4

I have the camera specifications from the manufacturer. The RGB quantum efficiency curve (I think it is also the spectral sensitivity curve) is provided. How can we compare these two cameras's quality of color reproduction directly from the curve? Or we need set up an experiment to do this? Thanks. enter image description here

enter image description here

  • 1
    What exactly do you mean by quality of color reproduction? At the end of the day, we 'reproduce photos' by storing just three pieces of data for each pixel: RGB. In general, there will be (uncountably) infinitely many combinations of frequencies which will be assigned the same RGB value. It's the pidgeon-hole principle. – Myridium Feb 16 '17 at 14:02
  • @Myridium Human eye also divide whole spectra to three-colour + brightness signals. That's why RGB colourspace works. If the eye would work with n types of colour receptors, we would have n-dimensional colour spaces for display and n+1-dimensional spaces for so-ho printers. – Crowley Feb 16 '17 at 14:09
  • 1
    You could use a monochromator to generate a beam of light with a very narrow wavelength range and measure the output of the sensor as you change the wavelength. Or alternatively you could use a diffraction grating and look at the range of colours across the sensor - I would expect a smoother transition for better colour capture. No model has a true RGB value as such, RGB is just one model to represent colours. There's also HSL, CMYK and almost certainly others. – Rob Feb 16 '17 at 16:21
  • 1
    @Rob What you describe is roughly how her graphs were generated. They represent the efficiency of the sensor to reproduce a generated wavelength through as series of generated wavelengths in a device called a spectrophotometer. – Stan Feb 17 '17 at 3:17
  • 1
    It seems to me this entire question is based on a false premise: That every specific color we perceive has a specific wavelength. That is not the case. Further, the 'true' RGB value of various items in the scene varies based on the properties of the light illuminating them. Broad spectrum light centered on 5000K (essentially D50) will make the same things look a vastly different color than narrow band 2700K sodium vapor lights. The RGB channel overlap in camera sensors that you seem to think introduces inaccurate colors works exactly the same as human vision does. – Michael C Jun 26 '17 at 18:00
4

How can we compare these two cameras's quality of color reproduction directly from the curve?

"which camera can get the RGB value closer to the true RGB value of the object, less RGB channel overlap"

A comparison of color reproduction potential based on quantum efficiency of the sensor filters is possible.

As others have mentioned, there are many contributing factors to the final color reproduction of a full color camera system. However, the sensor RGB sensitives are perhaps the largest contributing factor to color reproduction accuracy and their color reproduction performance can be measured.

What is true RGB

First we have to answer, what is the "true RGB" of a scene? A good definition of "true RGB" would be the relative responses of a human's three retinal cones to the scene. These cones are called LMS, long medium and short.

enter image description here

A spectrum of light integrated over these three sensitivity curves yields three LMS values that can be thought of as human RGB values, these are the target RGB values we want to reproduce with our camera if our goal is accurate color reproduction.

More commonly, we could also target the sensitivities of the XYZ color matching functions. These are linear combinations of the LMS functions so they are effectively interchangeable with the LMS functions.

enter image description here

Color Correction

In a digital camera, when a spectrum is integrated against the camera sensitivities (like the ones you posted) the resulting RGB values are called "camera RGB".

In most digital cameras there is processing step where a color correction algorithm (M) will be used to convert cameraRGB into humanLMS (or XYZ).

M(cameraRGB) = humanLMS

In this case, humanLMS will be a guess. It will not be perfect, and the difference between the guess and the real LMS value a human would have perceived is your color error.

Designing a good M is difficult because it is an under determined problem, some cameraRGB values have multiple potential humanLMS values (this is called metamerism) so it's not always possible to know exactly what the correct LMS is, but we can use natural image statistics and machine learning to make a guess at the most likely correct answer.

The most common implementation of M is a 3x3 linear transformation matrix, but if the camera sensitivities are not linear combinations of LMS then the transformation will contain errors. If the camera sensitivities happen to be linear combinations of LMS then the color error would be zero, this is called the Luther Condition. In practice digital camera sensitives never satisfy the Luther Condition so there is always color error.

Comparing Color Reproduction

There are now two factors that play into how accurate our LMS guesses are.

