Sunday, June 14, 2020

Perovskite 3-Layer Sensor Fulfills Foveon's Original Promise

ResearchGate publishes SPIE presentation "Color imaging sensors with perovskite alloys" by Mohammad Ismail Hossain, Wayesh Qarony, Haris Ahmad Khan, Masayuki Kozawa, Alberto Salleo, Jon Yngve Hardeberg, Hiroyuki Fujiwara, Yuen Hong Tsang, and Dietmar Knipp from Hong Kong Polytechnic University, Norwegian University of Science and Technology, Gifu University (Japan), and Stanford University:

"The conventional optical color sensors consist of side by side arranged optical filters for three basic colors (blue, green, and red). Hence, the efficiency of such optical color sensors is limited by only 33%. In this study, vertically stacked color sensor is investigated with perovskite alloys, which has a potential to provide the efficiency approaching 100%. The proposed optical sensor will not be limited by color Moire error or color aliasing. Perovskite materials with suitable bandgaps are determined by applying energy shifting model and the optical constants are used for the further investigations. Quantum efficiencies and spectral responsivities of the described color sensors are investigated by three-dimensional electromagnetic simulations. Investigated spectral sensitivities are further analysed for the colorimetric characterization. Finally, the performance of the investigated sensor is compared with conventional filter based optical color sensors. Details on the used materials, the device design, and the colorimetric analysis are provided."


  1. Perovskites or organic semiconductors or quantum dots. Great progress on these materials in the past years. Now, the trick is low read noise without complete charge transfer during the question of interconnects and don't you still get color mixing signals on the interlayer conductors? For example, on the conductor between red and green layers, do you get photoelectrons from both layers, or is one layer contributing electrons and the other holes? Either way how does this get separated out without introducing the Foveon problem in color and noise? Maybe I don't know enough about how these work, or maybe you need a dual conductor layer between colors, one for each? But that will impact fill stuff but so many questions ahead not to mention manufacturing issues. Where will silicon image sensors be by the time they sort all that out?

    1. With modern Si stacking technology and per-pixel ADC they could implement a digital CDS and get fairly low noise.

      They talk about photodiode on one of the slides, so I'd guess the separation layers serve as the other side of p-n junction.

      To me, the important unknowns are dark current, image lag or other long-term memory, and variability in mass production.

    2. Seems very difficult to do. It takes msec to do digital CDS with per pixel ADC. Such long CDS time is not friendly with 3T pixels

    3. So far, I dont think noise achieved using CDS (analog or digital) for a 3T type of device is as good as a good 4T device. Also, I don't think you understood my question about noiselessly separating out the component photocurrents (3 diodes, 4 wires). I agree that the other material issues are also important but they will get better over time (assuming continuous investment) but I think the things I mentioned are more fundamental issues that need to be addressed.

    4. I have no doubt that you understand this better than I do but I believe that your questions are answered in these papers:

      Figures 2 & 4:

      Planer p-n homojunctions:

      Layperson explanation to the expert (so not very well explained): Each layer has an opposite direction of current flow which distinguishes between the contribution of each color, the electrodes can be thin, placed on the edge of their respective layers, and overall pixel size can be smaller using perovskites instead of silicon.

      To Vladimir's first two questions (last sentence): "... perovskite SCs have low dark current and relatively fast photoresponse ..." (Source: Nature article, 1st link above).

    5. I agree that 3T does not have that good CDS like 4T pixel. On the other hand, the signal in 3-layer pixel is 2-3x higher than in Bayer one. So, in the end, the photon-referred noise might be not that different.

      3 diodes and 4 wires can be resolved if charge amplifiers with input virtual ground are used in place of SF. One can pick up blue and red signals from top and bottom plates respectively, while mid wires would be mix of the colors that needs to be subtracted. In terms of photon shot noise, this subtraction should be close to ideal, as top and bottom plates also give us fully correlated shot noise in blue and red.

      As for the dark current, even if perovskite does not have one, we still have diffusions on Si that are challenging in terms of leakage.

    6. Bi-color photodiodes have been used in infrared sensor, but the connection for 3 layers is much difficult indeed. Any contact openning/via will be a fundamental setback to its performance.

