reduced to a common size (denominator), then on the right inscription they will still be more readable. Those. there is a sense in downsizing even though the conditional sharpness efficiency of the D700 according to DRL is 75% (9 mp out of 12 mp), and for the D3200 50% (12mp out of 24mp), objectively the D3200 will pull out more details, although by enlarging the picture, many can see it with the naked eye, that there is a lot of "water" (empty information) in it, while many "gasp" from the camera jeep with an average image size (having reduced it in the settings from 24 to 13 megapixels).
I read this topic, did not write, because I can not think of a clear test for assessing the level of noise. In terms of perception, the picture becomes much better (up to almost perfect) after successive noise reduction (preferably first to the native ones in RAW, then to the taste of blur / contour sharpening along the channels in Adobe, for example) and subsequent reduction in Adobe with a reasonable algorithm. I regularly do this at 850 when I shoot skaters :) It doesn't even eat much time, since you can consistently automate these actions. Separately, I dabbled with the frivolous raw editor on the network, it's good that you can have fun with the algorithms of debayerization and anti-aliasing right at the subpixel level ... but it's more like a hobby :) not for photographers. the result is very comparable to the first, except that there are more semitones left. But the fact is that all anti-aliasing algorithms somehow leave rudiments in the picture, and if you approach the question strictly, then what the eye does not see mathematically can be pulled back (the pixel, as it were, is smeared with waves in the picture, in the picture with a lot of information there are heavily smeared waves are not noticeable to the eye, but if you take a black or white ... monotonous picture, you can see that information about the noise remains. Therefore, I did not write in the subject. I have no confidence and I did not directly check on the numbers that the signal-to-noise ratio with my manipulations are getting better. About A7S-Sony only starting with AIII began to use the entire matrix to create a video picture, in the previous ones it processed various crops. I again tried to shoot video in 4K on AIII from a full frame and then in the editor to process and compress to HD. I must say that the quality is also higher :) but hemorrhoids :) many times more. More questions arise here about autofocus, to be honest, it is very bad when it follows the skaters :) strives to catch on to the contrasting background (board or spectators)
Rather, pixel size affects the ability to control pixel characteristics. The smaller the pixel, the more the spread will be, which only increases with increasing gain (iso). In fact, there and in addition to the characteristics of the pixel itself, a lot of things affect, from the purity of the power supply and the noise of the transistors in the matrix to the noise of operational amplifiers and ADCs, and in addition, different color channels also have different gains, which naturally leads to an increase in noise with increasing gain. And the photosensitivity has been decent for a long time, some cameras have a matrix controlled ND filter with its help, for example, on some cameras they expand iso values below 100, of course we are not talking about cameras like d850 and z7 with native iso64, although there is such an ND-schnick there too , but it is used a little differently.
The point is this if you believe the pixel circuits from the patents. Pixel exposure is a controlled discharge of a capacitor, this process is affected by the capacitance of the capacitor itself, the transition resistance of the photosensitive element and the transistor, which additionally controls the discharge rate (shutter speed) and this is where most of the problems lie, firstly all photosensitive elements have +/- different characteristics, capacity capacitors are also far from ideal and also have a scatter of parameters, like the characteristics of transistors responsible for "exposure", this all leads to the fact that even in the case of ideal light, each subpixel will be exposed +/- in different ways, and then the iso was wound, that is, the signal received an even greater difference which we see as colored noise.
And there is also a stupid limitation on the number of photons in low light conditions. It's like a 3D rendering with ray tracing, if there are few photons, the image looks like a mosaic at first, then it's just noisy, and only then it becomes realistic. If two adjacent pixels are illuminated by 10000 and 10100 photons, the difference may not be noticeable. And if there are only 100 or 200 of them, this is a twofold difference in brightness. This is the main cause of noise in low light, and is fundamentally unsolvable.
Of course, the “intensity” of light also strongly influences, but in the case of “weak” light, we just additionally run into the noise bar of electronics and this is fundamentally not solved and the nonlinearity is likely to grow.
in the sense of who else ?? "Aunt" releases her a7c3 with 12 MP)
Forgot about the A7s and Nikon Df. Just modern and low-pixel ones. There is almost no gain at high ISO. 2-3 times there and not close
As proof that you are right, Df
DXOmark. I used to look at their website very often, but I forgot about them for a couple of years.
