LX3 sensor analysis

[Jeff Charles]

...The raw converter that understands that HTP was used,
simply boosts the output by a stop (apart from rolling off the
highlights) to get the image brightness back to what the metering
indicated at the time of exposure for ISO 200.

Raw Therapee, which would not know that it needs to boost exposure for LX3 raws, produces output that is about 1/4 stop less bright on the white patch than does SilkyPix. That's somewhat less than the difference that DXO reports. I don't know if this is significant one way of the other.

...
But the point is that you can't go by converted raw output to
determine the actual sensitivity of the sensor.
...

Understood. It is just one data point, but it is consistent with your findings and with DIWA's.

In the final analysis, though, the performance of the camera in actual use matters more than the sensor sensitivity. It's pretty clear that if you use the same ISO, f-stop, and shutter speed as the G10 or most other cameras, you will get a roughly equivalent exposure with the LX3. So, for practical purposes, the LX3's nominal ISOs are fairly accurate.

If DXO's results are correct, and in fact the jpeg and raw converters are boosting exposure behind the scenes, it does not seem to have had a detrimental effect on image quality.

Jeff

 

[xvimbi]

A larger photosite area captures more total photons, because it has a
larger total area. But brightness is measured by the flux of
photons per unit area; the larger number of photons divided by the
larger area of a large photosite will see the same flux as a smaller
number of photons recorded, divided by a smaller photosite area.

What a larger sensor allows by capturing more total photons is a
higher signal/noise ratio in the image data, which is a major reason
for the higher image quality of cameras with large sensors.

So, two sensors with the same size, but one of them having only half the total photosite area than the other sensor will produce same-size images with the same 'brightness'???

Is talking about 'brightness' actually meaningful when analyzing CCD sensitivities?

Best - MM

 

[ejmartin]

So, two sensors with the same size, but one of them having only half
the total photosite area than the other sensor will produce same-size
images with the same 'brightness'???

All other things being equal, yes.

Is talking about 'brightness' actually meaningful when analyzing CCD
sensitivities?

Well, "brightness" is relative to a calibration standard. The ISO standard is the accepted norm for determining "brightness" -- middle gray registers at a particular fraction of saturation for a particular flux of photons. So yes if the standard is adhered to there is a well-defined notion of brightness.

The problem is that camera makers play fast and loose with the definitions. But in the end of the day it doesn't matter so much -- it doesn't matter so much where a particular photon flux lies in the range of the digitized data (within reason). The important quantity is the signal/noise of the captured image data, and that depends on having clean, non-noisy electronics and on capturing lots of photons. And the calibrated brightness is really only important for raw converter makers to know what a particular raw value should correspond to in terms of brightness of the rendered image.

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[xvimbi]

So, two sensors with the same size, but one of them having only half
the total photosite area than the other sensor will produce same-size
images with the same 'brightness'???

All other things being equal, yes.

Maybe I am just dense, or I don't get what 'brightness' is really referring to...

I can see that each individual pixel, no matter what size, produces roughly the same brightness, but the images from the two sensors we are talking about must have different overall brightness. After all, one of the images has a lot more black space between the pixels than the other.

What am I missing?

Thanks. Best - MM

 

[tartamillo]

I can see that each individual pixel, no matter what size, produces
roughly the same brightness, but the images from the two sensors we
are talking about must have different overall brightness. After all,
one of the images has a lot more black space between the pixels than
the other.

What am I missing?

What black space?

--
Maki

 

[Leo]

I have read the above paragraph and need your help again but now to

..........

(So you get 4095-16) / 2.55 = 1600
And 2^10.6 = 1600, so DR is 10.6 stops.

........

Just my two oere
Erik from Sweden

Erik,
What is the meaning for 2.55?

(So you get 4095-16) / 2.55 = 1600

Leo

 

[Ehrik]

(So you get 4095-16) / 2.55 = 1600
And 2^10.6 = 1600, so DR is 10.6 stops.

Erik,
What is the meaning for 2.55?

(So you get 4095-16) / 2.55 = 1600

It's what the standard deviation of the noise in an unexposed shot (blackframe, shot with lens cap on) would have been if the black end of the raw data hadn't been clipped*. Emil has reconstructed this value from noise levels of consecutively more exposed areas.

It represents the "strength" of the noise, the strength that the signal has to compete with to be visible in deep shadows when we try to brighten them. The stronger the noise, the more signal we need to overcome it, and thus the less dynamic range is available to us.

