From DXOMark: More pixels offsets noise3

Started May 30, 2009 | Discussions
John Sheehy Forum Pro • Posts: 21,722
Re: From DXOMark: More pixels offsets noise3

DSPographer wrote:

John Sheehy wrote:

[snip]
I know exactly how they can avoid banding after capture, even if they
can't do it electronically. Banding is extremely simple. I don't
know what the H*ll is wrong with companies that leave banding in
their images - it almost seems like sabotage, to bank for future
upgrade carrots.

Really? I looked at the banding noise from my 5D2 and it was not as
simple to remove in software as I expected because it wasn't as
stable across the image as I had hoped. The techniques I can think of
to reduce would either not eliminate it completely or else they could
impact the signal details in deep shadows. Do you know of a technique
that would leave the signals untouched while completely removing the
banding?

Most of the banding noise energy is completely consistent across the frame in the RAW data. Conversion, and SNRs of different tonal levels may make it seem otherwise. Radio interference, when it occurs, will also be asynchronous to the lines.

Canon RAWs have 70 to 150 pixels to the left of each row of pixels which are masked from light; you can get a rough map of horizontal banding just from that.

Canon doesn't have many blind pixels on the vertical columns, though. They only have 13 such pixels on the 5D2; not a great sample.

If the sensor had, say, 180 blind pixels on both ends of each row of pixels, and 120 at both ends of each column, the banding could be very effectively removed by converters, even if the banding offsets were on a gradient, such as a line being +0.2 ADU too high on the left end, ramping to 0.4 ADU too low on the right end, due to electrical pulses or interference.

Figuring out the banding from the image itself can be a little tricky. Some images are easy; a low-contrast shot taken with extremely weak exposure has noise and banding stronger than signal, and simply shrinking the image to one pixel wide with downsampling, and then resizing it back creates a suitable mask. If you have a vertical strip of blackness or smooth, OOF solids, you can get the banding pattern from that as well.

Sometimes you can take even a non-linear image, and do local banding removal by using a high-pass filter on a copy of the image with perfectly horizontal or vertical mothion blur, as needed. You can paint any areas on the correction frame where you don't want correction in gray 128.

The best way to subtract banding, however, is to do it in a linear state, before conversion. You can not subtract it accurately from a converted image on a tone curve.

-- hide signature --

John

John Sheehy Forum Pro • Posts: 21,722
Re: Is DXOMark reliable? Looking at old cameras...

photo nuts wrote:

I used DXOMark to examine old cameras such as the Nikon D70, Canon
300D, Canon 350D etc and was shocked at the results.

Many of those so-called experts (some are fairly vocal in this
thread) often tout the superiority of OLD Canon sensors (e.g. 300D,
350D, 10D, 20D) against Nikon equivalents (D70, D100) at that time.
Mind you, they only use RAW images for comparison.

But the DXOMark tests do not seem to support those claims. In fact,
the performance of the Sony CCDs vs Canon CMOS sensors are nearly
identical for the most part!!!

It makes me wonder... who is lying? DXOMark? Those experts? RAW
Converters?

You're going to need to be much more specific than you were here, if you want answers.

The "Print" DR of the 20D at ISO 1600, for instance, should be far superior to that of the D70. That's the kind of stuff that the CMOS was doing at the time.

-- hide signature --

John

photo nuts Senior Member • Posts: 1,364
Sorry for my confusion and thanks for the clarification. /nt

no text here

photo nuts Senior Member • Posts: 1,364
Just did another check...

Sorry for the previous confusion...

This time, I compare the Canon 20D against the Nikon D70 and D200:

http://www.dxomark.com/index.php/eng/Image-Quality-Database/Compare-cameras/ (appareil1) 178|0 (appareil2) 197|0 (appareil3) 203|0 (onglet) 0 (brand) Canon (brand2) Nikon (brand3) Nikon

All across the board from SNR18%, dynamic range, tonal range to color sensitivity, the 20D is as good as, if not SIGNIFICANTLY better than the D200! The comparison is done in both pixel and image level.

Why then does the Nikon D200 sensor obtain a HIGHER DXOMark sensor score than the 20D???

Is DXOMark really being honest here?

John Sheehy Forum Pro • Posts: 21,722
Re: Just did another check...

photo nuts wrote:

Sorry for the previous confusion...

This time, I compare the Canon 20D against the Nikon D70 and D200:

http://www.dxomark.com/index.php/eng/Image-Quality-Database/Compare-cameras/ (appareil1) 178|0 (appareil2) 197|0 (appareil3) 203|0 (onglet) 0 (brand) Canon (brand2) Nikon (brand3) Nikon

All across the board from SNR18%, dynamic range, tonal range to color
sensitivity, the 20D is as good as, if not SIGNIFICANTLY better than
the D200! The comparison is done in both pixel and image level.

