Debayering a DSLR's Bayer matrix.

[GuillermoBarrancos]

Hi

The increase with microlenses is actually 15% at 550nm and 10% at 650nm in the graph.

Alistair

That is for this example CCD sensor. In this case the KAF8300.

It seems to be impossible tho to find similar graphs for the Canon CMOS sensors. So then it´s all speculation in how much effect the microlensing has across the spektrum.

The only graph I could find is this one from the 40D (which peak QE unmodded seems to be similar to 1100D (36%):

So following the trend of the average 10% drop by removing lensing at the 650nm, the effective QE would end up at around 20% in Ha.

Fascinating stuff!  Really brushing up my Google search skills lately with all these interesting discussions.

Edited April 29, 2014 by GuillermoBarrancos

[russellhq]

That is for this example CCD sensor. In this case the KAF8300.

It seems to be impossible tho to find similar graphs for the Canon CMOS sensors. So then it´s all speculation in how much effect the microlensing has across the spektrum.

I could try taking a flat with the SII filter. This would guarantee that only red photos are hitting the sensor and I could then make a direct comparison between the bayer matrix side and debayered side and not have to worry about double counting photons (as from the charts, it looks like the RGB filters overlap making it difficult to draw a comparison)

[Gina]

From the graphs the KAF8300 mono is actually more sensitive at Ha and SII wavelengths and the same at OIII   But as has been said this is a CCD sensor and not CMOS - whether they're the same or different in this respect I have no idea.

Edited April 29, 2014 by Gina

[GuillermoBarrancos]

From the graphs the KAF8300 mono is actually more sensitive at Ha and SII wavelengths and the same at OIII But as has been said this is a CCD sensor and not CMOS - whether they're the same or different in this respect I have no idea.

You have to look at the blue line with the crosses. That is the monochrome without micro lensing.

Which is more or less equal to the RGB version with micro lensing.

An important clue me thinks as to the effect of DeBayering.

Edited April 29, 2014 by GuillermoBarrancos

[Gina]

I think I've pretty much settled on a 4 position FW for my debayered 1100D with Ha, OIII, SII and L filters.  The L is useful both as high res luminance and for setting up.  I could use an 1100D with untouched CFA for RGB simultaneous imaging.  This would almost make up for the smaller, less sensitive pixels compared with mono separate RGB subs binned 2x2 if this were possible with a DSLR (which it isn't AFAIK).  This is a cheaper option than a single 1100D with 7 filters as I won't need to buy RGB filters.

[Gina]

You have to look at the blue line with the crosses. That is the monochrome without micro lensing.

Which is more or less equal to the RGB version with micro lensing.

An important clue me thinks as to the effect of DeBayering.

Yes - this is what I meant

[tommy_h]

Hi

Look what Central DS says to the sensivity of a monofied DSLR: http://www.centralds.net/cam/?p=6562

This is also my subjective opinion.

I have monofied a 500D with a wooden scraper and I have also noticed a dull gold layer below the bayer-matix and removed this layer easily with polishing paste. 

Then a glossy gold layer gets visible, which is hard to remove. Below there is a reddish layer (maybe the photodiodes?)

I have lost one sensor by polishing to much in the edges and removing the glossy gold layer (the sensor is still working, but the flat is really uneven).

Regards

Thomas

[russellhq]

Hi

Look what Central DS says to the sensivity of a monofied DSLR: http://www.centralds.net/cam/?p=6562

This is also my subjective opinion.

I have monofied a 500D with a wooden scraper and I have also noticed a dull gold layer below the bayer-matix and removed this layer easily with polishing paste. 

Then a glossy gold layer gets visible, which is hard to remove. Below there is a reddish layer (maybe the photodiodes?)

I have lost one sensor by polishing to much in the edges and removing the glossy gold layer (the sensor is still working, but the flat is really uneven).

Regards

Thomas

1. Green pixel – Before:  14008, > After: 13363  ……….    4.6% down

2. Red pixel    - Before:  7966 ,  > After: 13363  ……….   67.8% up

3. Blue pixel   – Before:  9381 ,  > After: 13363  ……….   42.4% up

With this modification we lost the micro lens layer and this is so serious damage in light gathering power. But the CFA removal gave us higher resolutions and increase of sensitivity in Red, Blue pixels area as Green pixels.

I'm not sure the conclusion is too clear. The mono sensor appears to be more sensitive to red, but this is only because it's also receiving green and blue light. If you exposed the sensors to just red light, you would find the microlenses are still more sensitive.

[sharkmelley]

Look what Central DS says to the sensivity of a monofied DSLR: http://www.centralds.net/cam/?p=6562

This is also my subjective opinion.

