Wednesday, October 05, 2011

Small Pixels and Machine Vision

Adimec blog published a post on small pixel potential in machine vision applications: "Can small pixel CMOS image sensors be useful in Machine Vision?"

The conclusion is: "There is a trend towards using smaller pixels for machine vision applications leveraging technology developed for the high volume consumer camera applications. However the requirements for machine vision are very different and it will take some time before the image sensor pixels have high enough performance to make it into the machine vision cameras."


  1. Small pixels for "amateur" astrophotography are a problem too (and making mostly color sensors too). Smallest pixels in a deep space astro camera are 4.4 x 4.4 (ICX274AL), while in most cases bigger pixels are required. And they don't have to be fast as exposures are in minutes (except planetary cameras).

  2. There are some good points in this article by Gretchen Alper, but I think the author falls prey to several common misconceptions. Chief among these is the conflation of the benefits/drawbacks of pixel size with those of sensor size.

    The article generally speaks from the premise that pixel count is fixed, while sensor size varies with pixel size (with some exceptions, such as where it's mentioned that the cost benefit of smaller pixels could be traded for more resolution). However, that premise can lead to incorrect conclusions about pixel size. For example, if the given application requires a certain pixel count, one might be tempted to conclude that a smaller pixel (same format, higher count) can't work. But if a higher pixel-count sensor can deliver the same frame rate, and the DSP can keep up with it and still downsample to the desired pixel count, then that conclusion would be incorrect. Furthermore, if the problem is that the DSP cannot keep pace within the design criteria (e.g. heat), then it would be important to include that caveat in the conclusion about pixel size.

    I would mention in passing that the "fixed pixel count" position taken by the author makes sense for his audience (machine vision), where it's common for the design criteria to be set by performance and the goal is lowest cost possible. But it's worth noting that in many other applications (outside the scope of the authors article, of course), the design criteria is set by cost and size (including optics), with the goal of highest performance possible. Those costs change at an extraordinarily slow pace compared to that of achieving a certain pixel count.

    > The larger pixel image sensors (greater than 5.5 um) can allow for the best accuracy (i.e. Full Well and Read Noise)

    Here I think the author meant "e.g." and not "i.e.", since accuracy is a lot more than just FWC and read noise. In any case, the article here ties pixel size to sensor size, so I agree that larger sensors have higher accuracy. But what if you consider the impact of pixel size and sensor size separately? Does a large pixel in the same sensor size always have better accuracy? It does not, and using pixel-centric characteristics alone can be misleading in that case.

    For example, let's say two 2/3" (or any size) sensors both have an ISP that delivers the same 4 MP, but one is a 2um pixel with 10ke FWC and the other is a 6um pixel with 50ke FWC. On the surface, it would seem that 50 is better than 10. But since nine 2um pixels fit inside one 6um pixel, the max SNR in the delivered 4 MP output is actually going to be a lot higher than the 6um pixel, equivalent to 90ke.

    Furthermore, many spatially-varying signals that would have exceeded the FWC of the 6um pixel would be localized to a smaller number of the 9 2um pixels, making it possible to retain that detail. So even if the FWC scaled linearly with pixel pitch, the smaller pixel could deliver output with a lower percentage of clipped signal.

    There are downsides too, such as when the spatial frequency of the signal (e.g. a star) is high enough (while avoiding any blur and flare) to fall on just one of the smaller 2um pixels, clipping it, when it does clip the 6um pixel. I don't know how common that is, but the point is that it's not so simple to say that large pixels (in a given sensor size) are more accurate due to higher FWC.


  3. The article also mentions read noise. In applications that aren't already dominated by photon shot noise, small pixels (for a given sensor size) are going to have a much harder time of it. The image sensor designer would have to achieve a much lower read noise (scaled in direct proportion with pixel pitch) to match the performance of a larger pixel sensor.

    As read noise improves, the percentage of applications where SNR is dominated by photon shot noise increases. And in those cases, it's only the QE that matters for SNR (and FWC for max SNR too, of course). So then the question is how much variation there is in QE among pixel sizes. I'm certain that QE can be maintained at pixel sizes much smaller than the 4.5um mentioned in the article.

    FWC, read noise, and QE are not the whole picture, of course. The article makes a lot of good points about many additional factors, such as angle of response, optical/electronic crosstalk, and more. The main point I wanted to make is the separation of pixel and sensor.

    The article also makes a great point about how many machine vision applications require features that aren't common in consumer electronics (CE) and currently require a lot of extra transistors (e.g. 8T), such as global shutter and high frame rates. I am hopeful that some of these features will come to CE in the future. For a lot of applications, a sufficiently fast rolling shutter is a viable replacement for global shutter. The new Nikon One system is an indicator to me that higher frame rates are coming too, but for now there is still a huge discrepancy between machine vision and CE.

  4. This comment has been removed by a blog administrator.

  5. Daniel, it looks like my spam filter initially decided that your first comment is spam. I'll tried to bring back the original sequence of comments, hope it's ok now.

    I agree with you that small pixels have a lot of innate advantages. One trade-off in 4T pixel design is between the read noise and the full well. All things being equal, big full well, like in large pixels, requires lower conversion gain to fit voltage swing into the supply budget. Lower conversion gain inevitably cause higher read noise, again all things being equal. So, larger pixels with high full well have higher noise as well. There are different improvements of the basic 4T pixel design to alleviate the problem, for example Aptina's dual gain DSLR sensor, but they involve other trade-offs.

  6. > Daniel, it looks like my spam filter initially decided that your first comment is spam.

    I'm not surprised, it happens all the time. Apparently my writing style must have something in common with Nigerians that are seeking your assistance to move the sum of $34,000,000. ;)

    > [I] tried to bring back the original sequence of comments, hope it's ok now.

    Yes, thanks.

    > One trade-off in 4T pixel design is between the read noise and the full well.

    I had read about the conversion gain trade-off before, but now it's confirmed for me.

  7. Hello all, i know this is probably the wrong place to ask for career advice, but i am hoping to get some help! I am planning to join a CMOS imaging company which has a reputation to be struggling, but all i have heard are rumors and i don't want to make a career decision based on rumors. Job security is important and i got another offer from Teradyne. So basically i am trying to decide. Please any candid input will be appreciated. Sorry for hijacking this thread with my digression.

  8. Almost all companies are struggling these days.

  9. From your comment about a struggling CMOS image sensor company, i am guessing Aptina. Sorry, having Aptina and Teradyne as a choice does not make much difference. I would probably go for teradyne, as the earlier will most likely die soon.

  10. @ "all i have heard are rumors and i don't want to make a career decision based on rumors. Job security is important"

    One thing you can to to make the process easier is to take away the gravitas of it being "a career decision". You're picking a job, not determining the precise arc of the rest of your working life. Whatever your choice is, it will almost certainly not be definitive and final. People change directions and make new starts all the time.

    I agree with the "almost all companies are struggling" comment, and I don't think there is "job security" in the sense that if you show up every day and put in your time, your job and paycheck will be there into the indefinite future. On the bright side, you have two job offers in a weak economy, so you are doing really well, stressful though the search might be.


All comments are moderated to avoid spam.