A histogram matching is a standard method of how to adjust color space from one image to another. For gray-scale images it is straight forward, but how to deal with color images? So far I found it done the same way per color channel independently like skimage - https://scikit-image.org/docs/dev/auto_examples/transform/plot_histogram_matching.html
I am not very convinced that the color information can be treated independently per channel at least not for all color spaces; for example RGB, personally some smoother color space would be better like CIE lab?
A model case would be having two color images, one would be only green other only red... the histogram matching independently per color channel does not make much sense, right?
Well, I think you can treat it each channel independently. You have a RGB reference image with different histogram pattern in each channel, so I think it makes sense adjust each channel of the source image to the reference image. I think it would be problematic if you have a grey reference image.
Well, I am talking about the colour enhancement of two colour images, for example, one would be only green other only red... the histogram matching independently per colour channel does not make much sense, right? You would try to green channel with values to zeros and the same for zeros red to some values...
The McAdam's Ellipse is a tricky thing , it is meant for generation. For matching I would think transformation to YCbCr or HSL and then correlate to CIE color coordinates and back to RGB , so that pink stays pink and does not become red (I have not done this practically but I am speculating)
My take on this problem is to consider the phenomena of proportionate and scalar value in any given colour comparison as fundamentally distinct elements of measurement. When the colours in two images are to be compared I would map the pixel RGBs onto a standard reference space such as Photoshop sRGB. Once you have a stable reference-basis you then normalize the two datasets onto a common lightness scale. The next step is to analyse their proportionate Hue Chroma colour differences at constant lightness. Again in Photoshop, you can use CIE Standard Observer data to quantify the colour differences three-dimensionalal in CIE L*a*b* colour space with its Hue circles and Chroma vectors that radiate out from a point on the grey scale .
I found also this discussion: Which color spaces are useful for histogram matching? https://github.com/mapbox/rio-hist/issues/3
There must be a reason why you want to perform histogram matching. I assume that you want to compensate for some color variation before you can compare two samples.
If the color variation is due the lighting conditions, I would think of a color twist matrix (if the light remains achromatic, histogram matching on Lightness alone should do). If it is due to the inherent color of the tissues, I guess that it would be interesting to observe how the colors in one image map to the other, in RGB space. This can probably be imaged as a deformation field and should give you some insight in the appropriate transformation.
Which color space is best, mainly depends on your object(s) of interest:
if the information available is independent of the color, matching of the L (from HSL)/B (from HSB)/Y (from YCbCr resp. YUV) artefacts might suffice. Differences in brightness would be easily adjustable where necessary.
Matching R with e.g. G might make sense in some applications, no sense in others. (I once had imaging where details could only be seen with green illumination; red illumination would have completely masked the information of interest.)
Si it's the prior knowledge (required) that dictates what and how to match - there is no such thing as a general rule.
- Badges
- Science topic