There are actually two reasons. First, there is no way to compare two or three lenses from different manufacturers that compete with each other. But you usually visit such sites before buying optics or cameras. Many times it happened that, for example, a lens entered the market, say Sigma and a little later Tamron. One segment (price, focal, class), but Tamron's test never came out. And only 5 years later, when Tamron had already updated the line, they tested the lens. By that time, Sigma had already updated its line. I have cited two brands as examples, but this is the case with all. This is such a common story that the position of the site to please both yours and ours (without pushing competitors head-on) is annoying. I understand that yes, “the independent press is an illusion”, but there are other examples of similar sites. And the fact that they completely switched their attention from photographic equipment to smartphones only strengthened me in my innocence. This is the second reason. As for me, they donated the photo to the audience in vain. Do they really think that the person changing the smartphone comes to their website to check the light sensitivity curve? Or delving into the details of other measurements?
It's good that there have been alternatives lately. By the way, radojuva tested so many lenses and cameras during the period of the pandemic, more than a year before! I understand what a colossal work it is and thanks to Arkady for this!
“And in the comments, do not forget to write about your attitude towards DXOmark. Peace. Work. CMOS! ”
Once trusted, stopped some time ago (after buying Canon cameras and scandal with measurements on 1DX mk3, where they had to change numbers after outraged yells in the comments).
I have no questions about their value as ISO workers for D700, D3s - I had these cameras on hand and I was convinced of these numbers. There are questions about the measurements of the D4 and D750, the D750 is more likely to get a less noisy picture at ISO3200 than the D4. I kept D4 for myself, D750 was and I sold it. But I liked how the D750 works in the dark more.
I do not believe their ISO measurements for D800 / D800E / D810.
IMHO, if you forget about DXomark, progress is still going on, but slowly. There are now more low-noise cameras available for ISO2500-3200. There is also progress in the value of DD, although it is progressing at a snail's pace.
Nikon D3 - 12mp
Nikon Z7 ii - 45mp
Draw conclusions
We have made conclusions. Needed Xiaomi Mi 11because it has 50mp. It will make less noise :)
The usual Mi11 has 108MP%)
Hush, otherwise all the mirrors will scatter
So they are not visible in the photo ... Well, except for the size of the image ....
Shchta?
Nikon Z7 ii - 45mp and at high ISO the detail drops by 6MP as in d70 noise and are silent
Sony a7s series - 12MP total. ISO workers speak up to 25600.
I think dho +/- gives true information, but ... I will never believe that d3200 is not worse than 5 dm3. In theory, yes, maybe the d3200 can sometimes be removed "not worse", but in conjunction with the optics, even the DXO itself shows that there is nothing to catch the d3200 versus 5dm3. The cool thing is that DRL has debunked many myths of the "fat" pixel of the gnarled professionals, and in fact it turned out that the fat pixel often loses to the weak one in ISO. Where to develop? It seems to me that you can use AI in data processing like in smartphones. In the already mentioned d3200, I don't need 24MP, it's better to do RAW output with 13MP using 24MP data to raise the iso and noise reduction. This is a cool topic if you can't improve the sensor, but there is an opportunity to improve the processor. It's just that one experiment in January 20 threw me into shock, we went to St. Petersburg and took a few photos on the Galaxy s8 smartphone: where my camera would take ISO 3200 s8 took ISO 360 and the pictures turned out. Yes, there was the wildest shumodav, but the pictures are bearable.
Would you put a normal glass with f / 3200 and an optical stub on your d1.7, and on the d3200 it would be ISO 320 (or even less), and the pictures would be obviously better than on the Galaxy))
And where are these "passable" pictures then? On Instagram? ;)
So ISO is just a number.
I had a D3200 and still have an s8. I had a wonderful opportunity to compare pictures from both in different conditions. s8, alas, is not suitable for a DSLR, even if the latter has a kit lens. And software processing of images by a smartphone looks good only on its screen, in a small area. If you open it on a computer on a large monitor, you can immediately see the oversharp and the colors twisted to the maximum and the excessively suppressed noises and other “joys”.
But what the s8 is really good at is panoramas. The quality comes out quite decent for printing on a page spread of an A4 photobook, I was even surprised at the result myself. Apparently, when stitching panoramas, there is not enough power to turn on the "enhancers" and the image comes out more natural. However, this is just my guess, I have not thoroughly checked it.
And a DSLR can't do automatic panoramas :(
Sony SLT A58 furry 2013 is great at automatic panoramas. Yes, and all similar Sony cameras of those years are able to. But the Nikon 3000 series does not even know exposure bracketing. So this is just Nikon marketing, not a technical impossibility to organize this function.
