Photography Chat

12:02 AM

and you might be thinking "hah, its just the street ligths that are yellow, my town is also yellow at night"

daytime is also yellow and blue :)

ah, about QE for nikons, its about comparable. I had jsut seen the D800 wooping 57%, but something like D5000 is also mid30s

I dont have any conception of the nikon line naming so I dont know what "translates" to xxxD,xxD,xD canon series.. its getting harder with canon too, when 60D is not continueing the xxD series but 7D is...

jrista

@MichaelNielsen There are other ways to improve S/N as well. Canon uses a bias offset, which allows them some freedoms Nikon does not have, which is probably where the slight increase in S/N at high ISo comes from.

@MichaelNielsen Also, the latest Canon sensors are around 50% as well.

The 1D X is 47%, the 5D III is 49%.

The D3s, D4, and D800 are the only ones that have higher than 50% Q.E., and the highest is 57%.

Nikon's best currently has less than a 10% Q.E. lead on Canon's best.

Which technically SHOULD allow them to achieve at least ISO 51200 natively, but there are some significant differences in how Sony Exmor works that seems to preclude that.

I believe the highest native ISO exmor can achieve is 12800.

It may be due to the way their on-die pipeline is split into analog and digital parts.

Most of Nikon's own sensors (the ones they manufacture(d), not Sony's) have worse Q.E. than Canon sensors.

Michael Nielsen

12:21 AM

and then tehres the read noise numbers on the site. 50D has less than 40D while reviews went crazy over that 40D has less noise than 50D, complaining they put too many pixels per area with the 50D

jrista

there is read noise and photon noise

read noise may be less

as an intrinsic property of the sensor

photon noise, in the abscense of a sufficient increase in Q.E., is a property of pixel density

Michael Nielsen

I thoguht read noise was so much higher than photon noise that photon noise was leglitiable in comparison

jrista

the 50D has smaller pixels, but only a 5% improvement in Q.E.

Michael Nielsen

igher = more annoying

jrista

@MichaelNielsen hell no...photon noise is dominant except in the lowest signal ratios

Michael Nielsen

12:22 AM

negli...

jrista

Photon noise exists at every brightness level

it is due to the random nature of light

It is why a blue sky tends to be noisy

Michael Nielsen

yes I know the poisson dsitribution

jrista

the sky itself may be fairly bright, but it still tends to be noisy, even on the best sensor

read noise, on the other hand...

at it's worst it is 38.2e- (in the 1D X)

the 1D X saturation point is 90367e-

the maximum read noise level is a mere 0.0004% of the maximum signal

It also only matters at the lowest ISO settings.

and then only in the shadows

Michael Nielsen

the noise is higher at higher brigthnes but you notice it less

jrista

the benefit of Sony Exmor is it's digital noise reduction in its CP-ADC logic

@MichaelNielsen photon noise, yes, read noise no

Michael Nielsen

12:25 AM

unlike the read noise that you notice in the blacks even if the iamge is well exposed

jrista

yeah

it is all a matter of S/N

in the shadows, your signal may only be 100e-, maybe even 50e- for deep blacks

at that level, noise is a significant percentage of the signal

so it is more apparent

since exmor effectively has a 3 e- read noise level at all ISO settings, read noise is the same percentage of a low signal regardless

another thing, Canon uses a bias offset

where as Nikon/Sony do not

nikon and sony effectively clip blacks entirely above the noise floor

and it is not possible to register a signal below that clipping level

Canon, on the other hand, since it uses a bias offset, renders shadows or low exposures frequently unusable because signal and noise become mixed

and some of their read noise only exhibits as high as 25-35e- in certain areas (such as FPN and HVBN)

Michael Nielsen

Ive been surpised at what my 40D can do today. I took iamges I thoguht Id need a fullframe to do, and all it took was a good lens and manual focus (well, manual everything).

jrista

so there can be usable image pixels in between horizontal and vertical bands

but because that signal is intruded upon by read noise, the whole thing is effectively unusable

Canon needs to rethink their sensor design, and find a way to either eliminate read noise (or at least reduce it to 3e- or less), or clip the black levels above the maximum level of read noise

Michael Nielsen

ah so thats why. sometimes I jsut have to go b/w to make it usable if its underexposed

jrista

Ironically, that is how most sensor designs work.

