Monday, October 03, 2016

Thesis on Image Sensor Characterization

Delft University publishes MSc thesis "Characterization of CMOS Image Sensor" by Utsav Jain with fairly detailed description of measurement setups for most essential image sensor parameters.

16 comments:

  1. It is unclear why some methods differ from the EMVA1288 method. It would be interesting to know why the standard method is not used. Or if it is indeed used, then why doesn't he just refer to the standard instead of describing the method in details. There are also some minor mistakes in the application of the standard.

    And btw, even if the standard is mentioned in the Abstract, it does not seem to be mentioned anywhere in the document or in the references. It would be nice to make clear where the standard is applied and where it is not.

    ReplyDelete
  2. I agrre with Arnaud, I also don't understand why they have to use their methods instead of applying the methods, which are suggested by EMVA1288.

    ReplyDelete
  3. Gents, the student worked for 10 months on the topic at CAELESTE, optimizing their measurement techniques, methods and algorithms. I guess that the people of CAELESTE are involved in evaluation of sensors much longer than the EMVA1288 standard exists, and that most of the CAELESTE evaluation platforms are "born" long time ago. That is probably the only reason why they sticked to what they had already ...

    ReplyDelete
  4. Hi Albert, nice explanation, but EMVA1288 is not just created yesterday and has passed some years of work and revisions. Since a masters work should be a scientific work under guidance (so I was told when I did mine 26 years ago), it is not on the student. And he achieved to quote the Janesick papers and lots of recent papers, so in my eyes, although the people at Caeleste have a longstanding experience in the field, I don't see that as a good argumentation not to mention EMVA1288 or discuss it in that type of measurements.

    ReplyDelete
    Replies
    1. From a technical point of view, the EMVA1288 is certainly useful once you have a polished, well-operating camera. But it makes way too many assumptions, known not to be generally true, to yield best learnings during development. So I think there will always be a need for non-EMVA1288 conform measurements ;-D

      Delete
  5. That's certainly true, that EMVA 1288 is not the only possibility to measure or characterize an image sensor. And there are many methods to do so, but honestly, I don't see this as a good argument not to mention it or compare to it in such a work. Even if the authors are not interested in industrial applications, most of the image sensors are made either for consumer and industrial applications, and the standard was created to help customers to compare. Therefore I don't see a good reason to ignore it.

    ReplyDelete
    Replies
    1. I don't think people in Apple, Google, Sony, or Samsung care about EMVA 1288 at all. They do the evaluations their own ways.

      Delete
    2. Another interesting argument why the EMVA 1288 standard, which is accepted and supported by AIA and JAII should not be mentioned. People from Sony have already been in the working group, so they are aware of it. Again I don't understand, why the assumption that companies like Apple & Google (US) and Sony (Japan) might or likely do their own testing is a good argument for a masters thesis in Belgium (Europe) not to mention or compare to the EMVA 1288 standard. In case your argument implies the standard is useless, not important or the like, I have to say that I obviously don't share this opinion, for the reason that I know the people who work in the EMVA 1288 working group and their efforts and being part of the group.

      Delete
  6. Hi all,
    Of course we want to honor the EMVA1288. The EMVA is a fantastic initiative and it deserves a much wider adoption than now is the case. The procedures described are all sound and well-thought. However for measuring the CVF of an image sensor we happen to know 5 competing methods (MeanVariance, (QE) Reference Pixel, CapacitorSwitch, ChargeInjection, Fe55). For several reasons we find (sorry guys) that the QE reference pixel method is the most accurate and the least subject to measurement flaws. The Photon Transfer method (aka Mean Variance method) is the best method when the response is linear over a range from dark to beyond the kneepoint readnoise/PSN, as interpreting non-linearity in the MeanVariance plot is ambiguous and at best not automatizable. It is perfect *if* there is no image lag that may smooth out the photon shot noise at low light levels, meaning that the image lag should be far below the read noise and that is hard to prove. Also consider that the slopes fitted to the MV curve are fits to a noisy measurement; you can only reach decent accuracy by averaging a large number of measurements. The (QE) reference pixel method does not have these flaws. However it requires the existence of a number of pixels that have a direct current output contact.
    Peng

    ReplyDelete
    Replies
    1. Apparently this discussion was triggered by one little sentence in the foreword of the thesis:
      “… hence the measurement procedures are more specific to Caeleste requirement and follows EMVA 1288 standards.”

