New DSLR with Exmor R CMOS sensor

[agorabasta]

That would be amazing if it could be done, RGB direct from the sensor.

It has been possible for ages. The P&S sensor density on a large sensor can do it easily.

The main problem is the very low yield of fabbing. The minimum pixel-array better be 3x3, means you face about nine times higher defect rate. Assuming about 50% initial yield, we look at 5.5%. So the cost of a big sensor is about 81 times the cost of a smaller one. Yet if the yield for small sensor fabbing hits 90%, the feasibility of such composite sensors grows immensely.

My speculation here is that Sony is really getting close to point that makes composite sensors viable. The backlit design effectively removes the last fundamental disadvantage of smaller photosites, as now a few smaller ones are just as good as one large of the same total area.

[agorabasta]

In that situation, you might as well just read out the sub-pixel values individually and end up with 49 times the pixel resolution in the final image . If you want to achieve full RGB per output image pixel, you can simply downsample a Bayer image by a factor of 2 linear (4 area). By keeping the original pixel data, you can choose to either have high resolution, or downsample later and get good per output-pixel rendition.

49 times amount of data crunching is a bit too much at current tech . My conjecture was considering existing tech.
Moreover, physically connecting a few photosites together would prevent them from saturating one by one; so you really get the same DR as with one large sensel of the same total area.
In fact, I described something that really can be done. Both the simple striped layout and parallel connection of the photosites do simplify manufacturing tasks. If they manage good fab yields - it's commercially viable; more so if they manage to extort some $15k for such a machine

Stacked pixels (RGB) have been patented by Fuji (using dye-coupler technology developed from their colour film experience, US2005205903 (A1)) and in 2006 they demonstrated green dye pixels. http://techon.nikkeibp.co.jp/article/HO ... 15/123642/ I haven't found anything more recent unfortunately. A 2-layer architecture would give you the majority of the spatial and colour resolution of a full RGB sensor, but with 33% less information to have to cart off the sensor and process, and 33% less circuitry to produce (less fabrication steps) and go wrong. It would also boost sensitivity since more light is being turned into photoelectrons. In fact, I could envisage an array where one layer was a chequerboard of cyan and yellow, and the second layer a chequerboard of red and blue, with red underneath the cyan pixels and the blue under yellow. You could interpolate an RGB image from this, but this array would ideally turn all the incoming light into photoelectrons.

By registering more than one primary colour per one data value you may actually gain some efficiency. Even a Bayer-kind array could be made of White/Non-Red/White/Non-Blue sensels; the colour info stays complete. The problem is that you may lose greatly in spatial resolution due to the lens CA! Your example is about the worst case in that regard.

[Greg Beetham]

Would a two layer sensor actually work? would the bottom layer be in focus? and what about divergence, the corner sensels are getting shaded now with some lenses on FF, making vignetting worse, or so I have been led to believe....
Greg

[David Kilpatrick]

Foveon's 3-layer sensor works, but it just happens to be a lucky chance - silicon lets red wavelengths penetrate better than blue (reverse of water), so the top layer is blue, middle green, deepest is red. This matches the natural way in which a non-apochromatic lens will focus the wavelengths, red tending to be behind the focus plane and blue in front. So with many lenses, the Foveon shows better results from longitudinal chromatic aberration.

What it doesn't do any better is to handle the lateral aberration - different image size created - which is linked to this. You get fringes, but they are very sharp

David

[Greg Beetham]

Ah, that's ok then, I like my colour fringes to be sharp...
Greg

[alphaomega]

I think that Olaf put his finger right on it by stating:

Don't be silly everyone! The first Sony Alpha DSLR cameras with Exmor R sensors won't arrive before 2010 or 2011, I'm sure. The DSLRs currently announced to be released within the next weeks will have regular Exmor sensors, just like the current ones.

I am convinced now that the A500/550 will both feature the A700 Exmor sensor in a development like the D300s just announced. Maybe the A500 will be w/o video and the A550 with. Who knows?
My son wants to purchase my A700 so I have 98% made up my mind to purchase a new A700 before it is too late. I just do not want video and I feel that we will see cameras with A700 features removed at a higher price than the £540 or so I have seen a body only A700 for. Anyone knowing a UK source with a cheaper price? I just feel we are going to be presented with 500 series DSLRs at higher price points and less features than the A700 and maybe a marginal increase in less higher ISO noise. A bit like what happened to the 300 series with a bit of performance improvement. If the Exmor R is imminent and it can gather twice the light of a standard sensor why is Sony majoring on FF? Surely if they bring out shortly an Exmor R as APS-c It will outperform the FF sensor in all respects.

