US patent
20050280727 describes a sensor that groups 3 or 4 detectors under a single micro-lens:
The variant that seems most interesting to me uses R, G, B and luminance (i.e. no color filter) photo detectors under the micro-lens. One of the main advantages listed in the patent description is the elimination of the anti-aliasing filter. A diffusion filter located between the micro-lens and the photo detectors makes sure they all see the same light color and intensity:
What would be the advantage of such a sensor? Let's first compare it with a traditional one that has a single photo detector under the micro-lens (assuming the same size micro-lens). The proposed sensor would have 4x more photo detectors, allowing it to record a complete set of RGBY values per pixel. It would have better resolution due to the lack of an AA filter, but the larger number of photo detector would seem detrimental to noise and dynamic range.
How bad would the noise be? Let's look at luminance noise first. The traditional sensor would capture roughly 1/3 of the incoming photons due to the color filter (ignoring other factors such as the photo detector's quantum efficiency for now). The proposed sensor would capture more light (roughly 1/2 of the incoming photons) thanks to the luminance photo detector not having a color filter. On the other hand it would have more readout noise because the photons have to be read out from 4 separate photo detectors. To evaluate luminance noise, let's assume that all 4 values are simply added (more sophisticated algorithms might be able to do better). This would cause some of the readout noise to cancel each other, with the total readout noise probably being around twice the amount for a single readout. So the proposed sensor would have twice the readout noise but a 50% stronger luminance signal, resulting in an relative noise increase of only 33%.
This does not sound too bad, especially since there would be no anti-aliasing filter. How much the AA filter costs in terms of resolution, and how much one gains by having per-pixel RGB values is hard to tell. However, based on what I gleaned in the Foveon vs. Bayer discussions in the Sigma forums it could vary from 1x to 4x depending on subject matter, with 2x being a reasonable "average" resolution gain for a true RGB sensor. In other words the proposed sensor could have a smaller number of micro-lenses (and therefore better noise characteristics) and still produce sharper images due to the lack of AA filter and to the additional color information.
How could this play out for the K20D? Assuming that this all pans out (i.e. comparable readout noise and micro-lens / filter / quantum efficiencies) Pentax might be able to build a sensor with e.g. 8M micro-lenses (32M photo detectors) to go against the D300 with its 12M Bayer photo detectors. Noise should be comparable (or maybe better for Pentax?) and the resolution should also be in favor of Pentax due to the lack of AA filter and to the per-pixel RGB values. Such a sensor would also be very interesting for B&W, as it would be only 1 stop worse than a true 8M B&W sensor. Image processing would require more horse power and storage due to the 4x increase in data values (due to the 32M photo detectors). However, to ease the pain I could imagine a "Bayer" mode where only one R, G or B photo detectors is sampled per micro-lens (i.e. per pixel). This would use a regular Bayer pipeline with only 8M values, providing higher processing speeds, but noise and dynamic range would of course be impacted.
What do you think?
Disclaimer: I don't really know what Pentax is up to. All I did was look for sensor-related patents from Pentax and dream for a while...