Only Exmor is really significantly different

But Exmor is being used in more and more of Canon's competition, and they need to step up their game.

Michael Nielsen

12:31 AM

even if it is nicely exposed than a person wearing a black jacket can be annoying

I have exmor in my ninja camera

Im not a fan

jrista

?

Michael Nielsen

washed out details

sony wx1 :)

but it records concerts really well

maybe my lumia 920 can replace it

will wsee at next concert

jrista

oh, well

that is a BSI small form factor sensor

it is an Exmor, but the pixels are TINY

that is like comparing an elephant to a flamingo

Exmor and Exmor R are not really comparable

Michael Nielsen

but still it seems to me that nikon vs canon, canon has more micro contrast underneath the noise, and nikons wash out some details, have you the same impression?

jrista

yeah, I do

regardless, images from a Nikon have far more usable shadows, washed out or not

Canon's deepest black levels (negative bias values) definitely seem to be more contrasty, with a slight red tint sometimes. Nikon's deepest black levels (down to zero, they do not have negative bias) appear to be blueish gray. I think it is the blueish gray tint that makes Nikon shadows appear a bit washed.

Michael Nielsen

12:50 AM

well, in post we can clip teh blacks above the noise floor too, if we choose to. but we get teh choice. why should sony do it it for you , so you can never get the details back?

jrista

well

that is not quite what I mean

they clip all noise

there is always a certain amount of noise in a circuit

in an Exmor, the value that registers as 0 is actually not at purely zero electrons.

there are always some electrons in the circuit

Sony clips whatever the average of the amount of noise in the circuit is

so, if it is say around 30 on average

then they would set the zero level at 31

now, there are some fluctuations, so there will be a low level of read noise, and there are still FPN and HVBN noise sources

the rest is eliminated by their CP-ADC

Canon differs, in that they effectively use zero electrons as the minimum possible signal level, and set the base image signal readout level at some bias offset.

Michael Nielsen

and if I udnerstood you correctly canon substracts divides the 30 by , say, 5 so teh 0-30 is scaled to 6 and clips at 0 so the signal details are intact, but added some noise?

jrista

Because of the bias offset, Canon's signal effectively ranges from some negative number (-512, -1024) up through the maximum signal, which is 2^14 minus the bias offset (1024, 2048, etc.)

Now, Canon can choose to clip blacks (set the black point) wherever they want in the bias offset, a range of 1024 or 2048 in most of their current cameras.

they could pick black to be -137

it could be -989

if it is -989, and read noise is 38.2

well, some noise will leak into the image signal

if they pick he full -1024, then there is the potential of 38 to 39 electrons worth of electronic noise in each pixel leaking into the image signal.

Michael Nielsen

if sonys adc has NR , that should be the reason they sacrifice detail

jrista

Canon uses an analog CDS, but it is not nearly as effective as Exmor's CP-ADC and digital noise removal

Michael Nielsen

1:01 AM

all filters filter signal details too

even teh advenced wavelets

jrista

Sony's CP-ADC uses the reset read to determine read signal, pattern noise, banding noise, and PRNU

it stores the "read noise level" of each pixel in CP-ADC at reset as a negative digital offset

Michael Nielsen

so they both have ofset, one analogue and teh other digital

jrista

then, when it reads the actual image signal, which will have roughly the same amount of read noise for each pixel, the negative offset previously recorded effectively eliminates read noise in the image, and once positive digital units are read, they represent nearly pure image signal

well, bias offset is very much an analog and signal power thing

bias offset is not much like the kind of digital offset Exmor uses

Exmor's digital offsets are usually quite small for each pixel

a few units or so

Canon's bias offset can be quite large, and it is effectively dynamic and arbitrary

Canon's offset effectively allows useful signal and electronic circuit signal to mix in a bad way.

Exmor's offset is explicitly designed to identify and eliminate electronic circuit signal as precisely as possible, leaving behind the purest image signal they can.