Yes, there is no technical limitation, you are right. Moreover, one of the D3xxx (I forgot which one) is still capable of panoramas (though only 1080 pixels vertically, if I'm not mistaken), but not D7xxx and not full-frame models. This deliberate restriction just infuriates me.
And this A58 could make panoramas at any focal length, even through a telephoto lens? And were there any restrictions on the size of the image?
I went to the Internet for instructions. Here are some excerpts:
• In [Sweep Panorama], it is recommended to use
wide angle lens.
• When using a lens with a long focal length
distance, pan or tilt the camera
slower than when using wide-angle
the lens.
And the Specifications section:
PANORAMA SIZE: MAX .. COVERAGE ANGLE (FOCAL DISTANCE 16 MM / 18 MM)
Wide: horizontal 12 x 416 (1856 MP), vertical 23 x 5536 (2160 MP), Standard: horizontal 12 x 8192 (1856 MP), vertical 15 x 3872 (2160 MP)
For a smartphone, being able to take panoramas is a really serious advantage, because 99% of its users will never fuss with specialized software. For a DSLR, not everything is so simple ...
I will support the commentators above, but I will add some of my observations. In fact, at the moment, ISO workers have reached such a level that they allow you to shoot in everyday life without noise on dark lenses such as 28-60 from Sony or new dark zooms from Canon, and some 24-105 / 4 can become the main lens for almost any conditions, because even iso 32-40k can be called conditionally working on the same nikon z7 and cameras of its generation. At the same time, for professionals working in difficult conditions, high-aperture and super-high-aperture optics are still needed. IMHO, due to the increase in megapixels in all segments, the same limit value is maintained, I think in the very near future (with the release of a7iv) in the most massive segment, not the averaged 24MP, but 32-38MP will become popular.
in most shooting conditions, 2000 asa is enough….
“You can't jump above your head and you won't be able to squeeze something significantly larger out of CMOS” Do not add, do not add. Precisely stated. Physics cannot be defeated. All attempts to further increase pixels will lead to a dead end. It will only get worse from now on. Or maybe already a dead end. Know physics. And that's it.
Physics does not need to be defeated. You have to put your head on. Stabilization, multishooting (like in Google Pixel), noise reduction with frenzied resolution, non-standard arrangement of color filters - and from old technologies it turns out to squeeze out a much more decent picture. Compare smartphones today and ten years ago - and the size of the sensor there has increased quite a bit.
But the size of the sensors has also grown - the difference between a ten-year-old typical camera phone and a typical smartphone 2020-21 is three times in the area of sensors, this is still more than the difference between crop and ff, for example.
Well, the software chips, of course, but the result is not always happy.
The main breakthrough in matrix technology took place in 2005-2012. Compare d70> d80> d90> d7000> d7100. Each next camera has a much better sensor: higher working ISO, DD, lower noise. And this is with a constant increase in resolution.
Recently I shot at the same time on d90 and d5600. And it's just heaven and earth. The d5600 has a much higher ISO worker. On d90, ISO 1600 is practically inoperative, and on d5600 you can put 3200 and not worry about the result. In addition, 24 Mpx gives more headroom for noise cancellation. Therefore, the real difference in working ISO is 2-3 steps
I once read an explanatory article where they tried to count the number of photons per photosensitive element and according to all these calculations it turned out that the best FF cameras at high ISO have an efficiency close enough to the physical limit.
So the matter is not in CMOS and lack of technical progress, just the limit has been reached and the only way today is matrix of a larger physical size.
The larger the sensor, the larger the lenses. The market is moving in a different direction. Again, mobile phones - even 5 years ago, no one would have thought to lay out a camera, leaving on vacation. Today it is a reality for the majority.
Also, look at the same fifty dollars for DSLRs and bzk: Nikkor Z 50 f / 1.8 is twice the 1.8G. Yes, telephoto lenses are finally able to use Fresnel lenses, but at wide and standard angles, almost all optics are larger than 30 years ago. Now interpolate that into larger matrices. I don’t think this is the right way, because the average photographer also has physical limitations on the weight and size of his equipment.
Well, we want to improve the performance, so either we buy lenses from 30 years ago, or heavier - larger - with a larger number of elements.
And as for laying out the camera and replacing it with mobiles ... Well, mobile phones have simply reached the limit of photographic skills of most users. They don't care how their favorite child is filmed on the beach with someone's fat belly in the frame - it's a full frame or a phone - there is no difference.