This seems to confirm that losing the microlenses has a very serious impact.  Using their figures, the R,G,B pixels have values 14008, 7966 and 9381 respectively.  The sum of these is 31355 and this is roughly the value we would hope to see in the debayered pixels.  But instead we see 13363 which is a factor of 2.3x lower than expected.  So removing the microlenses (and maybe some additional "unknown" effect) is reducing the sensitivty of the debayered sensor by a factor of 2.3

I could try taking a flat with the SII filter. This would guarantee that only red photos are hitting the sensor and I could then make a direct comparison between the bayer matrix side and debayered side and not have to worry about double counting photons (as from the charts, it looks like the RGB filters overlap making it difficult to draw a comparison)

That would be a very useful test.  Try it with each NB filter you can!

Mark

[alistairsam]

I'm not sure the conclusion is too clear. The mono sensor appears to be more sensitive to red, but this is only because it's also receiving green and blue light. If you exposed the sensors to just red light, you would find the microlenses are still more sensitive.

Thats a nice write up.

Here they talk about their experience with different methods and finally settling on the infamous wooden tool technique which is my personal preference.http://www.centralds.net/cam/?p=6435

I think we're focusing a bit too much on the loss of microlenses.

The logic is simple, the major benefit with a mono compared to an osc is the detail provided by the luminance framd where the whole mono sensor area is used to capture all the light that falls on the surface.

Rgb data captured with full size rgb filters don't contribute to detail which is why its usually binned 2x2 and even blurred in some cases.

So the details in a galaxy for example would be more defined in a mono dslr with just a lum filter or an LP filter.

Combine that with rgb from a filter modded dslr and you're achieving results close to higher qe ccd's at a fraction of the cost which is the essence of this mod.

Qe is very important and dslr's are usually between 25 and 35%.

So longer exposures with cooled mono dslr's will easily achieve a respectable snr even without microlenses.

The 8300 ccd has a rather low qe of around 50% compared to the sony ccd's that are close to 70%. Yet there are some remarkable results with the 8300 sensor with optimal exposure and processing.

Just my 2c..

Alistair

[Gina]

This seems to confirm that losing the microlenses has a very serious impact.  Using their figures, the R,G,B pixels have values 14008, 7966 and 9381 respectively.  The sum of these is 31355 and this is roughly the value we would hope to see in the debayered pixels.  But instead we see 13363 which is a factor of 2.3x lower than expected.  So removing the microlenses (and maybe some additional "unknown" effect) is reducing the sensitivty of the debayered sensor by a factor of 2.3

That would be a very useful test.  Try it with each NB filter you can!

Mark

I may be confused but surely you must compare pixel with pixel not the sum of 3 pixels with the value of 1.  If we average over the quad, the sum of the OSC version is R + 2xG + B = 38321.  Now lets add the mono pixels = 4 x 13363 = 53452.  So comparing like with like, before debayering the sum of 4 is 38321 and after debayering it is 53452.  This shows an increase in sensitivity of 1.36 times.

Of course if we could somehow retain or replace the micro lenses the benefit of debayering would be even greater.  Apart that is from the doubling of resolution in both axes.

Edited April 29, 2014 by Gina

[GuillermoBarrancos]

It was what I feared and why I have been skeptical about the whole debayering. I think we all have been ...how to say.... maybe too enthousiastic too quickly perhaps? Jumping on the hype after seeing some pretty mono Pictures from the first debayered cameras.

I think the effect of micro lensing has been (most likely accidently) overlooked and impact it has on QE / sensitivity underestimated.

Tho I don't think we immediately have to completely dismiss debayering.  If you are after serious Narrowband imaging with a DSLR, then debayering might still be worthwhile risk to take, as you will be able to illuminate all pixels during exposures, instead of just 25%. Getting increased resolution with narrowband filters.

But I Guess it's pretty safe to say that if you are just after extra QE / sensitivity, then debayering is a Complete waste of time and not worth the risk!

[JamesF]

I may be confused but surely you must compare pixel with pixel not the sum of 3 pixels with the value of 1.  If we average over the quad, the sum of the OSC version is R + 2xG + B = 38321.  Now lets add the mono pixels = 4 x 13363 = 53452.  So comparing like with like, before debayering the sum of 4 is 38321 and after debayering it is 53452.  This shows an increase in sensitivity of 1.36 times.

I think what Mark is doing is comparing the achieved sensitivity (without the lenses and CFA) with the extrapolated optimal sensitivity based on the readings with the CFA and lenses.  For example, if the RGB values for a given photosite total 31355 then ideally with no CFA the same photosite should record 31335.  Because the microlenses are no longer present then the measured value is considerably lower.

Gina on the other hand is comparing the actual signal values measured for (a group of four, in this instance) photosites before and after removal of the microlenses and CFA.  In this case the readings are better "after" compared with "before".