Michael Nielsen

so tehy a assuming that the read noise level is hte same teh usec later after the reset

which its not

as its not a smooth sine curve

that can be predicted

as far I remmber the read noise is uniform random generator

jrista

eh, photon noise is gaussian

read noise tends to have certain patterns

Michael Nielsen

1:09 AM

so it can be (at eth range -100 to +100) -99 at teh reset when teh measure and +100 the next

jrista

there is some random element to it

the exact charge in the pixel at reset time can vary a little

Michael Nielsen

and photon noise is poisson

jrista

but it is generally not significant

Michael Nielsen

and the elctronic noise of guassian

jrista

poisson, gaussian, same basic thing

Michael Nielsen

1:09 AM

no

gaussian is fixed std dev

possion has higher std dev at higher brightness

jrista

it is a probability curve

both are probability curves

the nasty kinds of read noise tend to be fixed

Michael Nielsen

uniform noise is a p cruve too, its jsut constant :)

jrista

read noise comes in many varieties

there is no simple gaussian probability curve for "read noise"

Michael Nielsen

has the "window box" shape

jrista

for reset current there might be

but reset current is only one type of noise

there is fixed pattern noise, which is a permanent attribute of a given sensor

Michael Nielsen

1:11 AM

the box pdf is a pdf no less than the guassian pdf

jrista

there is banding noise, which tends to be a mix of fixed and semi-random noise

there is also a slight differential between the response of each transistor

Michael Nielsen

are you taliogn about the fixed cmos pattern?

jrista

which might follow a gaussian distribution

there are thermal noise effects, which tend to occur in localized areas

and which also enhance the impact of FPN

Michael Nielsen

where each pixel has its own unique gain slifhtly differnt form its next?

jrista

well, no

There are a couple patents for variable per-pixel gain

Michael Nielsen

1:13 AM

not talking abotu that

jrista

but I don't believe any commercial sensor in a DSLR uses that yet

I mean response

not every pixel will respond the same way to stimulous

Michael Nielsen

Im takning about that evern since cmos was invented, they had the trouble veruss ccd that the brightness of each pixel varied

jrista

nor will every transistor at each pixel behave exactly the same way

Michael Nielsen

so they have to subtract a refrence frame

jrista

@MichaelNielsen CDS, Correlated Double Sampling

Canon uses an analog form, Exmor uses the digital approach I described

analog CDS tries to read out the reset current when the signal is read, similar to CP-ADC digital CDS

but it is not ideal...the reset charge can differ from the same charge when the exposure is read out

Canon has to deal with signal levels right down to zero electrons

Exmor, as I gather, has the benefit of most read noise simply being clipped, so the amount of reset signal to deal with in the first place is fairly low

and this has smaller variation

Michael Nielsen

1:17 AM

but the raw converter can choose its own black level, above the noise (ie you as user) ie yo ucan clip it like xmor does

but you get the choice

jrista

no, heh

in Canon, useful image signal and read noise are mixed

useful image signal may only go to -350, where as teh bias offset allows as much as -1024. Read noise may reach right up to signal level 0 or higher.

In an Exmor

image signal starts where read noise is clipped

I don't know how they do it

maybe it is all done through their CP-ADC

but the patents and documents I've read only describe the digital CDS and the ability of each ADC to compensate for per-column banding

Michael Nielsen

if we agre that they have lost details that explanation doesnt make sense,. if tehy can magically clip teh noise wihtout signal they wouldnt lose any.

jrista

but in all of the papers, the digital CDS reset signal tends to be only a few units

not thousands

maybe in actuality the reset signal is much greater

if so, then all noise removal is handled as digital units

but I'd expect them to have the same problems Canon does with a bias offset

and they do not

you are most certainly not losing anything with an Exmor

the simple fact of the matter is that Exmor retains two stops of increased shadow detail that, in a canon, is mixed with read noise

Canon uses analog CDS, has no FPN or HBVN removal

and they have problems with all three...dark current, FPN, and HVBN

Exmor has very slight problems with dark current noise, however it definitely seems to follow a more random distribution pattern and the maximum noise signal is rarely higher than 4e-

That means that you can recover useful image detail right down to nearly ZERO signal level.

Michael Nielsen

I hope canon refrains from adding a filter you cant turn off

I dont believe in a filter that "knows" waht is noise and what isnt

jrista

It may be a bit washed out once you get down that far, but the simple fact of the matter is no canon sensor on earth allows you to make effective use of signal levels that low.

Canon blacks may be more contrasty, but they tend to be unusable because they have bright red banding noise and dark current noise all over the place...half the time you can't tell if it is noise or image detail your looking at when you get into the blacks.

So it doesn't matter that you can't "turn off the filter", which it isn't really a filter.

