2
\$\begingroup\$

I am learning about light and color formation but I don't understand how does the sensitivity affects an image. Take for instance the different sensitivities for two "cameras" for each RGB channel

enter image description here

Would it be ok to say that camera one (blue plot) for each of the RGB sensors the brightness for colors around RGB is mostly uniform. Whereas for camera two (red plot) the brightness is concentrated only in a very short range in the RGB colors?

Is there more information (or probably I'm misinterpreting the graphs) that can be drawn from the plot? What would be the effects on an image?

\$\endgroup\$
4
  • 1
    \$\begingroup\$ What are the two axes? Wavelength and some measure of sensitivity? \$\endgroup\$
    – MikeW
    Sep 29, 2013 at 0:57
  • 1
    \$\begingroup\$ Seems odd (and undesirable) that a sensor would have 0 sensitivity to certain wavelengths, e.g. 600nm, if that's how to read the red plot \$\endgroup\$
    – MikeW
    Sep 29, 2013 at 0:59
  • 2
    \$\begingroup\$ Relevant on another level: xkcd.com/833 \$\endgroup\$ Sep 29, 2013 at 2:31
  • \$\begingroup\$ Yes x-axis is wavelength and the other is a measure of sensitivity. @MikeW I know is somehow odd but is only for learning purposes :) \$\endgroup\$
    – BRabbit27
    Sep 29, 2013 at 7:19

3 Answers 3

2
\$\begingroup\$

The subject is much more complex than the following "simplistic layman's answer"** , but ...

Simplistically, a sample of light at a point can be represented by a two dimensional position plus an amplitude, or by the amplitude of three orthogonal component vectors. It is traditional to use Red Green and Blue, approximately corresponding to the colour receptor wavelengths in most human eyes* but other component vectors could equally be used. (* Some eyes lack all or some sensitivity in one or more receptors and it is claimed by some that a few people have an additional wavelength receptor).

The "red" curve set lacks sensitivity on some areas - overall light measured would be less than the total.

The "blue" curve set overlaps very significantly. Light at a wavelength midway between the Red (right hand) and Blue (middle) pek response points would appear equally in both red and green channels and could not be distinguished in any one channel from monochromatic light of a slightly lower intensity.

I would expect that the narrow red curves would have some issues with low sensitivity and not dealing well with some light with distinct spectra peaks BUT
I would expect the wide blue curves to provide a munged pastel colour inaccurate mess.

Of the two I would expect the narrow curves to do better, but better still would be something with broader squarer non overlapping response curves which have both very little dead space and minimal overlap. Non-existeum selective interference filters would probably meet the need.

Many idea starters and examples here

What real people typically claim From here

enter image description here

BUT there are many variations depending on application. One "trick" is to use sharp squarush non overlapping filters plus a luminance channel that covers the whole spectrum.

A few other examples.

enter image description here


** Whether simplistic-answer or simplistic-layman left to reader's discretion.

\$\endgroup\$
1
  • \$\begingroup\$ I'd been led to believe that typical Bayer filters are not the simple, steep-edged, single bandpass filters that these graphs imply. Rather, it seems that all three Bayer filters pass considerable IR and UV, such that a "hot mirror" is needed in the sensor stack to keep IR from fouling the exposure. For this reason, it is possible to have "full spectrum" sensor conversions, which can be used with a variety of IR and UV filters for particular effects, as well as a number of specialized and ordinary visible-light filters. astrosurf.com/luxorion/Physique/spectral-response-ccd.jpg \$\endgroup\$ Dec 12, 2022 at 3:34
1
\$\begingroup\$

Its not uniform. uniform would be equal sensitivity to frequencies. None of those plots are desirable, albeit the blue one is more true in bayer pattern sensors. For technical shooting I would prefer the "red" (black spiky) plot.

The blue one has cross talk between the channel. If the subject has more green, both red and blue values go up. There is a calibration called "spectral sharpening" that will give the illumination of no crosstalk. It calibrates a simple linear combination where col_x= a1*col_x+a2*col_y+a3*col_z , etc.