But they're not the same thing, so they can't be compared directly.  Both are probably valid ways to look at the data.  The difference is probably that with the microlenses and CFA present the resolution is lower even if more photons contribute to the readings for each photosite.

(Obviously there's some approximation going on with the CFA example too, because at a red photosite the G and B values never actually happened.  They're interpolated from the values recorded at the surrounding pixels.)

James

[Gina]

As I said above, if you want to use the camera for RGB only there would be only a small benefit i debayering and not worth the effort.  But if you have the actual resolution from the sky and optics you are using then using a debayered camera for luminance will double the resolution in both axes and if this is used to control intensity in the RGB colour image obtained from an (unmodified) OSC camera then the overal perceived resolution is increased.

Narrow band is a totally different matter.  The sensitivity of the red pixels is increased for Ha by removing micro lens and red CFA filter.  The increase is even more for SII and for OIII it is increased slightly due do the wavelength falling on the sides of the G and B spectral responses.  Having two green pixels per quad halps with OIII in the OSC version but by my reckoning this doesn't quite make up for the attenuation from the colour filters.   The improvement in sensitivity for Ha and SII is considerable - being around 2x.  Then on top of this we have four times as many pixels receiving photons giving the aforementioned increase in resolution.  If binning were possible with a DSLR we would gain 3 stops for the same resolution by binning 2x2 over the OSC version.

[GuillermoBarrancos]

 The sensitivity of the red pixels is increased for Ha by removing micro lens and red CFA filter. 

How can sensitivity in red pixels for Ha increase by removing micro lens and CFA filter, when sensor QE graphs show the complete opposite? I don't follow.

Pretty much all sensor QE graphs I have seen show that a Mono sensor without micro lens has the exact same QE in the red band as an OSC sensor with bayer Matrix. Some actually show even lower sensitivity due to absense of micro lensing!

Below graph of the KAI 11002 is such an example. This sensor (both OSC and Mono have micro lensing) shows no difference in QE in red band between OSC and Mono Version:

As Canon doesn't produce Mono CMOS sensors of the Digic series, we don't know if a Mono will have a higher QE than a OSC at all and that it might have the same graph as above example of the KAI 11002 sensor. 

Which would mean removing the micro lensing might have an even more profound impact on QE in the red band with debayering.

You end up with a much lower sensitivity. The only gain again being an increased resolution with narrowband.

Edited April 29, 2014 by GuillermoBarrancos

[Gina]

I was going by figures and graphs posted above but once I get a part debayered 1100D sensor I shall be able to do some tests.  Then we shall see   Someone else could well beat me to it though

[russellhq]

I was going by figures and graphs posted above but once I get a part debayered 1100D sensor I shall be able to do some tests.  Then we shall see   Someone else could well beat me to it though

Indeed

Here are the results from my tests using Ha, OIII and SII filters. You can see that from my sensor, removing the microlenses roughly halved the sensitivity of the photosites. But you can also see that, since all photosites on the debayered side are sensitive to the filtered light, the sensitivity of the sensor improves by 75-110% for SII & Ha (ignoring signal recorded in the B & G regions). But the results from the OIII are a bit more interesting. This showed no improvement in the sensitivity of the sensor after debayering (ignoring signal recorded in the B & R regions). Also, if you could use the signal from the B pixels, the debayering process actually makes the sensor only 75% as sensitive as the original with microlenses.

Here's the table of results:

Here are the RAW files if anyone is interested

https://www.dropbox.com/s/l0ctwa7ve17zqky/IMG_5508_Ha_7nm.CR2

https://www.dropbox.com/s/pdv66522z911rxc/IMG_5523_SII_8nm.CR2

https://www.dropbox.com/s/vgj7t4uz57w98oq/IMG_5510_OIII_8.5nm.CR2

[Thalestris24]

Hi

When I (eventually!) get the debayered 1100d I ordered ages ago... I'll be able to do a direct comparison with my standard, unmodded 1100d. I'll probably just do some shots through my canon 75-300mm lens, initially. At the moment I'm not sure how best to process the mono raw files but I'll figure something out

The mono 1100d has been ordered from JTW but they are a bit slow...

Fwiw, during discussions they said:

"The debayered sensor loses the microlenses but gains resolution and sensitivity. There is a bit loss of about 30% from the microlens removal but an even bigger gain, just under 1 stop is about average."

The proof of the pudding, as they say. 