1) the design of our color correction algorithm M

2) how similar our sensor sensitivity curves are to the LMS sensitivities

This gets at the heart of your question: some sensitives will result in quantifiably more accurate colors than others because they are closer to the LMS sensitivities which makes it easier to guess the LMS value, which is the "true RGB" we desire

Or we need set up an experiment to do this?

What might be helpful is "ISO standard 17321, Sensitivity Metamerism Index". This calculates color reproduction accuracy based on spectral responses.

https://www.dxomark.com/About/In-depth-measurements/Measurements/Color-sensitivity http://www.iso.org/iso/iso_catalogue/catalogue_ics/catalogue_detail_ics.htm?csnumber=35835

This index tells you the average perceptual difference between colors recorded by your camera that have been linearly corrected by a optimized 3x3 matrix, and the known colors of a test scene.

The only problem is this procedure is done with a full camera so it's measuring the color error of the the sensor and the color correction matrix and the optics etc, not only the sensor.

If you truly only want to quantify the error of only two different sensors you could do the SMI procedure with the same camera and only change the sensor. Or instead of a physical experiment with a real camera you could simulate your camera in software and not include any optical or demosaicing contribution to the simulated cameraRGB values.

There are many papers on camera simulation for more info on that: http://color.psych.upenn.edu/simchapter/simchapter.pdf

"CIE Special Metamerism Index: Change in Observer" is another relevant standard meant for comparing color reproduction in humans with slightly varying spectral responses. I think you could apply this to camera spectra as well.

http://link.springer.com/referenceworkentry/10.1007/978-3-642-27851-8_322-1#page-1

  • I read the articles you cited. The links you provide dispute your claim. The question was about the use of the sensor data, not about the post processing required for human perception which your articles address (most notably the last one cited). The last article explicitly states that it is not possible without a "calibration" phase. (my term and quote emphasis). According to the articles, it is only theoretically possible when the sensor has characteristics that exactly conform to the "ideal independent observer." – Stan Feb 17 '17 at 19:39
  • @stan As far as I can tell, the fundamental question here is: given two RGB sensors with different spectral responses curves can we say anything about which can reproduce colors more accurately. That question is well understood and the answer is yes, the sensor with spectral curves closest to a linear combination of human LMS spectral responses will have less uncertainty when you go to apply a color correction from camera RGB space to CIE perceptual space. – Chandler Feb 17 '17 at 20:59
4

tl,dr: No, you have to set up proper experiment and evaluate it using calibrated display/printer.

Spectral sensitivity is only one part of the whole process. Between the capture and the print things that alters the "image":

  1. Filters (UV, polariser, colour, neutral grey)
  2. Lens
  3. Sensor mask and sensitivity
  4. In-camera algorithms
  5. Postprocessing
  6. Print

Let's say we may neglect filters, postprocessing and print effects.

Still there may be differences among lenses and lens manufacturers regarding spectral transmissivity of their products.

Note that when the sensor is exposed the signal is analoguous - the higher charge in the well, the more light was captured. This signal is then digitalised and converted to RGB values. The digitalisation method and RGB conversion algorithm may differ among manufacturers.

  • Is there any way to compare the camera sensor chip only? not include the lens and filters, etc. Just compare how good the sensor can response to light – Zhao Huang Feb 16 '17 at 14:19
  • 1
    @ZhaoHuang Yes, but any differences are pretty meaningless with regards to photography since one must use at least a lens, sensor, demosaicing algorithms and other postprocessing steps to obtain a viewable image. – Michael C Feb 16 '17 at 18:52
  • 1
    @ZhaoHuang You already have that data presented with your question. Your two graphs compare the two sensors and their response across the visible spectrum. They tell you nothing about accuracy. You cannot use photography to accurately record colour; but, it's great for perceptual rendition. – Stan Feb 17 '17 at 3:27
  • I would repeat items 4 and 5 about three times each :-). Those are (one hopes) specifically designed to compensate for all differences between the Bayer filter spectral curves and the human eye's retinal spectral curves. (yes I know about the many magical neurological adjustments to perceived color.) – Carl Witthoft Feb 17 '17 at 13:16
1

You have the answer in the curves you present for comparison.

You can compare the response of one sensor to another directly at any common point to the two sets of data. For example, the efficiency of each of the sensors at 550nm can be given and compared. That's the only statement that can be made with the given data.

A statement about whether one sensor is "more accurate" than the other cannot be made. The only statement that can be made is already given (plotted)—the relative quantum efficiency.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.