[Javelin]

Sony says the technique used for exmore R will have little or no effect on the larger sized sensors. so you might get 2x the sensitivity on the tiny sensor then you might only get a tiny fraction of that benefit for a much more expensive sensor. maybe though they can use the exmore R sensor to provide video in some way without making it too complex

[dkloi]

Would a two layer sensor actually work? would the bottom layer be in focus? and what about divergence, the corner sensels are getting shaded now with some lenses on FF, making vignetting worse, or so I have been led to believe....
Greg

We are talking a separation between layers of the order of a micron or two, an order of magnitude less than the depth of focus (not to be confused with depth of field), hence no significant additional problems with longitudinal CA. In comparison, negative film emulsion is of the order of 20 microns in thickness.

Vignetting can be an issue but this depends on good design of microlenses and the circuitry. With the Fuji stacked pixels, the colour filtering/absorption is in the photodiode region hence you could get rid of the filter array usually deposited on top of the sensor, upon which the microlenses are placed. Without this filter layer, you could place the microlenses closer to the photodiodes and minimise acceptance angle issues, might even be an improvement over current sensors. Being BSI also helps alot.

Foveon shows stacked pixels work. I haven't heard of significant shading problems (but I haven't looked that hard). From what I can find, they are about 5 microns deep (600nm light has a 1/e extinction length of 2.6 microns, hence you get about 85% absorption in 5 microns). Going to organic dyes would allow you to increase colour separation and wavelength selectivity and pixels wouldn't need to be that deep. Hence a two layer organic dye array would be better than current Foveon sensors in this regard. http://www.streetsofdublin.com/photogra ... ensor.html

Cheers,
Daniel.

[dkloi]

In that situation, you might as well just read out the sub-pixel values individually and end up with 49 times the pixel resolution in the final image . If you want to achieve full RGB per output image pixel, you can simply downsample a Bayer image by a factor of 2 linear (4 area). By keeping the original pixel data, you can choose to either have high resolution, or downsample later and get good per output-pixel rendition.

49 times amount of data crunching is a bit too much at current tech . My conjecture was considering existing tech.
Moreover, physically connecting a few photosites together would prevent them from saturating one by one; so you really get the same DR as with one large sensel of the same total area.
In fact, I described something that really can be done. Both the simple striped layout and parallel connection of the photosites do simplify manufacturing tasks. If they manage good fab yields - it's commercially viable; more so if they manage to extort some $15k for such a machine

I take your point about existing technology, but that's boring . I like to speculate but bounded by the current research.

The question is whether it would give you a significant improvement over a conventional sensor to justify the extra complexity/cost. The fact is that Bayer sensors work very well, you have to do quite a lot to get improvements for pictorial use. Foveon is one end where there are demonstrable advantages in certain areas but it involves three times the data for a 40% improvement in linear resolution. David's idea of a single microlens coupled to a triplet of RGB pixels sounds like it would do exactly the same as your 49 cross-linked sub-pixel scheme, namely produce a RGB sample at a pixel location, but with 1/3rd the sensitivity of an ideal RGB stacked pixel. I think cross-linking photosites is a good idea for on-chip (pre-ADC) binning to reduce read noise when downsampling. I'm not aware of this being done on CMOS sensors (old technique for monochrome CCDs where binning is relatively easy to do, less easy to do with Bayer CCDs) unless the Fuji EXR sensor implements this (or maybe they do the binning post-ADC).

A two-layer approach would give you most of the 40% increase in linear resolution but with 33% less complexity of Foveon. Conversely, you could manufacture a two-layer sensor with half the pixels as a conventional Bayer, the same linear resolution, same amount of data, better sensitivity (due to larger pixels and higher QE), and better colour resolution (though slightly less than Foveon).

Stacked pixels (RGB) have been patented by Fuji (using dye-coupler technology developed from their colour film experience, US2005205903 (A1)) and in 2006 they demonstrated green dye pixels. http://techon.nikkeibp.co.jp/article/HO ... 15/123642/ I haven't found anything more recent unfortunately. A 2-layer architecture would give you the majority of the spatial and colour resolution of a full RGB sensor, but with 33% less information to have to cart off the sensor and process, and 33% less circuitry to produce (less fabrication steps) and go wrong. It would also boost sensitivity since more light is being turned into photoelectrons. In fact, I could envisage an array where one layer was a chequerboard of cyan and yellow, and the second layer a chequerboard of red and blue, with red underneath the cyan pixels and the blue under yellow. You could interpolate an RGB image from this, but this array would ideally turn all the incoming light into photoelectrons.

By registering more than one primary colour per one data value you may actually gain some efficiency. Even a Bayer-kind array could be made of White/Non-Red/White/Non-Blue sensels; the colour info stays complete. The problem is that you may lose greatly in spatial resolution due to the lens CA! Your example is about the worst case in that regard.

Not sure about this pattern being any worse for resolution loss due to CA. It shouldn't be any worse than a conventional Bayer.

Kodak's trumpeted "white pixel" pattern does have the problem that the white pixels would saturate faster than the (occasional) non-white pixels scattered through the pattern, assuming that they didn't do something really complicated like resize them differently. In a two-layer scheme, there's no additional complexity to make the top layer pixels of larger capacity to the lower level pixels leading to saturation of both layers (under white light) at the same exposure level.