      This expression is a bit unfortunate and should have been something like “… hence the measurement procedures are more specific to Caeleste requirement and are similar to EMVA 1288 standards.”

      Indeed, Caeleste does not adhere strictly to the EMVA 1228 standard and never claims to do so. The standard is aiming at camera characterization and does not cover all measurements required to characterize an image sensor. Often al lot more detail is needed, for instance to determine the real source of observed noise, non-linearity inside the pixel or sub circuits. (see also Peng's comment)

      Besides, Caeleste mainly works for specific customers and does not offer standard products. Many measurements are specific and adapted to the custom design at hand or to the specific requirements of the customer. The issued test reports always contain a detailed description of each measurement executed: the method, the equipment used, the calculations, the traceability, …

      For Caeleste, the EMVA 1288 is a good guideline and many measurements are done in a very similar way. However, Caeleste focuses on custom design requiring custom measurement and wants to keep the freedom not to strictly comply to a standard. Then again, if customers require so, measurements can be performed to the EMVA standard.

      Delete
    2. Hi Dirk, if I screen the literature list, then there is no reference to the EMVA1288. I know that you are more interested in procedures and measurements that answer your design questions on the single pixel level. If you use methods, which give sufficient results for your work and for your customers, that's fine. The EMVA1288 standard was created not because the members occasionally had some spare time and did not know what to do, but because if there are many, many elaborated methods to evaluate quality parameters, customers have no chance to compare. And it doesn't play a major role if the customer is a camera manufacturer purchasing an image sensor, or if the customer is a camera user purchasing a camera. The standard should enable the customer to compare quality parameters, which was very difficult before.
      In my understanding it would have been correct in a thesis about "Characterization of CMOS image sensor" which certainly compiles the methods you use at Caeleste, to mention and quote the standard, even say where it is not so good because the linear camera model fails, but not mentioning it or not referencing it is in my opinion not correct, and that's what I wanted to say.
      But unfortunately in most replies, it sounds like I wanted to say you should only use the standard, which I certainly did not intend.
      On the other hand everybody is invited to join the group and help improving the standard, it is all about helping the customers achieving better qualified decisions for their applications.
      Unfortunately I had my share of experiences with sensor developments, where the single pixel measurements performed absolutely well while the total image sensor was not working at all, therefore I might be a bit picky at that.

      Delete
  7. There are those that follow standards, and then there are those that create the techniques from which standards follow. I am not even sure I understand the point of most of these comments.

    ReplyDelete
  8. Well, if I may explain for me the point was, that I was surprised and a little disappointed that in this recent work there was no mentioning, no quotation and no reference to the EMVA 1288 standard. I never tried to say, that it is the only way to measure or characterize image sensors. I am aware that image sensor developers favorize the reference pixel measurement (maybe because it is nearer to their way of looking at the the semiconductor), on the other hand the EMVA standard tries to approach it from the perspective what the image sensor is delivering - the image.
    Unfortunately the discussion turned a little into which is best.

    ReplyDelete
    Replies
    1. I guess that would be a different MSc thesis! My feeling is that standards are useful when they are useful, and useless when they are not. I don't think EMVA1288 has ever been mentioned in any of my labs. Not that there is anything wrong with such a standard...it is just not useful to my work.

      Delete
  9. I know that due to your achievements this statement has an important weight. If you are convinced that the standard has no use for your work, I didn't mean to say it has. But, although being less important than you in the field of image sensors, I don't share your opinion about mentioning in that thesis, but maybe we have a different view on scientific work. I apologize for the discussion, I will stop from here, because it doesn't change anything, anyway.

    ReplyDelete
    Replies
    1. Definitely no need for an apology. Comment away!

      Delete

All comments are moderated to avoid spam.