a and b channels in CIELAB color space

I am trying to convert an image from RGB (uint8,uint16) to LAB color space , I want to rescale it to [0 1] based on the max and min value of the color space not the image itself.. I would like to know the range that is used when I use rgb2lab..

in the documentation, it is mentioned that it can be [-100, 100] or [-128, 127) for a* and b* channels,

the question is how to know which one is used so I can use it and why these two different ranges??

ref: https://www.mathworks.com/help/images/ref/rgb2lab.html

Any references or explanation will be appreciated,

Answer 1

[Image Analyst]

There is no max and min for the colorspace. a and b do not have defined range, like L (which represents percent reflectivity). I don't believe there is any need for you to rescale your a and b images. Why do you want to do that?

[MatlabUser]

for clustering, the used clustering features ranges are widely different and I need to rescale them to avoid beeing biased to the largest range.

[Image Analyst]

1. What clustering features are you using? The L values, a values, and b values?
2. Can you use colorcloud() and show me what your gamut looks like? Attach screenshot.
3. What are these "ranges" you have defined in advance, and why would those ranges bias where the cluster centers are found, or what cluster each data point belongs to?
4. What clustering algorithm are you using (kmeans, knn, discriminant analysis (see attached demos), or decision trees)?

[MatlabUser]

Thank you for replying @Image Analyst:

1. I have ranges [0,50], [0,100], [ a, b ranges], [ -3,3]. for position, color, and other feature that has range [-3,3].
2. I attached screenshots for RGB before convert to LAB
3. I mentioned in point 1 the ranges, I noticed from the results the clustering is biased to the position feature, and the feature with range [-3,3] can not affect the results with its small range (unless I multiply by a large weight value).
4. K-means (manual implementation not matlab function), and I read many articles saying since K-means is based on distance metric then the features should normalized to have same ranges.

Note: when I print the max and min value it is always within [-100 ,100] for 6 different images, but for two images b channel can reach to -101.2, so I am wondering if -100,100 is the range that I can use?

[Image Analyst]

I don't believe you need to do that, though you can. That is necessary if for example one of your features is in the range of 1 million to 2 million, and another feature ranges from 0.01 to 0.1 - several orders of magnitude different such that you'd have a problem with truncation or roundoff errors. You don't have that situation. Yours are all in the range of -100 to +100 or so, which is not disparate enough to cause the problem. I don't see that normalization would be needed. That said, if you still insist on using it, you can call rescale on each column before doing kmeans on it, then undo the scaling on the centroid locations you found to get back to the original space.

But looking at your gamut, I don't see any real clusters that kmeans would do a good job at finding. You can see my demo and see that kmeans doesn't do such a great job with RGB images that don't have well defined clusters in the gamut (like yours).