Thursday, July 22, 2021

A (Wrong) Attempt to Improve Imaging

University of Glasgow and University of Edinburgh publish a paper "Noise characteristics with CMOS sensor array scaling" by Claudio Accarino, Valerio F. Annese, Boon Chong Cheah, Mohammed A. Al-Rawhania, Yash D. Shaha, James Beeley, Christos Giagkoulovitis, Srinjoy Mitra, David R.S.Cumming. The paper compares SNR of a large single sensor with an array of smaller sensors having the same combined area. The conclusion looks fairly strange:

"In this paper we have compared the noise performance of a sensor system made using a single large sensor, versus the noise achieved when averaging the signal from an array of small independent sensors. Whilst the SNR of a smaller physical sensor is typically less than that of a single larger sensor, the properties of uncorrelated Gaussian noise are such that the overall performance of an array of small sensors is significantly better when the signal is averaged.

This elegant result suggests that there is merit in using sensor arrays, such as those that can be implemented in CMOS, even if the application only calls for a single measurement. Given the relatively low cost of CMOS and the wide availability of CMOS sensors, it is therefore beneficial to use arrays in any application where low noise or multiple parallel sensing are a priority."


  1. So many wrong equations in the paper, I don't even know where to start

  2. I do not get if the same image was projected multiple times onto the smaller sensors, or the smaller imagers probed parts of the larger image. If the former, could the optics explain the results obtained? Nevertheless, I think it might have some important results for the industry especially when large image sensors are required for certain cases

    1. The paper implies that the same image is projected on each sensor in the array. They say nothing about the optics.

  3. Are the expectations and expert opinions, that the averaging can't compensate the higher read-noise caused by smaller pixels?

  4. I am pleased to see that we can now treat all image noise sources as Gaussian. The researchers have down-played this remarkable discovery.


All comments are moderated to avoid spam and personal attacks.