I'm not expecting the mono 1100d to then miraculously become a great camera - it will still be uncooled and noisy. It will still need dark and flat calibration frames. There are bound to be imperfections from the debayering - flats will be essential to help cover them up, and for vignetting, of course. But, hey, it'll still be 4272 x 2848   That alone is pretty cool

Cheers

Louise
 

[Thalestris24]

Indeed

Here are the results from my tests using Ha, OIII and SII filters. You can see that from my sensor, removing the microlenses roughly halved the sensitivity of the photosites. But you can also see that, since all photosites on the debayered side are sensitive to the filtered light, the sensitivity of the sensor improves by 75-110% for SII & Ha (ignoring signal recorded in the B & G regions). But the results from the OIII are a bit more interesting. This showed no improvement in the sensitivity of the sensor after debayering (ignoring signal recorded in the B & R regions). Also, if you could use the signal from the B pixels, the debayering process actually makes the sensor only 75% as sensitive as the original with microlenses.

Here's the table of results:

DebayerResults.png

Hi

If I'm reading that right, when using the original sensor with say light via Sii, which is just a narrow band of red light, then the responses from G, G2 and B pixels are really colour 'noise' since ideally you'd want the G, G2 and B pixels to be 0?

Cheers

Louise

[russellhq]

Hi

If I'm reading that right, when using the original sensor with say light via Sii, which is just a narrow band of red light, then the responses from G, G2 and B pixels are really colour 'noise' since ideally you'd want the G, G2 and B pixels to be 0?

Cheers

Louise

Sort of. If you look at QE charts of OSC sensors, you'll see the R, G & B pixels overlap each other. Thus, the G and B pixels are still capturing some of the Red SII light. It's not like mono sensors with RGB filters where you get a sharp cut off between the colour bands.

[Thalestris24]

Sort of. If you look at QE charts of OSC sensors, you'll see the R, G & B pixels overlap each other. Thus, the G and B pixels are still capturing some of the Red SII light. It's not like mono sensors with RGB filters where you get a sharp cut off between the colour bands.

Hi again

Yeah, that's what I was thinking. Makes it hard to do 'before' and 'after' comparisons. Also, the green region in colour sensor graphs are the response of two pixels rather than just one - Oh, I think someone's already said that!

I think if the debayered version performs reasonably well then that's what you'd hope for and can be happy

Cheers

Louise

[russellhq]

Hi again

Yeah, that's what I was thinking. Makes it hard to do 'before' and 'after' comparisons. Also, the green region in colour sensor graphs are the response of two pixels rather than just one - Oh, I think someone's already said that!

I think if the debayered version performs reasonably well then that's what you'd hope for and can be happy

Cheers

Louise

I think the charts are for only 1 green pixel, otherwise I would expect the QE of the green to be much higher than the R or B, but instead, the peak is roughly the same.

You should be able to do before and after comparisons with the narrowband light. Since you know that 100% of the incoming light is passing through the OSC filter, therefore you can make a direct comparison with the debayered sensor and evaluate the impact of debayering on a pixel by pixel case.

But as you say, if you're happy with the results then that's really what matters.

[Thalestris24]

Ah, yes, you must be right as it's absolute QE. For my own interest, I was thinking of just doing a simple, subjective comparison. Two different cameras but with the same lens and same scene and with same manual exposure and ISO. It's surely the result that counts . I don't have any narrowband filters at the mo. I thought I'd do some short daytime exposures and some long astro exposures. I could simply desaturate the colour images and just compare side by side. I'll think about that a bit and I don't have the debayered camera yet. Hopefully it won't be much longer!

Cheers

Louise

Edit: of course, it will be cloudy when I get it!

Edited April 29, 2014 by Thalestris24

[russellhq]

Ah, yes, you must be right as it's absolute QE. For my own interest, I was thinking of just doing a simple, subjective comparison. Two different cameras but with the same lens and same scene and with same manual exposure and ISO. It's surely the result that counts . I don't have any narrowband filters at the mo. I thought I'd do some short daytime exposures and some long astro exposures. I could simply desaturate the colour images and just compare side by side. I'll think about that a bit and I don't have the debayered camera yet. Hopefully it won't be much longer!

Cheers

Louise

Edit: of course, it will be cloudy when I get it!

The trouble I see with that approach is you have no control over the scene which will be constantly changing. And that will have a significant effect on your image which could give misleading results.

I have debayered only half my sensor, so when I get the time to do some real world tests, I'll be able to do a direct comparison that will be able to negate the effects of the changing scene.

[Thalestris24]

Hiya

True, but for a daytime scene it can be an indoor still life or something A long exposure night time view would be trickier but a widefield sky view probably won't change much. At the end of the day, I'm only looking to see roughly how different the responses might be. Say I do a 2 minute exposure with both cameras but find the mono is a bit fainter but can get a similar result to the colour camera by exposing for an extra 20s. That will be good enough for me. I don't much care about gathering precise data, I only want to see if the mono camera is still useable, really.

Cheers

Louise