(* With cameras from Canon and Oly and some Sonys, you can just measure the std.dev. straight off since there is no clipping.)

Just my two oere
Erik from Sweden

 

[ejmartin]

I have read the above paragraph and need your help again but now to

..........

(So you get 4095-16) / 2.55 = 1600
And 2^10.6 = 1600, so DR is 10.6 stops.

........

Just my two oere
Erik from Sweden

Erik,
What is the meaning for 2.55?

(So you get 4095-16) / 2.55 = 1600

Leo

2.55 is the read noise in raw levels; 4095 is the raw saturation level, and 16 is the black point offset. So the maximum signal above zero is 4095-16, while the lowest detectable signal must be above the level of the read noise of 2.55 to be detectable. Thus the dynamic range between the maximum possible signal and the minimum detectable signal is (4095-16) 2.55. As mentioned, this is a quite liberal definition of DR; one may wish to set a more stringent lower bound, for instance that the signal be four times the read noise. This would yield a lower DR for higher minimum quality.

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[Pierre Sottas]

interesting analysis and very impressive result.

Does someone know the pixel architecture for that sensor? Thanks.

Pierre

 

[John Sheehy]

2.55 is the read noise in raw levels; 4095 is the raw saturation
level, and 16 is the black point offset.

Are you sure about the bottom level?

The FZ50 data is 16-bit, and has 16 as the lowest value, but that is actually "1" in 12-bit. Could you be mixing the 16-bit minimum and the 12-bit maximum?

Anyway, this value of 2.55 - This is the actual ADUs in 12-bit, with ISO 80 metered gray only 2.5 stops below green saturation? That indeed is very low. That means that in DSLR-speak, this is the read noise of ISO 160, quite a bit better than many DSLRs.

Does the read noise exhibit the short-streak banding like that in my FZ50 pushed examples, or any other kind of banding?

Other than the short streaks in my FZ50, the sensor is near-perfect. There are no very low frequency fixed pattern noises like those in my G9 or even in my 50D (only relevant at ISO pushes into the hundreds of thousands or millions).

--
John

 

[John Sheehy]

(* With cameras from Canon and Oly and some Sonys, you can just
measure the std.dev. straight off since there is no clipping.)

If I get an LX3, maybe I'll put some weak LEDs inside of it to raise the black level above clipping

--
John

 

[John Sheehy]

16 to 4095 is just labelling; you could call it 0 to 4079 so it's not
the ratio of those numbers we are interested in. It's the noise
rather than the signal resolution that sets the limit. The DR
definition Emil works with (the engineering definition) is the ratio
between the saturation level and the read noise. (Where signal and
noise are equally strong.)

Actually, that isn't exactly true. It's a carry-over from terminolgy where nothing like shot noise exists. Signal and noise are equally strong (statistically) a little bit above the read noise "floor".

--
John

 

[John Sheehy]

Very cool results emil. How did Pany make such a leap in sensor
capability? From darn poor noise with the LX2, to damn good with the
LX3?

Well, the Olympuses should be seeing this too, as I believe they use the same technology, just bigger.

Anyway, somebody PLEASE do an LX3 vs D3 or 1D3 comparison in the style of my FZ50 vs XTi and 10D comparisons. That should be very interesting.

--
John

 

[John Sheehy]

Interesting read. Thanks for you input on this.
Looks like the highlight tone priority mode on some dslrs. There's
been many complaints by users of not only noise in p&s cameras, but
also dynamic range. This appears to be Panny's approach to giving
some highlight room and seems to work quite well for what the camera
is.

Well, for starters, most people complain about DR when the real problem they are complaining about is highlight headroom. DR and highlight headroom are two completely different things. DR also includes the "footroom" or the usability of shadows. Most compact cameras destroy the shadow areas when creating the JPEGs. I can take a G9 image, exposed for ISO 6400 while set to ISO 80, and get a web-sized image that hardly shows any noise, doing all the steps manually from the RAW. From the camera's JPEG, it is pretty hopeless, as the shadows are completely desaturated, and void of detail.

--
John

 

[Ehrik]

Here are replies to some of John's posts.

The FZ50 data is 16-bit, and has 16 as the lowest value, but that is
actually "1" in 12-bit. Could you be mixing the 16-bit minimum and
the 12-bit maximum?

