Jiaju Ma et al. of Gigajot Technology, Inc. have published a new article titled "Ultra‑high‑resolution quanta image sensor with reliable photon‑number‑resolving and high dynamic range capabilities" in Nature Scientific Reports.
Abstract:
Superior low‑light and high dynamic range (HDR) imaging performance with ultra‑high pixel resolution are widely sought after in the imaging world. The quanta image sensor (QIS) concept was proposed in 2005 as the next paradigm in solid‑state image sensors after charge coupled devices (CCD)and complementary metal oxide semiconductor (CMOS) active pixel sensors. This next‑generation image sensor would contain hundreds of millions to billions of small pixels with photon‑number‑resolving and HDR capabilities, providing superior imaging performance over CCD and conventional CMOS sensors. In this article, we present a 163 megapixel QIS that enables both reliable photon‑number‑resolving and high dynamic range imaging in a single device. This is the highest pixel resolution ever reported among low‑noise image sensors with photon‑number‑resolving capability. This QIS was fabricated with a standard, state‑of‑the‑art CMOS process with 2‑layer wafer stacking and backside illumination. Reliable photon‑number‑resolving is demonstrated with an average read noise of 0.35 e‑ rms at room temperature operation, enabling industry leading low‑light imaging performance. Additionally, a dynamic range of 95 dB is realized due to the extremely low noise floor and an extended full‑well capacity of 20k e‑. The design, operating principles, experimental results, and imaging performance of this QIS device are discussed.
Ma, J., Zhang, D., Robledo, D. et al. Ultra-high-resolution quanta image sensor with reliable photon-number-resolving and high dynamic range capabilities. Sci Rep 12, 13869 (2022).
This is an open access article: https://www.nature.com/articles/s41598-022-17952-z.epdf
it's a gigajoke? they come back to classic dual-gain pixel!
ReplyDelete163 Mpixel photon-counting sensor - congrats to the Gigajot and TSMC engineers! This advancement enables high-resolution low-light imaging for many applications.
ReplyDelete0.35e is not really photon counting?
DeleteSo many puzzling points in this paper! 1. Why is it called QIS? What is its definition? So confused now. 2. In which sense that it is different from our CIS? Can we now call all low noise image sensor QIS? 3. Can 0.35e- noise sensor really called QIS? 4. What is the definition of photon number resolving? Can 0.35e- noise really do photon number resolving? and it even said 'reliable'? Haha... You can have 'reliable photon number resolving' capability only when your noise is below 0.15e-
ReplyDelete@anonymous While I think these are all good questions your raise, this is not my paper so maybe one of the authors will reply eventually. There is really no reason to insult the company and the authors if you have genuine questions about the contents of this paper.
ReplyDeleteIn my opinion, 0.35e- rms median noise in a 163Mpixel sensor is great for low light applications, no matter what you call it. As discussed in the recent review paper, you could call this a low-light CIS using technology developed for QIS.
Also, DCG has a heritage that started with cascaded integration at Micron (by me), followed a year later by a similar approach for PPD devices by Rhodes at Micron, followed several years later by what you refer to as classic DCG that came out of Omnivision after Solhusvik and Rhodes changed companies from Micron. LOFIC came out later and seems to have been independently reinvented. Gigajot has not published the details of their version of DCG and frankly, I don't know the details.
There is no insulte for the authors at all. When you use a title with "...reliable photon counting...", all the guys have questions to ask. Compared to Eric Fossum's hard comments to contributions of other researchers/engineers, I think that these questions are more reasonable.
ReplyDeleteGreat great job Gigajot! It's funny and at the same time sad to see some (jealous) academics and others post insulting comments. I'm amazed to see a small startup is capable of making a sensor that largest sensor makers are unable to make. more than 150MP, 0.35 read noise and at the same time 95dB dynamic range with tiny pixels. And it's a product not a research project from a university lab that never will be commercialized. I'm old enough to remember how the legacy members of imaging community resist against success (for example: transition from CCD to CMOS).
ReplyDeleteN.D.
People are not jealous as you said. Simply against abus of language in a scientific paper. If they can publish as a nice design with a tiny CIS pixel with a crazy noise performance, we will applaud. We appreciate their achievement, no problem, but is it a "reliable photon counting" with 0.35e after CMS8?
DeleteWhat is the threshold to be considered reliable? 90%? Based on this paper JEDS.2013.2284054 0.35e- read noise translates to more than 92% accuracy. Pretty reliable. The right question you should ask yourself: what is the closest competitor to this sensor? Gigajot has other lower resolution sensors with 0.2e- read noise. IMO, the beauty of this sensor is you get all the good specs in one sensor. And the small pixel size makes it desirable for consumers and other high volume applications.