Cheers,
Daniel.

[agorabasta]

Not sure about this pattern being any worse for resolution loss due to CA. It shouldn't be any worse than a conventional Bayer.

Kodak's trumpeted "white pixel" pattern does have the problem that the white pixels would saturate faster than the (occasional) non-white pixels scattered through the pattern, assuming that they didn't do something really complicated like resize them differently.

The cyan & yellow consist of two primaries each, so they are blurred by the CA.

Kodak's approach was plain dumb. As soon as they go for white sensels, there's nothing else to lose in resolution. They should've tried white-cyan-white-yellow pattern - it would saturate even smoother than classic Bayer. WCWY collects about twice the photons than GRGB at the expense of CA-induced blur in all channels. But then you get almost no CA left to correct

[agorabasta]

While we are discussing the improbabilities, here's one more thing that's very easy to implement though totally unrelated to matters discussed above.
As long as a cam has got the sensor-shift drive (SSS or whatever), they could've got rid of the ugly optical LPF. All you have to do is simply shake the sensor around at about 30kHz at about half-pixel swing during exposure. That would provide the same blur as the LPF.

But such a system can be controlled as to the amount of blurring and it could also be switched off completely if you don't need the LPF.

The cost of such arrangement would be negative, btw

[dkloi]

Not sure about this pattern being any worse for resolution loss due to CA. It shouldn't be any worse than a conventional Bayer.

Kodak's trumpeted "white pixel" pattern does have the problem that the white pixels would saturate faster than the (occasional) non-white pixels scattered through the pattern, assuming that they didn't do something really complicated like resize them differently.

The cyan & yellow consist of two primaries each, so they are blurred by the CA.

Kodak's approach was plain dumb. As soon as they go for white sensels, there's nothing else to lose in resolution. They should've tried white-cyan-white-yellow pattern - it would saturate even smoother than classic Bayer. WCWY collects about twice the photons than GRGB at the expense of CA-induced blur in all channels. But then you get almost no CA left to correct

You're forgetting the second (primary) layer which also has to be taken into account when doing interpolation/reconstruction of the final image. There's nothing magical about CA. If it's there at the image plane, it'll be treated exactly like any part of the image. I'm not really following you when you say "blurred by CA".

Cheers,
Daniel.

[PhotoTraveler]

For that to work, you need you 30kHz movement to be on top of the travel of the sensor, which wouldn't be that easy, but more importantly you would have to figure out what the linear SSM motors can do. A quick search says they can do upwards of 5MHz. But that's apparently the traveling wave, how fast you can reverse direction isn't clear.

While, yes, they can probably do it and it would look like some orbital wiping pattern as it moves. Who knows what other issues it may cause, might really short the life of the SSS system. Could be a good case for the pentax system where it's magnetic.

While we are discussing the improbabilities, here's one more thing that's very easy to implement though totally unrelated to matters discussed above.
As long as a cam has got the sensor-shift drive (SSS or whatever), they could've got rid of the ugly optical LPF. All you have to do is simply shake the sensor around at about 30kHz at about half-pixel swing during exposure. That would provide the same blur as the LPF.

But such a system can be controlled as to the amount of blurring and it could also be switched off completely if you don't need the LPF.

The cost of such arrangement would be negative, btw

[David Kilpatrick]

Modern low-pass filters do not simply blur (if that was required, it could more easily be done by a focus offset linked to focal length and aperture data). They are based on thin layer coating diffraction, and produce an exact blur pattern from incoming rays which is tuned to the standard de-Bayer convolution of the sensor. The technology used is related to the Minolta Portrayer P filters, which used diffraction from multiple thin layer coatings in a pattern over ten years before any lo-pass filters were developed from the same technology.

The early lowpass filters on camera like the Canon Ion were refractive - the Ion filter used prismatic faces very similar to the Duto soft focus filter developed in the 1950s as an alternative to the Softar. The Duto Disc has concentric rings of slightly angled faces. Today' DSLR lowpass filters can not be made using refractive bluring, diffraction coatings are the only adequate solution. These can be computer-designed to provide exact blur patterns. An early use of this (with polymer films) was the B+W Spectrastar filter. A point source was rendered as a precisely placed ring of wavelength separated dots. This is also the basis of rainbow diffraction foils and decorative finishes. It's hundreds of times more coarse than thin layer coating, and aimed at producing rainbow effects, while the lowpass filters are of course aimed at a pure diffusion.

To get even close, the movement of the sensor would need to be a circular one, or covering a specific circle with deceleration to the edges and a complex pattern of diametric movements.

David

[kingfisher]

hello everyone

i can tell you that both camera`s , the A500 and the A550 have CMOS EXMOR R

iàm sure

greetings
gustav