Friday, July 10, 2020

Image Algorithmics on RGBW Color Filter Misconceptions

Image Algorithmics kindly sent me a presentation with the company view on RGBW CFA advantages:

"There is a strong preconception in the market that RGBW does not work well. This is understandable given the failure of previous attempts. What's worse, many engineers now believe that it fundamentally cannot work well. I am attaching a slide deck to address this misconception.

We have tested our algorithms on 0.8u, 1.0u, 1.12u and 2.8u RGBW sensors. RGBW has a 6dB+ SNR advantage over Bayer in low light, read noise limited conditions and 3dB+ SNR advantage over Bayer in bright light, shot noise limited conditions. RGBW also has a 6dB dynamic range advantage.


  1. The Aptina implementation already used the W (or actually Y, in their case) to denoise the RGB.

    One remaining issue not discussed in these slides, is an increase in intensity and chromaticity of ghosts that are incident on the sensor at angles well outside the CRA + margin ray angles, due to the higher off-angle crosstalk through the W pixel.

  2. Aptina indeed used W/Y to denoise RGB, as did Sony, but the myth of poor bright light performance still persists. Different implementations have solved different subsets of the problems mentioned in the slides but, unfortunately, none of them solved all the problems.

    Your ghost issue is interesting. I have not encountered it, but I'd like to know more about it. Please contact me offline (tsingh at imagealgorithmics dot com). I will be happy to share our image set with you.

  3. I think these tiny pixel sensors could be the future, if considered only in the terms that you collapse the RGB pixels into one, thus finally getting true RGB values instead of the bayer upscaling.

    The pixels are small enough, that no optics can ever resolve them. Thus the RGB values would be accurate.

    And RGBW would be the best option, as then you would also have a real monochrome sensor. So finally a sensor that would give true RGB colors, no upscaling and as a plus a true monochrome sensor all in one!

    You would of course need a lot of pixels, at least 100MP to hit 3x25MP resolution. But considering we already have 100MP sensors, this should be possible. You could still have the option for the basic bayer upscaling. As people probably are too fond of their fake upscaled megapixel ratings.


All comments are moderated to avoid spam and personal attacks.