“More matrix - more lenses” here I completely agree.
But lenses in general are becoming more compact, this is an unambiguous trend, especially if you look at the glasses that are now being produced by Sony, and not some Pentax thread there.
We look at the fresh Sony 14 f / 1.8 and compare it with a similar Sigma.
We are looking at the latest fifty kopecks Sony f / 1.2.
Let's look at the fresh 35 1.4 Sony and compare with the old Zeiss 1.4, which is huge and not very good optically.
Etc. etc.
Yes, some lenses have grown in size, but this is only when compared with all the old ones, which were terrible in quality and were not even close to being able to resolve the modern multi-megapixel matrix.
And if we talk specifically about fifty dollars, then recently there was just a gorgeous article on the topic on dpreview “Why are modern 50mm lenses so damned complicated?”.
And yes, photographic equipment more and more goes into the hight end zone. Otherwise, it is impossible to compete with smartphones. Therefore, xs, maybe in 10 years all old-school camera lovers without exception will switch to medium format.
Please don’t drive to Pentax, they have a lot of miniature "pancakes" in their lens line. And they were there long before any sleepyheads.
+1, the pentax had a revered compactness even before it became fashionable.
Who cares what happened a hundred years ago?
Looking at last year:
Pentax-D FA * 85mm F1.4 ED SDM AW - three hundred kilo
Sigma 85mm F1.4 DG DN Art - 600 grams.
At the same time, Sigma, at least, is not worse optically.
And right there next to the 77 / 1.8 weighing 270 g.
How old is this lens? twenty?
“And yes, more and more photography equipment goes into the hight end zone. Otherwise, it is impossible to compete with smartphones ”
Smartphones are not qualitatively competitors even for the initial DSLRs with whale glasses. Here the point is different.
Smartphones give a good result in one click; in the second click, this result is available to all your friends. An initial DSLR with a whale is already a lot of buttons and twists, and then sitting at home on a campotera pulls, because a normal jeep is hard to do right away (the camera doesn't think for the user as it should). And then more heavy pictures to load somewhere to someone. In short, it is long, inconvenient - literally “for an amateur”. So the extinction of inexpensive cameras is understandable and logical.
We still speak for amateurs, not amateurs, who bought a DSLR “to shoot on vacation, because Vasya's neighbor also has one”.
If a person is not quite an amateur and has an idea of what a converter is, where to turn the sliders, the difference will be very noticeable, no one has yet been able to deceive physics.
Nevertheless, if we talk about the 21st year, the majority of ordinary people still view content on the displays of mobile devices, and there the difference is already erased, yes.
competitors, because mere mortals bought exactly this kind of camera until they switched to smartphones
And you'd better look at lenses over 50mm, and it is advisable not to compare SLR (from sigma in your example) and mirrorless. There, the super-width is huge only because of the very large working distance, which is three times less in Sony.
Yes, probably. Moreover, high ISO is the result of amplification. There you can play with the noise of the op amps used, but the inhomogeneity of the image due to the lack of photons cannot be overcome in any way - there is nothing to register simply.
Yes, nothing waits, smartphones will kill cameras completely
It still seems to me that the point is in the rating scale. DxO puts 30 dB and says that this is the equivalent of an excellent photo. But I know a good saying: the best is the enemy of the good. Where, well, just good photos are quite enough (namely: ISO6400 and higher), modern cameras have made a huge leap forward. Yes, pictures in good light from an old Df or even my D600, when compared head-on with modern noise carcasses, can be plus or minus about the same thing. But one has only to start a conversation for high sensitivity, as it turns out that everything is not so rosy. What kind of ISO workers does Z have - 12800, 25600? I have a complete mess on the D600 in this mode, but I would just need a good shot. Therefore, modern cameras are better than old ones.
Open the Instagram feed from your laptop and you will see a big difference between the “phone” shots and what you see through the smartphone screen. The great success of mobile phones as a means of obtaining photographs is due to the development of social networks and smartphones in general - most users surf the web through mobile devices, on the screen of which pictures from smartphones look noticeably better than on large screens of home computers. No matter what they say, most people will not give up the prospect of arranging a photo session with a camera at all, although, of course, they will still say behind the eyes that the smartphone can shoot just as well.
