Thursday, January 26, 2023
Towards a Colorimetric Camera - Talk from EI 2023 Symposium
Tripurari Singh and Mritunjay Singh of Image Algorithmics presented a talk titled "Towards a Colorimetric Camera" at the recent Electronic Imaging 2023 symposium. They show that for low-light color imaging it is better to use a long/medium/short (LMS) filter that more closely mimics human color vision as opposed to the traditional RGB Bayer pattern.
Subscribe to:
Post Comments (Atom)
Ummm... This was all explicated about 20 years ago by Foveon, which had the additional advantage that demosaicing was not required. The main problem with the Foveon sensor was that its read noise was very high so it did not work well at low light levels. They even figured out in later versions that the blue resolution could be cut in half with little effect on the images.
ReplyDeleteFoveon is famous for not needing demosaicking, as you noted. However, Foveon is notorious for poor colors, especially skin tones. The spectral sensitivities of Foveon isn't close to satisfying the Luther condition.
ReplyDeleteCan you give us a reference for Foveon's color performance?
Here is one reference https://www.jstage.jst.go.jp/article/photogrst1964/66/1/66_1_57/_pdf.
ReplyDeleteI'm one of the authors and would like to clarify that this is not just for low light. We get better color and better luminance SNR in all lighting conditions. The presentation focuses on low light as noise has been the primary issue with LMS filters in the past.
ReplyDeleteThanks for reminding me of this work, I do recall reading about it when it was first published. If I may summarize it, Foveon X3's color accuracy was competitive with the CFA based image sensors in the market at the time. Furthermore, its color accuracy can be improved with a pre-filter.
ReplyDeleteCFA based sensors' color accuracy can also be improved with a pre-filter, for example see https://ueaeprints.uea.ac.uk/id/eprint/73655/2/Accepted_Manuscript.pdf
The problem with pre-filters is the loss of sensitivity and decrease in SNR. Other colorimetric cameras have been made for applications where light is plentiful, such as archival work at museums. Indeed tristimulus colorimeters can be viewed as single pixel cameras where noise is less of an issue.
Unlike previous colorimetric camera designs, ours does not trade SNR for color accuracy. Instead it improves both color accuracy and luminance SNR. This is indeed possible and explains why the Human Visual System has so evolved.
There's a baseline demosaicing only matlab code for a LMS filter, but using a more aggressive factor named LMS 12:12:1.
ReplyDeletehttps://ww2.mathworks.cn/matlabcentral/fileexchange/124890-lms-12-12-1-color-filter-array-demosaicing