The LX3 uses Panasonic's new RW2 raw format that is more compact
(with improved shot-to-shot times). I don't know details, but it's probably
not storing 16 bits per pixel any more.

Does the read noise exhibit the short-streak banding like that in my
FZ50 pushed examples, or any other kind of banding?

I asked Emil about that and got this reply:

The LX2 exhibited ugly horizontal streaks when the shadows were pushed.
Did you have any chance to check for pattern noise?

Yes I did have a look, it is extremely well controlled. Essentially none.

John, about DR definition:

Actually, that isn't exactly true. It's a carry-over from terminolgy where
nothing like shot noise exists. Signal and noise are equally strong
(statistically) a little bit above the read noise "floor".

Ah, yes, there's a wee bit of shot noise too.

John, on the sensor improvements:

Well, the Olympuses should be seeing this too, as I believe they use the same technology, just bigger.

The DSLRs are not using Panasonic CCD chips but NMOS chips. That architecture doesn't seem to control read noise (1D and 2D) nearly as well as the LX3, unfortunately.

It's quite interesting, actually, that the 10Mp NMOS cameras have a peculiar "knee" in the DR performance, with more loss of DR going from ISO 100 to 200, than from 200 to 400.
See here:

http://www.pages.drexel.edu/~par24/rawhistogram/OlympusE510Test/OlympusE510Test.html
And it looked the same at dxomark's curves for the E-510, L10 and E-3.

Anyway, somebody PLEASE do an LX3 vs D3 or 1D3 comparison in the style of my FZ50 vs XTi and 10D comparisons. That should be very interesting.

Yes indeed, I love those. But I wonder if the lenses' lp/mm in those tests are favouring the small pixels? If one did a simulation starting with a clean image that was resampled and then noise added according to measured values, would that lead to a somewhat different conclusion do you think?

Just my two oere
Erik from Sweden

 

[John Sheehy]

Yes indeed, I love those. But I wonder if the lenses' lp/mm in those
tests are favouring the small pixels? If one did a simulation
starting with a clean image that was resampled and then noise added
according to measured values, would that lead to a somewhat different
conclusion do you think?

Well, I've thought of doing that in the past, but the thing is, so many people are resistant to proofs from actual cameras, that it's hard to imagine that they'd be receptive to a simulation.

Sometimes I wonder how many of the people who saw my empire state building at ISO 13,000 understand its significance. Many of the reactions suggest that this has nothing to do with anything "real", when in fact it is about as real as you can get; a clear demonstration that with current pixel technology, tiny pixels outperform big pixels in read noise and DR at base ISOs.

--
John

 

[Leo]

Emil,

Thank you. I have got it. Few days ago I have purchased a used E-410... so small and light but noisy at ISO 800 and at ISO 400 the shades if pulled up are are also noisy. However, it is pleasure to use this camera. Compared shape, weight and size with my two film SLR cameras I still have - so close - like twins

In the past I have use a film camera with 50 mm lens and no higher than ISO 200. The E-410 is as film except more often (very often) blown highlights. However, every camera (film or digital) will blow highlights with different starting point
Leo

 

[Leo]

Well, for starters, most people complain about DR when the real
problem they are complaining about is highlight headroom. DR and
highlight headroom are two completely different things. DR also
includes the "footroom" or the usability of shadows. Most compact
cameras destroy the shadow areas when creating the JPEGs. I can take
a G9 image, exposed for ISO 6400 while set to ISO 80, and get a
web-sized image that hardly shows any noise, doing all the steps
manually from the RAW. From the camera's JPEG, it is pretty
hopeless, as the shadows are completely desaturated, and void of
detail.

--
John

John,

You underexposing over 6 EV leaving very little for the image. Let's say G9 RAW DR is 10 (which is too high) then you squeezing the whole image into 4 EV. How it can work. In addition to reduce the noise the goal is to keep the image.
Leo

 

[bstring]

Sometimes I wonder how many of the people who saw my empire state
building at ISO 13,000 understand its significance.

I searched, but could not find this. Do you have a link?

Brian

--
Phanfare supporter-

 

[Ehrik]

Brian, here you are:

http://forums.dpreview.com/forums/read.asp?forum=1018&message=28607494

http://forums.dpreview.com/forums/read.asp?forum=1018&message=28760503

Just my two oere
Erik from Sweden