Wednesday, August 01, 2012

Mobile Image Sensor Progress vs Other Industries

PR Newswire: PWC Mobile Innovations Forecast predicts that between 2011 and 2015:

  • Image sensor: Megapixels per dollar (MP/$) will improve 20% CAGR.
  • Device connectivity speed: Megabits per second per dollar (Mbps/$) will improve 37% CAGR.
  • Infrastructure speed: average Megabits per second will improve 54% CAGR.
  • Processor speed: GigaHertz per dollar (GHz/$) will improve 53% CAGR.
  • Memory: Gigabits per dollar (Gb/$) will improve 48% CAGR.
  • Storage: GigaBytes per dollar (GB/$) will improve 35% CAGR.

So, whoever complains that we are shrinking pixels too fast, should compare it with Mbps, GHz and GB of others.


  1. people are complaining you guys are shrinking pixels too fast because in many cases, the picture quality actually decreases as pixel count increases. As we seen some 1/2.3" 12Mp CMOS parts have better picture quality than a 1/2.3" 14Mp CMOS from the same P... company, even when the 16Mp picture is re-sized to 14Mp for a fair comparison.

    1. The development of small pixels goes by steps. For example, the first generation of 1.4um pixels was clearly worse than the 1.75um pixels of that time. The second generation improved them, and now most big companies are at 3rd or 4th generation of 1.4um pixels. They are much better then the first one. Now the pixel advances achieved in 1.4um pixels are scaled back to 1.75um pixels and they become better too.

      So, it depends on what exactly pixels you compare. It also depends on optics, signal processing chain and even on camera assembly quality and assembly house.

      Said that, the progress in other industries has its own trade-offs. For example, all things being equal, bigger RAM in your smartphone means shorter battery time. Faster processor in your laptop means noisier fan. Remember old phones with B&W displays with zero power and excellent visibility in sunlight? Now compare it to modern smartphone displays under the sun. Remember times when you charged your phone once a week? Now we traded it for colorful displays, fast CPUs and Mbps-rate data connectivity.

      Small pixel has its own logic of development. Its progress is combined with progress in image processing, optics and fab processing tools. Even if you do not see an immediate advantage now, it might think differently in five or ten years from now.

    2. Meant to be "you might think differently in five or ten years from now."

  2. Image sensors and displays are the only technologies shown that are really rated by a perceptual quality assessment, rather than a single performance measure.

    Tracking MP/$ makes is disingenuous and only serves to continue to drive the industry and consumer marketing in the wrong direction (or at least in a direction that does not optimize for the image quality).

  3. With few exceptions, pixel producers have to produce what will sell in consumer applications. Some want 1.1, or hell, 0.9! It's more along the lines of "they want 1.1, so let's do our best to make a 1.1 device that matches 1.75 in other performance metrics". If you try to go too much into educating the consumer, you're on the loosing end of the curve. Remember Zima and Betamax?

    1. Once we get an acceptable quality 0.9um pixels, one can get 1 Giga-pixel sensor in full-frame 35mm format. Do you say consumer does not need this?

    2. As a photographer I can say I do not feel I need this. Want this? Sure do. Interested in this, absolutely. What I really want is sharp images, amazing dynamic range and colour rendition with full colour sensor, no bayer or other pattern and amazing low light performance, oh and 8K video at 120 FPS. Not fussy at all ;).

    3. Consumers do need this, Vladimir, and as you know, they want it now! The only way we get better small pixels is by trying. I'm just responding to the sense that we are shrinking too fast. Inevitably, the next shrink has relatively poor performance until we learn what it takes to improve it. Do we need 1.1 micron pixels in a cell phone? Probably not, the 1.75 and larger have better signal to noise. Will we sell imagers if we can make 1.1 affordably with acceptable trade-offs? Yes!