Nowadays, not everyone already has home computers. All through a mobile phone
Instagram is a very unfortunate example, since it compresses images to 1080 pixels in width and jacks them down to ~ 100-200KB, so that in addition, small details are completely eaten up, the sky is posterized into all fields. displays the capabilities of modern smartphones.
I wonder what Panas has with the type of “different” native ISOs?
There is only one native ISO, and how the noise is reduced and how it is presented to the consumer (read, inflate consumers) is a separate question.
As I understand it, there is some kind of hardware chip
Yes, controlled light transmission, for this there is a special layer on the matrices, up to certain ISOs, less light is given to the matrix, and after the amount of light is increased, this allows you to reduce the gain and get less noise at high ISOs, that's the whole focus. This is used by Nikon and Sony, etc. only it is called differently or not advertised at all and the goals are slightly different.
Not really, see below. Two sets of amplifiers with different characteristics.
Two sets of analog op amps. The matrix has a certain sensitivity to light. So she picked up a certain amount of photons and gave out some voltage, depending on this amount. Then there is an operational amplifier with its variable gain, which multiplies the analog voltage. Further ADC digitizes all this. Naturally, the amplifier introduces some noises, but, naturally, they do not grow by a multiple of the signal amplification, a flatter dependence. But when we have already received the number, then you can multiply all the values by some factor and raise the exposure (or divide by the factor) and lower the exposure - here the noise grows proportionally. All these 1/3 ISO, Hi ISO, Lo ISO are received in digital form. Like there ISO 640, ISO 320, ISO 12800. The basic one needs to be clarified, which depends on the matrix, Panasonic seems to have 800, the others have 400 or 200 there, so 100-200 multiples of it are analog gain-attenuation.
Panasonic (but as far as I understand not only they) put another set of op-amps with a deliberately high base gain and lower noise. So they get one native ISO 800, which is from the matrix, from it 100-200-400, 1600-3200 due to analog amplification, from them 160, 320, 640 due to digital. And the second set - 5000 and what multiples are there, amplified by another part of the analog circuit, perhaps even with its own separate ADC. Something like this.
But as I understand it, they are not the only ones, the same Canon Magic Lantern is able to shoot in Dual ISO mode, perhaps Panasonic is the first who voiced this explicitly and sold it as a technology.
This is how, after reading the Internet, or rather after watching beautiful color pictures, people have false ideas about 2 different operational amplifiers in matrices, especially when people do not know what an operational amplifier is and how it works))) By the way, digital op amps do not exist, they are always analog)) )
Actually, this Internet waste paper is designed, because 2 analog amplifiers sound cooler than some kind of film with variable light transmission.
In simple terms, to make it clearer what digital noise is, though we have already written about this above. Take an ideal source of white light and take a picture on a base ISO, let it be ISO 100, if you hope that you will get conditional 0.1 volts from the matrix for each of the three color channels (for example), then this is a delusion, in reality it will be, for example, R = 0.12 G = 0.13 B = 0.128 and for each subpixel this value will float up or down due to the imperfection of the elements that make up the matrix, for the base ISO when digitizing, you will not notice the difference, but if you amplify the signal 10 times, it will already be 1.2, 1.3 and 1.28, what will be the result after digitization? That's right, the result will already be a pixel that is not white, and the more we enhance the more the dominance of which one color will be as a result. How can this difference be reduced? Using another mythical analog amplifier? No, by increasing the amount of light and decreasing the gain.
I expressed the thought clumsily. Look. If, for example, at some light intensity, we expect to receive from the matrix conditional 1 volt voltage for each of the color channels, and at a reduced intensity of 0.1 volts, which then need to be amplified 10 times to obtain the same result, then in both cases in the signal from the matrix to the input of the amplifier, the difference in the channels will remain at the minimum level and will be practically equal, but as soon as we start to amplify the signal, the difference in the channels increases proportionally, and the more we amplify, the greater the difference in the channels will be, which is our noise. This is not solved by any wonderful amplifiers with other characteristics, because the source of this difference is the matrix itself, even before the amplification. For the same reason, increasing the shutter speed produces less noise than increasing the gain (ISO).
"By the way, there are no digital op amps, they are always analog" - thanks, cap. I used the term “analog” to emphasize the essence of the processes. Analog gain, not digital, no bit limit. The quantization level is set only by the number of photons and amplifier noise.
“Some kind of film with variable light transmission” - I have never seen such an explanation. I will be glad to see.
https://sites.psu.edu/vsg5016/files/2015/09/ProductDescription-CMOSImageSensor-1w9kspy.pdf - here is a schematic diagram of the CMOS sensor. One of the Dual Native ISO options that I once came across involves the use of two non-pixel amplifiers with different characteristics.