Example, where I guessed values, that assume that red has a bit of green added to it, so I subtract a bit of green, and green has a bit of red and blue in it, so I subtract that, and blue has a bit of green. you see for example that the blue in the eyes pop out more and the yellow gets less blue, and the reddish blond hair gets a bit more of the red. The neck in the shade gets messed up though. Mixed lighting is difficult.

Example of manual spectral sharpening

The spiky plot has good separation between the channels. Just have to use it on things that distinguish themselves on those 3 wavelengths. And that often what we do, using monochrome cameras and adding a narrow band filter. This fictive camera you plotted is luxury as it has 3. Agricultural researchers' wet dream is to have three spikes like this on green, 680nm red (for red edge computations) and 780nm NIR. I know some that custom made a filter for the RGB camera that would "Sharpen" the spikes up for the green and red-edge and then use a split prism for NIR on to a monochrome camera.

\$\endgroup\$
1
\$\begingroup\$

Spectral Sensitivities of your sensor/lens combination have more of an effect on how much processing is done to an image before a jpeg or raw file is created. Today's camera's process the file for the look your desire given the camera profile (depends on brand). So camera manufacturers measure the spectral sensitivities of their sensors and create those profiles. We can do the same if we have access to some highly calibrated technical tools and have the programming abilities to create ICC profiles.

Reviewing your graphs, I can see you may have some mis-information. First let's start with relative human vision:

Color Matching Functions

So these are the relative measured wavelengths and intensities we humans can see. Some may notice I have removed the negative green node and plotted everything relative for comparison purposes.

Now we compare that with the spectral sensitivities of a CMOS sensor camera. In this case a Canon XTI I had measured a wile back:

EOS XTI Spectral Sensitivities

So clearly humans see differently than cameras. This difference is called the Luther-Ives condition and the result is a lot of noise in images and image processing to make up for our inability to manufacturer sensors that can satisfy that condition. The result in color is metamerisem. Because our human RGB vision inherently sees better and differently than camera sensors, and we have built in Auto white balance called chromatic adaptation. Cameras just can't mimic what God created very well at all.

So to obtain camera spectral sensitivities, you need some specialized equipment. There are a few ways to do this so I'll outline my method and others can chime in with their personal fav's: 1. Telespectro Radiometer or Monochrometer (I prefer a monochrometer) 2. A bench mount

Some foundational information: Three input channels are assumed for the spectral sensitivity data. The spectral data for the scene adopted white and illumination "perfect reflecting diffusers" must be absolute or normalized using the same factor. The camera spectral sensitivities are measured.

The RGB primaries for the white balancing of the training colors must be normalized so that equal amounts of the primaries combined together produces equal CIE XYZ values. The primaries may be selected to produce corresponding colors (simulating chromatic adaptation), but this is not required. Then we calculate the white balancing matrix from the scene adopted white XYZ to the destination color space adopted white XYZ. Then simply calculate the linear aim values. Then Calculate the linear camera response to the adopted white and the white balance channel multipliers. After that then we calculate white balanced linear camera response signals for the training colors. Then all that's needed is to calculate the color conversion matrix from white balanced linear camera RGB to the destination linear RGB based on least squares error in the destination color space with the gamma parameter nonlinearity.

Plotting the values in XYZ we see a pretty big difference between camera RGB and human vision. enter image description here

So what conclusions can we draw from this to make our art the best it can be? Well like everything in art, it depends on the goals of your specific project or the desires of the artist. If the goals are to be able to reproduce human vision, forget about it. But if you have a specific need that you must calibrate your camera to photograph and render accurately, then it may be possible using this method, depending on the exact spectral sensitivities of your particular camera &lens & lights. The plots show a CMOS sensor which will have very different sensitivities than the CCD array plotted above. Most good cameras today are CMOS, so CCD technology is relegated to come scanners.

So to sum it all up you can't overcome spectral sensitivity issues beyond the limits of the sensor, but you can modify the signal somewhat to compensate for a set of factors that remains static for the time you need it to. This can work for some artwork photography, and product photography to reduce color correction times and provide more efficient workflows for large catalog manufacturers. But the metameric effects of the difference between products given their dyes and pigments can still cause increases in time to color correct images to match.

So I hope this information proves helpful.

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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