  4. Pixel scaling is desirable for non-professional, majority.
    Imager acts as informaiton storage.
    Therefore, large pixel number means large memory,
    which is so usefull for a lot of user like me. (^^)

  5. Vladimir, I am far from alone in being dismayed to discover that the image quality in consumer compact cameras has consistently deteriorated over the last 5-7 years as the pixel area has shrunk from some typical ~5 um^2 down to 2-3 or less. I would welcome an increased MPx count but only provided that these are "clean" pixels, not a noise garbage. And here is what puzzles me.
    Suppose, the technology has been perfected to a point where fundamental physical limits are the only ones that matter. Let's assume that: 1) All other noises where reduced to a level well bellow the photo electron shot noise; 2) The "other" noise has no correlation between pixels; 3) I have an access to unprocessed raw data. If so, then I have no reason to complain about too small a pixel.
    Indeed, a sum of Poissonian random variables is Poissonian, too. If I sum up a signal from every 4 adjacent pixels then the result is an equivalent of a twice larger pixel (in linear size) with a twice better SNR. In this way, I don't need to insist on a physically larger pixel in order to improve the SNR, because the math will give me an equivalent. But the above reasoning is only valid if the assumptions (1-3) are holding.
    Thus, my main question is: do present day photo sensors satisfy the assumptions (1-2)? Or is there still a way to go before the total noise will be limited by the PE shot noise only? If the latter is true, then I am afraid that large enough physical pixels are the only hope to reduce the image noise back to an acceptable level, even if at the price of the lower MPx count.

    Regards, LB.

  6. I suppose #3 is your main complaint since read noise is 2-3e- rms in contemporary CMOS APS devices and if that is a problem for you, you need to stop shooting in the dark. As far as row and column noise, that is typically undetectable in most compact camera quality image sensors.

    You could complain about the optics being as good as the pixels. Or you could possibly complain about color crosstalk but unless you get your #3 wish, you won't see that, and it is small in any case.

    Frankly, my Sony 1.4 um BSI sensor compact camera works very well in my opinion -- far better than anything I ever could shoot with film in the same size camera, and the noise is also quite low. Of course I keep pushing the camera in low light farther and farther because it is so tempting, and then noise issues start to surface again. But, compared to 15 years ago, 10 years ago, and 5 years ago, I think you have nothing to complain about in the land of compact cameras -- except for lack of access to the raw image file. I would like that too. I don't like getting JPEG'd files out, and I especially don't like what my iPhone camera does to the images (my other very compact camera). Apple camera engineers have created a DR problem in an otherwise very nice camera. (and I know you guys read this so c'mon, fix it already!)

    1. Just to be clear, smaller pixels are still coming along, and between digital integration sensor (DIS) and quanta image sensor (QIS) technology, I think there is a roadmap down to 100 nm pixels with gigapixel (gigajot) count. Such small pixels also help make plenoptic cameras more competitive in output resolution.

    2. Thank you for the answer. It is really reassuring if one can assume that, except for an extreme underexposure, the total noise is, essentially, a PE shot noise.

      While on the subject, 1) is it true that a "specific" (per unit area) Full Well Capacity is some 800 e/um^2 ? 2) Does it vary a lot between different sensors? 3) Does fill factor figure in the FWC? Also, 4) is there a way for an outsider to learn which sensor is used in a given camera and then get a hold of the specs? Say, what is a 5) FWC? 6) noise measurements? 7) Light utilization efficiency (including losses in sensor optics, filters, etc)?

      Back to the image noise, I was suspecting that an in-camera processing could be a culprit, but wasn't sure. My personal experience is a comparison between a 2005 Canon Powershot SD450 (CCD/2.5", 5 Mpx, px=4.9 um^2) and a 2012 Panasonic Lumix ZS19 (CMOS/2.5", 14 MPx, px=2 um^2). To get an idea of zs19's noise, take a look at the photo of the surfer girls (bottom of the page)

      Look at the full resolution image and magnify girls' heads to the full screen. The faces, hands and suits are severely mottled. At a lower mag, the picture looks as if printed on a sandy paper. This is very annoying. Only my images are even worse. And no, I am not talking about a low light high ISO. I use a base ISO 100 under California sun and see blotched faces (at similar magnifications). SD450 never did anything remotely of the sort.

      The visible noise is very different from a typical Poisson at low ISO. This is more like a fat tail "salt and pepper". Perhaps, a mighty shameless in-camera unsharp masking is failing to tell apart edges from the noise outliers, so, the latter get amplified. I would even forgive Panasonic JPEGing if only I could opt out of their tampering with the noise/sharpening.

      Regards, LB

    3. @ "1) is it true that a "specific" (per unit area) Full Well Capacity is some 800 e/um^2 ? 2) Does it vary a lot between different sensors? 3) Does fill factor figure in the FWC?"