Why then does the Nikon D200 sensor obtain a HIGHER DXOMark sensor
score than the 20D???

Is DXOMark really being honest here?

Most people consider their composite scores of minimal value.

The biggest differences between sensors of the same size are in read noise; their composite scores don't weigh that in too heavily.

18% SNR at pushes to ISO 10,000, for example, might be more differentiating than the ISOs they do it at. 0.25% SNR at the normal ranges, too.

-- hide signature --

John

OP berleconi Contributing Member • Posts: 720
Re: A hint..

what was that article .
comment from bob2

started me on the path to destruction. The combination of events, the NYT article where PA posed as an expert on all of our behalves, the bogus downscaling article (with no comments

Also is dpr on one side of the argument.

Berl.

Steen Bay Veteran Member • Posts: 7,418
Re: A hint..

berleconi wrote:

what was that article .
comment from bob2
started me on the path to destruction. The combination of events, the
NYT article where PA posed as an expert on all of our behalves, the
bogus downscaling article (with no comments

Also is dpr on one side of the argument.

Berl.

I never saw the NYT article, but think that Phil was talking against the 'MP-race' in it, and that seems to be (or have been?) the 'officiel' DPR stance on that issue, like for exampel in the 50D review, then comparing 40D vs. 50D.

Iseewhatyoudidthere New Member • Posts: 2
This is where Phil settled the question

Steen Bay wrote:

berleconi wrote:

what was that article .
comment from bob2
started me on the path to destruction. The combination of events, the
NYT article where PA posed as an expert on all of our behalves, the
bogus downscaling article (with no comments

Also is dpr on one side of the argument.

Berl.

I never saw the NYT article, but think that Phil was talking against
the 'MP-race' in it, and that seems to be (or have been?) the
'officiel' DPR stance on that issue, like for exampel in the 50D
review, then comparing 40D vs. 50D.

http://www.nytimes.com/2008/11/13/technology/personaltech/13basics.html?fta=y

DSPographer Senior Member • Posts: 2,464
That is what I have seen also

I did see that if I take a shot with the lens cap on that I can make a correction using just the sum of the rows and the sum of the columns that removes the visible pattern noise. The problem I had was in the amount of masked pixels as you mentioned. I had hoped that the vertical noise followed a repeating pattern so that a correction could be formed by something like aggregating every fourth masked column value but the column data varies too irregularly from column to column for that.

 DSPographer's gear list:DSPographer's gear list
Canon PowerShot G7 X Canon EOS 5D Mark II Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
ejmartin Veteran Member • Posts: 6,274
Re: That is what I have seen also

DSPographer wrote:

I did see that if I take a shot with the lens cap on that I can make
a correction using just the sum of the rows and the sum of the
columns that removes the visible pattern noise. The problem I had was
in the amount of masked pixels as you mentioned. I had hoped that the
vertical noise followed a repeating pattern so that a correction
could be formed by something like aggregating every fourth masked
column value but the column data varies too irregularly from column
to column for that.
--

What is the degree of correlation of the average of the masked pixels in a given column/row with the average of the active pixels in that same column/row? If it's over 50% won't you do better by subtracting the masked pixel average from the active pixel data?

-- hide signature --
Daniel Browning Senior Member • Posts: 1,058
Myth busted: performance of small pixels

It is possible for small pixel sensors to have worse performance per pixel, but the same performance when actually displayed or used for the same purpose as a large pixel sensor. This fact may be unbelievable or at least counter-intuitive to many people who work with digital images, but I believe that is only because of the following five types of mistakes that are frequently made in image analysis:

  • Unequal spatial frequencies

  • Unequal sensor sizes.

  • Unequal processing.

  • Unequal expectations.

  • Unequal technology.

This thread is getting pretty full. I posted the rest of my thoughts on the matter in another thread. Six consecutive posts starting with this one:

http://forums.dpreview.com/forums/read.asp?forum=1018&message=32064270

Hope that helps.
--
Daniel

OP berleconi Contributing Member • Posts: 720
Re: From DXOMark: More pixels offsets noise3

Thanks.

Stan Prevost Contributing Member • Posts: 650
Re: Diffraction cutoff frequency

DSPographer wrote:

Not sure what you mean by "much coarser", it is 22 percent coarser, 1.00 vs 0.82 of the Airy radius, assuming that you are using the Rayleigh Criterion as the commonly cited diffraction blur (24.2% contrast).

Not a big deal, just that "much coarser" sounds ominous, like factors of two or something.

The rest of the numbers are good (to the precision implied). And certainly good for the purpose.

Stan

DSPographer Senior Member • Posts: 2,464
Re: Diffraction cutoff frequency

Stan Prevost wrote:

DSPographer wrote:

Not sure what you mean by "much coarser", it is 22 percent coarser,
1.00 vs 0.82 of the Airy radius, assuming that you are using the
Rayleigh Criterion as the commonly cited diffraction blur (24.2%
contrast).