  2. Is there a true need for 3 color channels per pixel???

    We know the need for high color resolution (vs luminance resolution) is quite low in human visual system.
    And even compression image/video chose to undersample the hue/saturation/color resoul
    So maybe the idea to get 3 color sensing per pixel is a high effort for low benefits ?

    Am I wrong ?

    1. I think they are trying to improve sensitivity, rather than a color resolution. If we take RGGB superpixel, G signal is collected from 50% of the total area, while the other 50% is lost. R and B signals use only 25% of the area, while 75% is lost. The 3-layer pixel uses the whole area, so that the signal is 2-4 times larger, depending on the color.

    2. Anonymous, the presentation in this article says: "❑ Characterization is performed by a linear transformation matrix method. ❑ The method transforms colorspace of color sensor to colorspace of human standard observer. ❑ Finally color error is calculated according to a procedure outlined by CIE".

      They are using something similar to the Finlayson tristimulus matrix:

    3. There isn't any particular advantage to stacked photodiodes in most consumer products (as the Foveon experience showed) but there are plenty of industrial and scientific applications where the misregistration of the colors in Bayer sensors causes computation problems. Demosaicing only makes the situation worse because this replaces the raw, misregistered data into numbers that are all approximations. That is partly why there are still robust markets (currently expanding) for prism-based cameras.

      Even in photography, where the goal is pretty pictures instead of accurate data, the lack of need for demosaicing when using stacked photodiodes provides an advantage in the accurate visual delineation of transitions between objects, especially those at different distances.

      Basically, the answer depends on the market under consideration.

    4. Yes, there is - the advantages will be an increase in resolution and sharpness, a reduction of demosaicing artifacts, and an increased sensitivity (reduced noise).

      If the claims from the presentation hold true, it will a quite significant development in sensor technology.

  3. A couple of points:

    First, regarding Dana's comment on the Bayer pattern. The drawing in this slide set is not a Bayer pattern, which requires repetition of the same 2x2 pixel block (usually RG/GB). Demosaicing the drawing here would be complex because the repetition interval is 6x6 pixels.

    The original Foveon promise was to supply a 3-chip prism-based camera configured as an SLR for professional use. They built a few of these (the Studio Camera) but found there wasn't much of a market. Then, they decided to rescue the business by taking over the stacked photodiode patents held by National Semiconductor (one of the original investors) and moving Richard Merrill, the inventor, from NI to Foveon to help get the technology into production. The result was the F7 device in the Sigma SD-9.

    Incidentally, as Dana mentions, sharp color boundaries cause problems in color accuracy. This has been known at least since the first 3-tube cameras used the original Philips prism in the 60s. The Foveon bands overlap too much but no overlap is as bad.


  4. Those are some fairly thick light absorbing layers, even compared to quantum dot technology. I doubt MTF would be comparable.
    Since this is a spin coated process, with multiple electrodes reaching out to various heights, how are they creating the connection on a per pixel basis? That must be a bear to perform pixel patterning. Quantum dot technology relies on single layer charge transfer and can utilize top common and bottom pixel electrode. How do you transfer 3 layers of charge when at least 2 of the layers need to pass and imply pixel pattern through the third?

  5. I think what they mean by the 33% max efficiency of conventional sensors is a simple fact that the CFAs filter out 2 of 3 color components, so only 1/3 of the total light incident to pixel reaches the well. In reality it's a bit more complex, but this assumption is a fair starting point.

    This is an interesting development in sensor tech, worth watching.

  6. There is a looong way from a university paper to a technology and product that sells.

    Most ideas and companies die on this road to commercialization.

    So, claiming that they fulfilled the promise of Foveon, is a strong overstatement.

    I agree with Eric that they way their diagrams are drawn, the photocurrents from the neighboring detectors will add up, and will be indistinguishable.

    Foveon solved this problem quite easily, with a single "ground plane" (p-type "substrate"), with individual access to (fully) depleted photodiodes (n-type regions). Of course, this adds complexity and cost to the process.

    The bigger problem I see is a material system, its manufacturability and material properties.
    In this regard, nothing can beat silicon.

    1. @ will be indistinguishable

      Top and bottom electrodes have B and R signals respectively. The middle electrodes have G-B and G-R signals. I do not see a problem in extracting G from that.


All comments are moderated to avoid spam and personal attacks.