Here is another diagram - https://www.canonrumors.com/forum/attachments/429a0821-dad2-43c7-be27-51ce7f37f2ee-jpeg.182242/ - is also based on two switchable (or connected together) analog sections of the circuit (that is, connected before the DAC).
Varaints with “deliberately darkening the film so as not to let the photons pass, so that later we still need to amplify the weakened signal with our hands” did not come across.
There are no amplifiers with different characteristics, all operational amplifiers work the same, especially when it comes to amplifying a constant voltage, and from a matrix they amplify a constant voltage. I have friends over the hill who have studied the device of matrices from different cameras in detail, including the b-es-ai matrices. I think it's no secret that many offices, to one degree or another, study the products of competitors or the products they are interested in, and they told me something and did not ask to keep it secret, but I will not say what kind of office it is. You are not looking at beautiful pictures with two amplifiers, include logic and head. If there were 2 different amps, why not use more “ideal” gains across the entire ISO range, rather than ISO 400 for crops and ISO 800 for full frame and how they generally do an ISO 50 extension on my D750 while doing this. there is no increase in dynamic range and noise, despite the fact that the gain formula is 1 + R1 / R2, that is, less than 1 in any way, and according to these friends there is not an inverting amplifier.
ISO50 is purely software. As well as others extended upwards. As well as intermediate 1/3. We simply multiply or divide by the coefficient with all that it implies.
One amplifier - two amplifiers - the point is that there are two elements of the analog part that work differently in different lighting conditions and have different characteristics. Perhaps by noise, perhaps by the capacitance of the capacitor, with which they hold the charge before reading, it is necessary to dig.
While the answer with links to moderation, I will continue.
“Take the perfect source of white light” is not just a clumsy thought, it is a wild lol. https://upload.wikimedia.org/wikipedia/commons/thumb/b/ba/PlanckianLocus.png/1200px-PlanckianLocus.png
Here you have a color temperature curve marked on the locus. It's all white. Which one do you prefer at this time of day? The camera has no concept of white. It is obtained by setting a white point after digitizing, where you put it, there will be 255, 255, 255 (in the selected color space, of course).
“But if you amplify the signal 10 times, it will already be 1.2, 1.3 and 1.28, which will be the result after digitization” - the result will be an amplified signal. The proportions between the channels are the same. Indeed, a clumsy explanation. And subsequent attempts are no better.
“The source of this difference is the matrix itself even before amplification” - the matrix has a lot of different noises. What you are wretchedly trying to explain is called photon noise. When the number of photons in low light is too small. You have 20, 30 and 40 photons hitting three channels, with an amplification of 1000 times it will be 20, 000, and 30. But under normal illumination, you got 000, 40 and 000 photons. With an increase of 1000 times it will be 1100, 1200 and 40. But the matrix has nothing to do with it, it physically cannot capture photons that are not emitted, so yes, no amplification will save, you need to increase the shutter speed, and with it the number photons.
And then there is the matrix's own noise, including thermal noise. And there, on the one hand, raising the shutter speed, we increase the number of photons, and on the other, the heating matrix makes more noise, so the telescope matrices are cooled. And then there is the amplifier, which is also noisy. And its noise is added to the noise of the matrix. But if he amplified the signal 1000 times, this does not mean that he added his noise a thousand times, there is a nonlinear dependence. But when we start playing with the already digitized signal, there we increase ALL noise at the same time. And photonic, and thermal, and noise of the amplifier circuit.
Yes, but the biggest fat and unpleasant one is reading noise. And with different hardware fillings (different I - V characteristics of the op amp), it will be different. So, despite the criticism above, things like Dual Native ISO are quite interesting.
Well, again, answering the question above - why there are only two alternative schemes, and not one for each ISO - it is likely that there will be more of them in the future.
You confuse different things a little, but oh well. it's pointless to argue.
“Well, again, answering the question above - why there are only two alternative schemes, and not one for each ISO - it is likely that there will be more of them in the future.”
