It was brought to my attention that Brandywine Photonics sells a special sensor said to be optimized for multispectral imaging, the FBX-2. The sensor has a number of interesting features. While the main version has 2056 x 256 resolution with 15um pixel pitch, it's said to be available in HDTV, 25MP, 1 GigaPixel(!), and butt-able array formats. The broad spectrum QE looks very good, especially 30% at 1000nm:
The read noise is said to be <10e-, and the "well depth" is 300K (not sure what are the units of this one).
Thanks to OSP for the link!
Full well capacity is measured in electrons (e-), and 300K is a lot of them :-)
ReplyDeleteFor some reason the datasheet calls it "full depth" rather than full well. Also, a combination of 10e noise with 300Ke full well would mean the linear DR is 90db, quite unique number. This is the reason that I'm not sure that "full depth" is measured in electrons.
ReplyDeleteI don't see why 300k electrons is not reasonable. 3 micron pixels with low fill factor have maybe 10k electrons and the 48 micron pixels from Radicon (Dalsa) hold 2.8 million electrons. The FBX-2 number seems right in the middle of these.
ReplyDelete300Ke alone might be reasonable. 300Ke in combination with 10e of noise starts to be difficult. Still possible to achieve though, just needs careful design.
DeleteBTW, state of the art full well of 3um pixels is closer to 20Ke:
http://www.imagesensors.org/Past%20Workshops/2005%20Workshop/2005%20Papers/52%20Agranov%20et%20al.pdf
Yes, the long wavelength over 900nm QE is good. But the pixel size is large and thick low doped Epi layer more than 10um could be applied to increase NIR QE in that case.
ReplyDeleteOne needs epi of order 100um to reach 30% at 1000nm. And even then it's not easy. Look, for example, at e2v QE curves:
Deletehttp://www.e2v.com/products-and-services/high-performance-imaging-solutions/imaging-solutions-cmos-ccd-emccd/qe-curves/
I think 90dB instantaneous dynamics at 300 K is over CIS state of the art. There should be a typo about RON in the data sheet. RON should be 42 e- for gain 0 (300 000 e- FWC) and 10 e- for gain 1 (30 000 e- FWC), and not the opposite as edited.
DeleteActually, it's not over the art. For 300Ke full well one needs conversion gain of order 6uV/e. Then 10e noise floor means 60uV of equivalent noise on floating diffusion. This is quite possible, just needs to be designed with care.
Delete50-60 µV rms is about the noise floor reached by Teledyne Hawaii detectors (SFD in pixel circuitry) using their low noise read out chain (<500 kpixels/s) when operated at 77 K and sampling one Fowler pair (off chip digital CDS). This corresponds to about 35 nV/sqrt(Hz) noise density taking a factor 4 for the bandwidth. The high speed read out chain (up to 5 Mpixels/s) features a slighly higher noise density. Is there really any example of existing CMOS read out chain (say few Mpixels/s rate)operated at 300 K reaching a noise density lower than 10 nV/sqrt(Hz) even for in-pixel circuitry pitch as high as 15-20 µm (such as for Hawaii and Brandywine devices)? This would be of strong interest for low light/high dynamics applications.
DeleteYou are right that low noise is harder to achieve at high speed. Still quite nice noise results have been reported in the papers. See, for example, here:
Deletehttp://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1643511
Thanks for correcting our data sheet. It should read noise <10 e- and FWC=30,000 e- for Gain level 0, with read noise <42 electrons and FWC=300,000 for Gain level 1; both in global shutter mode. I had the numbers switched. The FBX-2 is designed for dispersive hyperspectral imagers whereby each row sees a different wavelength, with the gain setting being row-selectable. Thus, you would typically set Gain=0 for the rows with low light levels (typically wavelengths <400 nm and >950 nm), and Gain=1 for the rows (wavelengths) with higher signal. Hope that helps.
ReplyDelete