      Full well per unit area is a moving target. It improves over the time and some companies are better than others, depending on the pixel generation. It also depends on how one measures it, for instance, linear full well or saturating. In any case, the current state of the art is many times bigger than your number.

      @ "4) is there a way for an outsider to learn which sensor is used in a given camera and then get a hold of the specs? Say, what is a 5) FWC? 6) noise measurements? 7) Light utilization efficiency (including losses in sensor optics, filters, etc)?"

      I afraid there is no simple way to get this info, other than paying to some company to do the measurements for you. DxOmark has a lot of useful info on many sensors, although it's not exactly what you want.

      Also, as you noted, a single number does not tell the whole story. Some examples: Salt and pepper noise is much more visible than the Poisson one. Some sensors suffer from image lag, where the noise depends on signal. Color crosstalk correction reduces the DR, so high full well and low noise combined with high crosstalk might give a worse final image than slightly lower full well and higher noise combined with low crosstalk.

    4. So, unless published by a camera test lab, the full specs remain proprietary. Too bad, because sample image galleries are a poor substitute to a good test drive.

      Not to abuse your patience anymore, short of going full weight into scientific publications, is there a literature to familiarize oneself with the stuff you mentioned and the current state of art?

      Many thanks, with kind regards, LB

    5. For basics one can read a book:

      For state of the art the best source is International Image Sensor Workshops archive:

      The modern low noise imaging is also covered here:

  7. @ "the image quality in consumer compact cameras has consistently deteriorated over the last 5-7 years as the pixel area has shrunk from some typical ~5 um^2 down to 2-3 or less."

    To check this I went to dpreview tests and compared this year's 12MP Canon Powershot 510HS with 5MP Powershot SD400 circa 2005:

    The older Powershot uses 1/2.5-inch CCD, while the newer one has 1/2.3-inch BSI CMOS sensor. The Macbeth chart patches at ISO 400 look a little better in the newer Powershot, in spite of smaller pixel size. It does not seem to support your claim.

    Actually, I recall the same complains repeat every time the new pixel generation arrives. In 2000, when 3MP compact cameras start appearing, many people said they prefer to stay with 2MP due to the low noise. Before that the same talk was at 1.3MP to 2MP transition. One can imagine how the compact camera would looks these days, if the industry has listened these customers.

    The sensors steadily progress toward a sub-electron noise level. So, the camera might be able to implement your 1-2-3 proposals in a couple of generations from now.

    1. Meant to me "the camera makers might be able to implement your 1-2-3 proposals in a couple of generations from now."

    2. Thank you. One counterexample is enough to dismiss my premature generalization. I agree that those sample images for 2011 Canon 510HS are tolerable. They are not nearly as bad as my Panasonic zs19 (see a followup to Dr. Fossum's reply). I still think that my 2005 Canon SD450 is perceivably better than 510HS at a base ISO and a high light.

      To my excuse, I did see a plenty of similar customer complaints on Amazon. I checked out the sample images for their cameras. It just so happens that their better cameras were the older CCDs with larger pixels while the disappointing ones were the latest CMOS'es with the smaller pixels. As an owner of SD450, I was not even aware about such a beast as an image noise. But a mere $300++ for the latest Panasonic have bought me a good lesson on the subject.

      Regarding the pixel count in compacts. How much does one need? Considering the mountains of images from the travel and life events, most are never to be printed, but only viewed onscreen. If onscreen, I can't help but play with zoom for details until pixelation is a distraction at which point I resample at a 1/2-1/3 px step to obtain a smoother image even though without a sub-pixel detail. Say, a field of 100 x 125 px is enough to hold a good enough portrait of a person to be immediately recognized in real life. I can stretch such portrait to a full screen and it looks fine (after resampling). I routinely take a 5 Mpx image (landscape) from SD450, fit it to a full screen, and then magnify some 16 times more and still get a reasonable quality portrait fitting a full screen. But this works only so long as my pixels are clean.

      In Panasonic, I have roughly 3 times the number of pixels. Paradoxically, I can not allow same magnification, because a person's face comes out "carved" by horrendous "pimples" which are now highly visible and distracting. There is a setting in Panasonic to save at a 5 Mpx resolution. One might think that at the same sensor size and same pixel resolution an image could be same quality as in Canon SD450. It is not. The "salt and pepper" is still there.

      Regards, LB.


All comments are moderated to avoid spam.