Because there are no two analog amplification circuits and there will never be more than one. The matrix works very simply and the principle is based on a controlled discharge of a capacitor, on the one hand, it is a photosensitive element, on the other hand, a field-effect transistor that corrects the discharge rate (shutter speed), just look at the circuits you have given, as well as patents, and all this is clearly visible. Under normal conditions, the thermal noise will be below the 14-bit addressing limit, which can be seen even in the ravs, the lion's share of the problems is the spread of the characteristics of the channel pixels, followed by linear amplification with a proportional increase in the spread, while the noise of opamp and ADC is not even on the second or third place, put there at least super precision low-noise operational amplifiers for 100 bucks apiece, nothing will change, their noise even with a high gain will also be outside the 14-bit ADC, the maximum is at the border. Thermal noise begins to significantly affect only long exposures and is caused by the temperature drift of the characteristics of the pixel elements due to which the discharge rate of the capacitors floats. I am a radio amateur (and not only) with an experience of under 50 years and the corresponding specialized education, and I perfectly understand what parasitic phenomena are in this implementation, and how, and what they affect, and I also know that another mythical analog amplification circuit cannot reduce nonlinearity, it can only be reduced to an amplifier.
"Because there are no two analog amplification circuits and there will never be more than one."
https://andor.oxinst.com/learning/view/article/dual-amplifier-dynamic-range - sCMOS sensor offers a unique dual amplifier architecture, meaning that signal from each pixel can be sampled simultaneously by both high and low gain amplifiers.
The only thing is that the signal is still summed and not switched between two amplifiers. Canon solves the same problem with amplifiers operating in two modes - Hi Gain and Lo Gain. https://nofilmschool.com/sites/default/files/whitepaper_canon_dgo.pdf
Arri does the same. The Dual Gain Architecture simultaneously provides two separate read-out paths from each pixel with different amplification. The first path contains the regular, highly amplified signal. The second path contains a signal with lower amplification to capture the information that is clipped in the first path.
Aptina does the same, attaching an additional capacitor to each pixel for good lighting conditions - https://www.photonstophotos.net/Aptina/DR-Pix_WhitePaper.pdf (i.e. at low ISO).
That is, the process differs from the one described by me (I will still look for the source, I did not take it from scratch), but it comes down to about the same - hardware support for different modes of operation of the analog part. Far enough from what you described "Yes, controlled light transmission, there is a special layer on the matrices for this."
Oooh, the holivar topic has been touched upon ...
I don’t know for the measurement methods and how much I like 30 dB in a single scene, but I can say this: modern cameras allow you to shoot without any problems in difficult conditions of lack of light with dark zooms. And I'm talking about the crop.
Now dive into the past and remember the gnashing teeth with which you set ISO 800 on cameras 12 or 14 years ago.
There is nothing to argue here, even now any adequate camera will take a wonderful picture even in dubious conditions. And on a clear day, a 15-year-old mammoth will give an adequate picture
They are marking time, apparently because there is nowhere to develop further. Of course, when completely different technologies are used, which we have no idea about now, then progress will begin.
Nothing, they will finish the plenoptics. Learn to make good diffractive lenses. They will learn how to make an aspheric of a complex profile. For monochromatic radiation, they have already been able to correct all aberrations with one lens within the sane angle. Further, they will throw more computing power. And there will be no more bulky lenses or large matrices. And at the same time, there will be no chatter about drawing, bokeh, color, etc. - this will be calculated with a couple of sliders in real-time or at least in post-processing. In general, this is not the most difficult mathematics. There will be no front and back focus. And in general - focus. Why is he in a plenoptic camera?) There will be no large matrices - firstly, it is expensive, and secondly - they will not make much sense. Physics will be deceived by "wiser" methods by accumulating a signal, or even purely calculations. Photography will become the fiercest casual game with the dominance of software even more trenchant than it is now. People with good old matrices and closures, as well as large tubes around their necks, will be perceived as people now who shoot on film SF or collodion.
Here is a straight knife on Zeiss and a sickle on the matrices.
Something, I suspect, Zeiss conducts research. And Canon has had high-volume optics with diffractive lenses for a very long time.
And only the Krasnogorsk Order of Lenin, the Order of the Red Banner of Labor, the personal plant of His Imperial Highness ...
I understand that it is easier for manufacturers to “give birth” to devices with 100, 200 and more megapixels than to give “honest and clean” 10 megapixels at once.
First, devices with XNUMXx megapixels sell better; secondly, you can't fit a full-frame sensor into a smartphone.
This is about half the topic. 100 megapixels downsized to an honest 10 will look better than an honest 10, including in terms of noise. Perhaps this is the answer to Arkady's old question.