Exmor simply retains far more shadow detail than Canon's do, and the detail they retain is usable, without interference from banding or FPN until you get down to the bottom few levels of black (i.e. a pixel with RGB values 0,0,0; 1,1,1; and maybe 2,2,2)

If you want more contrast with an Exmor, tweak the tone curve, and your set.

Michael Nielsen

1:27 AM

and yet, you still use canon?

ergonomics?

jrista

Ergonomics, glass, yes.

But, I also tend to shoot at ISO above 800 95% of the time anyway.

So the benefit of Exmor is not applicable.

Exmor can only retain shadow detail that it can capture in the first placew.

You lose roughly a stop of dynamic range per stop of ISO.

At ISO 100 a 14-bit ADC will allow a bit less than 14 stops of DR.

Michael Nielsen

ooh another strange thing about the graphs. canon has a expetected DR curve

nikon data goes upa nd down

jrista

At ISO 200, you get a bit less than 13 stops of DR. At ISO 400, you get a bit less than 12 stops, which is right in line with most Canon sensors...which sit between 11.5 and 12 stops of DR for ISO's 100-400. At ISO 800, you are limited to just less than 11 stops of DR, etc. etc.

Michael Nielsen

I thoguht iso 200 was gain=1.0

jrista

Gain = 1.0 is called Unity Gain

it is very rare that you achieve unity gain at a selectable ISO setting

gain at ISO 100 is usually greater than 1

Michael Nielsen

1:29 AM

ok.. I mean iso 200 = std gain

and iso 100 = attenuation

400 = boosting

jrista

um...no

Michael Nielsen

so ideal DR = 200 iso

jrista

ISO 100 is base these days, for everyone

no, hell no

ISO 100 tends to give you the best DR, regardless of brand

Michael Nielsen

I read many laces that 200 iso is hte best DR

jrista

ISO 100 is the highest gain

in the 1D X, gain is 6.7 e-/DU

at ISO 200, gain is half that

so 3.35e-/du

ISO 400 half again, so 1.675e-/du

ISO 800 is 0.8375e-/du

Michael Nielsen

1:32 AM

higher iso = higher gain

jrista

eh, no

Michael Nielsen

higher gain = less needed e / du

jrista

um, no

lets say we have a full well capacity of 100,000 electrons

Michael Nielsen

your cd player has a certain fixed output and when you turn up teh amp (= higher gain) you trandfer that same signal to higher output, ie. less db per output db

jrista

and a maximum digital signal level of 2^14, or 16384

your maximum digital signal as numeric units at low ISO is LESS than your maximum analog signal as electrons

so, you MUST use multiple electrons per DU at low ISO

in our case, 100000/16384

Michael Nielsen

1:34 AM

nt a signal that gives output 256 to give output 16000 so you need more gain

you want....

jrista

that would be 6.1e-/DU at ISO 100 for a 100,000 electron FWC

Now, ISO does not actually change the amount of light at the sensor...it can't. Changing ISO does only one thing: reduce the number of electrons that represent maximum analog signal.

So, at ISO 200, our maximum analog signal level is 50,000 electrons.

50000/16384 = 3.05e-/DU

in an image sensor, gain drops as ISO increases

Michael Nielsen

thats my ppoint, the light on teh sensro is fixed like the output of the cd player

ergo you need more gain

jrista

because you are allocating fewer and fewer electrons per pixel, and effetively "wasting" more and more of the pixels full capacity

Michael Nielsen

in indeustrial cameras gain is actually called gain

and is not in 1 stop increments

jrista

well, all I know, from reading a gazillion patents and technical papers and blogs

Michael Nielsen

1:37 AM

and if you increase the gain you get the same effect as increasing teh iso on the photo camera

jrista

is that "gain" as it is used in a CIS device is the ratio of electrons to digital units

Michael Nielsen

and on those cams teh default gain is higher than teh minimum gain

jrista

here, Roger Clark's (Ph.D. in this stuff) analysis of the 1D X: clarkvision.com/articles/evaluation-canon-1dx/index.html

Michael Nielsen

and the data sheets have the formula from teh "parameter" values to actual DB

dB

jrista

in the little table near the beginning, he has a column: Gain e/DN

Michael Nielsen

1:38 AM

and guess what: minimum gain = -3db

= -1 stop

jrista

well, that doesn't seem to apply to commercial grade DSLR sensors

ISO 100 -> 100000/16384

ISO 200 -> 50000/16384

ISO 400 -> 25000/16384

ISO 800 -> 12500/16384

Michael Nielsen

well, here on this site I read the reply about iso from one of the gurus near quote out of memory "actually iso 100 attanuate the singal by a stop effectively reducing DR, so you should be using 200 by default"

jrista

ISO 1600 -> 6250/16384

@MichaelNielsen That was Nikon manufactured sensors, and that was only the case over four years ago.