DeleteIs this great sensor for imaging society? Yes! Is this great product from a start company? Yes! All true. But I don't like the title they used in such a scientific paper. There are huge differences between CCD and CMOS, fundamentally different! Is QIS a good concept? Yes. It is. But could you tell me which point in this paper make it as QIS that CIS does not have? If you want to give something a new name, fine! But you need to give us something NEW that we don't have before. At least I don't see what makes this sensor different from CIS.
Delete~ 92% counting accuracy sounds good, but not, just assuming an average of 1000e-/frame/pixel, i.e. count 1000 times, PSN +- 33e- , counting BER of 92% translate to ~80e-. In geneal, BER should be in order of or less than *e-6. The 0.15e- mentiond in Eric's paper is indeed the first step to go ==> BER = ~5e-6. But, there is always a but, don't forget other nonidealilities, beside for example the max counting speed(DR), the power. I guess an even bigger problem should be the uniforminity of the 0.5e- treshold in a small pixel (appreciatted if people from Gigajot can comment,^_^). This paper/ sensor spec is no doublt very impressive, as well as the techology, espectially for a startup.
DeleteSince Prof. Fossum also kind of (at least not denying) agrees with 'anonymous', and no response from GigaJot team yet, I guess they also kind of accept that this is not so appropriate to call this QIS, it is more like an 'enhanced' CIS. Regarding the name, I guess this is related to how funding works in US. If you said that you'll develope a CIS, no funding; but if you said you'll develope a QIS which is very different from current existing CIS, you'll get a lot of funding. Like MEMS/NEMS story since probably more than 10 years ago, no way to get funding for MEMS, but certainly OK for NEMS though it is not that different from MEMS.
DeleteRead my comment above. You are obviously an academic. You can call it Cheese Burger Image Sensor. It doesn't change anything. It's still a great sensor unlike anything else existed before. It doesn't look like Eric Fossum is part of Gigajot development. I didn't see his name in any Gigajot's publications. I just checked and his name is not on Gigajot's website https://www.gigajot.tech/team
DeleteD.N.
IMO Anonymous' questions are quite valid. The paper is more of a "incremental improvement" to the traditional CIS. There may be many minor circuit design tricks to reduce the read noise inside for sure , but I don't see any "breakthrough" from it though.
ReplyDelete- J
I consider a 5x reduction in read noise over the state of the art (as it appears in the table) as more than "incremental" improvement. Would a 10x reduction (to 0.15e- rms) be a breakthrough? Incremental to me means perhaps a 1.5x improvement, which in most industries is still gigantic (e.g., fuel economy).
DeleteFurther, this improvement in read noise has a large impact on low light imaging performance - something we were all sort of surprised by, thinking that shot noise was already the limiting factor at even 1.6e- rms read noise. So this is sort of breakthrough in community thinking, that hasn't really sunk in yet.
I will also note that the pixel is NOT a PPD but a different sort of design with better performance, well-reported in the last few years, although still a "CMOS active pixel with intrapixel charge transfer" in general terms.
The breakthrough reported in this paper is showing that photon-counting in very high resolution image sensors has become feasible. Teranishi argued in 2012 that about 0.30e- rms is the threshold for photon counting, and I have argued that for low BER, the threshold is 0.15e- rms. One could argue is it the median or the average or the histogram peak that we should consider, but this demonstration with this read noise and resolution is quite impressive to me, no matter how much "anonymous" detractors with a bone to pick want to minimize the significance.
I'm disappointed that Eric Fossum, inventor of gigajot tech, was not included as an author in this pioneer Nature publication.
ReplyDeleteThanks but I made no technical contribution to this project so a totally appropriate author list. Also please note that at least one of the co-authors is a major co-inventor on the Dartmouth QIS IP. Lastly, I have intentionally stayed away from technical work with Gigajot due to conflict of interest with my Dartmouth research activity and IP assignment obligations. These guys deserve all the technical credit for improving and commercializing the licensed Dartmouth IP.
DeleteEric Fossum name is not in any Gigajot's publications and patents and it shows he did not have any role in Gigajot development. My guess is he promotes Gigajot because he may have some shares in company and why not just get some free credit of the other people work?
DeleteI have been following Gigajot work and Dartmouth work. The management of Gigajot made a very clever decision not to follow Dartmouth path to make binary or multibit QIS, not practical at least with today fab technology. What Gigajot is doing in developing CIS with great improvements in low light and dynamic range is practical and has applications in real life. Very different than Eric Fossum's QIS delusional idea.
DeleteAre there any commercially available qis sensors or qis camera at the moment?
ReplyDeleteThere isn't any product information on their own website: https://www.gigajot.tech/product.html . So, I don't think one can eventually buy these sensors.
DeleteI tried accessing Gigajot's website, but it appears to be not working. Is Gigajot still operational?
Delete