Still no one complains that the picture from the new phones looks worse than the picture from the old ones. The capacities allow you to grind these megapixels. There is no conspiracy.
For a long time now, I believe that Nikon, confidently occupying the second place in the global ranking of photographic brands, always gives a little more and a little cheaper than Canon with the exact same camera.
Now both major brands have waited, looked at the mistakes of Sony, FujiFilm, made sure of the expectations of the amateur photographers market and began to “cut cabbage” on mirrorless cameras ...
More: https://udivil.com/canon-ili-nikon-d7500-eos-200d/
“Now both major brands have waited, looked at the mistakes of Sony, FujiFilm”
Sonya made so many "mistakes" that the owners of Canon reached out to her, began to sell their cameras and hang L-ki on Sonya through adapters.
Fuji's “mistakes” are also curious. GFX is a good “mistake” that costs the same as Canon R5 (or rather, on the contrary, the full frame has pulled up at a price to budget SF cameras).
About SF now, of course, "experts" will say that this is a sub-crop, well then there is a Pentax 645Z on exactly the same matrix.
"More:"
But this link is not needed here.
terrible article
So this is just a barely disguised advertisement for the promotion of a blog with a variety of parsley of the same "articles" (not informational, but for the sake of attracting an audience)
I would react accordingly.
I have a big suspicion that noise is more about a property of light than a matrix.
Only so far I still can't get myself together to mathematically substantiate this (or refute)
How can light be noisy? In my opinion, the noise just depends mainly on the matrix and its binding. Two different matrices transmit the same light with different noise levels :)
> And how can light make noise?
Easily! It's called shot noise. After all, the photon energy is quantized, and if the cells literally catch several photons, then a difference of one photon leads to the presence of some irreducible difference in the levels of the read signal. But this is about astronomical matrices floating in liquid nitrogen (and, maybe, about A7s in the light of stars), and not about everyday photography, when there are a lot of photons.
Data from DxO does not always correlate with reality. For example, if we compare Nikon D80 and D200. DxO rated a conditionally good ISO for the D80 at 524, and for the D200 at 581. It would seem that the difference is somewhere in 1/6 of a stop. But in reality this is not the case: the D80 is noticeably more noisy than the D200 - by about 2/3 of a stop. I own both cameras, I made a personal comparison under the same conditions, the difference in noise was estimated by eye. ISO500 on the D80 is about the same noise as ISO800 on the D200.
Second example: Fuji S3 Pro. DxO has given this camera an ISO Sports rating of 346.
For the Nikon D40, it is rated as 561. So, in fact, at high ISOs (400, 800), Protroika has visually less noise.
I suppose that the point here is that DxO, when calculating ISO Sports, takes into account not only the signal-to-noise ratio itself, but, as Arkady wrote above, also DD of 9 stops and color depth.
Third example: DD cameras Nikon D700 and D3. DxO estimates them about the same - 12,2 stops.
However, if we read the camera reviews on DpReview, there the D3's DD is 3/4 stop higher: 8,6 eV versus 7,8 eV for the D700.
And the very difference between the data with DpReview and Dxo Mark raises questions - the difference with DpReview is 3 steps.
Based on the experience of shooting on the D700, I can say: I don’t feel a lack of DD. Highlights, of course, likes to lose, but we make allowances for age. Yes, and here it’s more likely a photographer’s mistake, and not an imperfection of the camera :)
Arkady and colleagues, what do you think about Foveon matrices and the prospects for Full frame foveon, which Sigma promised to roll out just now.
https://www.dpreview.com/news/6995633043/sigma-says-its-full-frame-foveon-x3-sensor-will-be-ready-sometime-this-year
Is it worth starting to look towards Panasonic Sigma as a promising photosystem of the future? :-)
Camera production at Sigma is just an image hobby.
They've been wanting to design a full-frame sensor for a long time, but technological issues are getting in the way. The latest model on Foveon is Sigma Quattro H with APS-H matrix.
The problem with cameras on Foveon is that there is a base ISO and all the rest. At the base ISO, the best result. According to DD at the base ISO, Sigma Quattro H was inferior to Sony A7Rm2 / A7Rm3.
Another Sigma problem is the lack of free sale of cameras, lenses for this mount, batteries, battery handles / blocks and the lack of service. Another negative point is gluttony.
And in terms of service - there is 1 in Moscow for the whole of Russia. I had a Sigma SD1 Merrill.
FF should be released. But due to the peculiarities of the technology, high ISOs cannot be seen there.