Since Nikon started using Exmor and sony tech in their own sensors, that is no longer the case.

Nikon has not used ISO 200 as base ISO for a long time.

Canon has always used ISO 100 as base ISO.

Michael Nielsen

and on sensorgen the data for 40D is equal for iso 100 and 200 showing that something "is up": sensorgen.info/CanonEOS_40D.html

jrista

yeah, read noise is "up"

from sensorgen:

ISO 100: Read noise 19.6e-, sat 40311e-, 11.0 stops

ISO 200: Read noise 11.7e-, sat 24668, 11.0 stops

read noise dropped at ISO 200, but so did maximum saturation

the ratio remained the same

thus...11 stops

Michael Nielsen

1:43 AM

ofc teh info could be outdated and at some point they changed the refernce iso to 100 , makign the "expanded ISO 50" be teh -3 db gain

jrista

Canon's CDS is over a decade old

Expanded ISO 50 doesn't change the power or gain

it is actually a shift in exposure and digital compensation

Michael Nielsen

bit shift?

industrial cams have that

jrista

ISO 50 is basically ISO 100 with some behind the scenes exposure trickery and a digital exposure shift to compensate

you LOSE DR at ISO 50 because of that

but the power levels and gain at the sensor are exactly teh same as ISO 100

every current commercial (consumer, prosumer, or pro) DSLR on the market these days uses ISO 100 as base ISO

some have an Expanded 50, some have Expanded ISO's above their maximum native

from ISO 100 through the maximum native ISO, the number of electrons per digital unit drops by a factor of two for each stop of ISO increase

so, fewer and fewer electrons are used to generate single digital units

at some point you have a crossover, where one electron is converted into one digital unit

that is "unity gain"

unity gain is usually somewhere between ISO 200 and ISO 800

it might be closer to a third stop setting

it is rarely ever an exact selectable ISO setting

I guess technically speaking, the per-pixel amplifiers are still putting out voltage

rather than DU...digital units only come into play at the ADC

so, assuming per pixel amplifiers in a CMOS sensor convert input voltage to output voltage

g = log(Vout/Vin)^2

I figure the formula would be:

if input voltage is 100,000 and output voltage is 50,000, gain would be -0.6dB

at ISO 200

Michael Nielsen

nope

that bells

deci is 10*

and its amplitude not enerry

jrista

yeah...that is log base 10, not log base 2

Michael Nielsen

1:52 AM

so its not power2

so its 3db

jrista

well, I don't know what the power input and output would be

pixels register charge

voltage

I'm sure there is a way to convert the voltage to power, but I don't have all the information to do so.

Michael Nielsen

Im pretty sure I looked it up recently if its 6db per stop or 3db

and found it was 3db

jrista

well, wouldn't that depend on what the full well capacity was?

I mean, all sensors are 14-bit these days

Michael Nielsen

I always get shaky about when it is 20-log vs 10*log

jrista

but not all have the same full well capacity in terms of charge

oh, I guess my formula above was wrong

it should be:

g = 10 log (Vout/Vin)^2

in 20 log:

g = 20 log (Vout/Vin)

that would mean 6dB

Michael Nielsen

1:57 AM

you can convince me that 100 iso is refernce, but you will never get me to see iso 100 is "higher gain" than 200 :)

jrista

no, in terms of dB:

Michael Nielsen

it is the power2 thaqt becomes 2* with log

jrista

ISO 100 = 0dB, ISO 200 = 6dB, ISO 400 = 12dB, ISO 800 = 18dB, etc.

Michael Nielsen

if you compare voltages is it amplitudes

ie no power2

jrista

hmm, I just looked up gain: en.wikipedia.org/wiki/Gain

Michael Nielsen

1:59 AM

if its P (watts) theres the power2 and it becomes 20*

jrista

Gain = 20 log (Vout/Vin) dB

Gain = 10 log (Vout/Vin)^2 dB

Transcript for