No, the Nyquist spacing is two pixels per cycle at the diffraction cutoff which is 0.41 of the Rayleigh criteria.

Not a big deal, just that "much coarser" sounds ominous, like factors
of two or something.

The rest of the numbers are good (to the precision implied). And
certainly good for the purpose.

Stan

 DSPographer's gear list:DSPographer's gear list
Canon PowerShot G7 X Canon EOS 5D Mark II Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
Stan Prevost Contributing Member • Posts: 650
Re: Diffraction cutoff frequency

DSPographer wrote:

Stan Prevost wrote:

Not sure what you mean by "much coarser", it is 22 percent coarser,
1.00 vs 0.82 of the Airy radius, assuming that you are using the
Rayleigh Criterion as the commonly cited diffraction blur (24.2%
contrast).

No, the Nyquist spacing is two pixels per cycle at the diffraction
cutoff which is 0.41 of the Rayleigh criteria.

Yes, but that is not what I said. At that point, I was talking about the diffraction blur, not how we sample it, as is clear from the last phrase in the paragraph.

I agree with two pixels per resolution cycle, as was clear from my agreement with your pixel spacing numbers.

Anyway, we seem to agree on this, my only point was to perhaps moderate the "much coarser" characterization, and to verify what you were using as a reference as the commonly cited diffraction limit.

Thanks,

Stan

DSPographer Senior Member • Posts: 2,464
Re: That is what I have seen also

ejmartin wrote:

What is the degree of correlation of the average of the masked pixels
in a given column/row with the average of the active pixels in that
same column/row? If it's over 50% won't you do better by subtracting
the masked pixel average from the active pixel data?

It is right about 50%. The first masked row has ramps rather than black level and the second row is also contaminated. That leaves about 50 masked rows but the masked rows don't all have the same mean value as the dark frame image: only the last 9 rows match the image black level. Using only those last nine rows the correlation coefficient between the masked and unmasked means is about 0.46 but using masked rows 3 through 52 raises the correlation to about a 0.54 coefficient. Unfortunately this is not really enough to make a visible difference in the pattern noise. For comparison the correlation between the mean of the rows in the top half of the unmasked black image and the mean in the bottom half is about a 0.97 coefficient.

 DSPographer's gear list:DSPographer's gear list
Canon PowerShot G7 X Canon EOS 5D Mark II Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
ejmartin Veteran Member • Posts: 6,274
Re: That is what I have seen also

DSPographer wrote:

ejmartin wrote:

What is the degree of correlation of the average of the masked pixels
in a given column/row with the average of the active pixels in that
same column/row? If it's over 50% won't you do better by subtracting
the masked pixel average from the active pixel data?

It is right about 50%. The first masked row has ramps rather than
black level and the second row is also contaminated. That leaves
about 50 masked rows but the masked rows don't all have the same mean
value as the dark frame image: only the last 9 rows match the image
black level. Using only those last nine rows the correlation
coefficient between the masked and unmasked means is about 0.46 but
using masked rows 3 through 52 raises the correlation to about a 0.54
coefficient. Unfortunately this is not really enough to make a
visible difference in the pattern noise. For comparison the
correlation between the mean of the rows in the top half of the
unmasked black image and the mean in the bottom half is about a 0.97
coefficient.

Thanks. BTW, what software are you using to get at the masked pixels?

-- hide signature --
John Sheehy Forum Pro • Posts: 21,722
Re: That is what I have seen also

DSPographer wrote:

For comparison the
correlation between the mean of the rows in the top half of the
unmasked black image and the mean in the bottom half is about a 0.97
coefficient.

I've been making that point for a long time; this banding noise in Canons is EXTREMELY simple. There is no excuse for its existence.

-- hide signature --

John

John Sheehy Forum Pro • Posts: 21,722
Re: That is what I have seen also

ejmartin wrote:

Thanks. BTW, what software are you using to get at the masked pixels?

If you convert to uncompressed DNG, the entire RAW image sits after the header, and you can load it into programs that load 'raw" images with supplied dimensions and header offsets.

Ilyah Borg's raw library package has an example executable that spits out the masked borders with an option.

-- hide signature --

John

DSPographer Senior Member • Posts: 2,464
Re: That is what I have seen also

John Sheehy wrote:

ejmartin wrote:

Thanks. BTW, what software are you using to get at the masked pixels?

If you convert to uncompressed DNG, the entire RAW image sits after
the header, and you can load it into programs that load 'raw" images
with supplied dimensions and header offsets.

That is what I did. I just read the last portion of the uncompressed DNG as int16 values with Matlab for analysis. I also saved the result as a png in case anyone would like a copy.

 DSPographer's gear list:DSPographer's gear list
Canon PowerShot G7 X Canon EOS 5D Mark II Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
Keyboard shortcuts:
FForum MMy threads