Dear Amir,

First of all, it is an interesting question. From a physical point of view, each  color channel represents a particular wave length or frequency and each frequency is sensitive a particular material or physical properties of the reflective radiation. In remote sensing, we do have more than the 3 channels (R,G,B) and each channel is connected with a particular wave length.

Regards,

Gamal

Dear Amir,

Why RGB? ... because we have red,green and blue sensitive photo

receptors in our eyes!

The RGB colour space is based upon the portion of the electromagnetic spectrum visible to humans (i.e. the continuous range of wavelengths in the approximate range(400–700 nm). The human eye has three different types of colour receptor over which it has limited (and nonuniform) absorbency for each of the red, green and blue wavelengths. This is why, the colour to grey-scale transform uses a nonlinear combination of the RGB channels.

Although a grey-scale image contains less information than a colour image, the majority of important, feature related  information is maintained, such as edges, regions, blobs, junctions and so on.Feature detection and processing algorithms then typically operate on the converted grey scale version of the image

Please refer,

The Image Processing Handbook, Sixth Edition,John C. Russ

CRC Press,ISBN 9781439840450  pp 49-52

http://www.cse.iitb.ac.in/~sharat/icvgip.org/icvgip00/I-71.pdf

http://www.pasteur.gr/wp-content/uploads/Basics-of-Quantitative-image-analysis.pdf

Regards,

VK

Thanks a lot! insightful.

Best regards,

Amir

Amir, I think the better question is what do you hope to achieve by analyzing a RGB image.  Keep in mind that RGB is fundamentally just a another three dimensional space: (R, G, B).  It is not uncommon to use RGB colors to extend representable data from the normal three spatial dimensions into a 6D space defined as (x, y, z, r, g, b).  Not all forms of information can be represented this way, but a surprising number of applications can gain from such an extension.  Note, this is in contrast to a black and white image that only has one color channel: dark.  In two spatial dimensions, this translates into a 3D space such as (x, y, intensity).

In terms of video and image processing, it has long been known in the machine learning community that the higher the dimensional space of your data the easier it is to detect relevant information.  This is reflected in biology in that certain plants use different wavelengths to signal to different species about the ripeness or desirability of their fruit and nectar.  By using different wavelengths, plants can hide their ripeness information from species that may not be as effective as spreading their pollen or dispersing their seeds as the targeted species.  Similarly, information that you are interested in may be hidden from you if you are only looking at a black and white image.

But rgb isn't just about cluster or object identification.  rgb can also be used to determine the chemical composition of material.  For example, both astronomy and chromatographic spectroscopy utilize color to determine the chemical composition of planets and material samples, respectively.

So, why use rgb? ... well, what are you interested in doing?

The reason why I ask "what are you interested in doing" is two fold.  The first is that while rgb adds information it may not add relevant information, and each new dimension that you add to your problem will increase the computational complexity of your problem.

The second reason is that the use of rgb is such an open question it is hard to know whether it is a good or bad idea to use it without knowing the nature of the problem.  There is always a benefit and a cost.

Color can tell you a few things about an image. Pixels can be analyzed to show the intensity of light that is being portrayed, as well as the amount of contrast is in the image. Generally, 3 things you can look at are Energy, Entropy, and Contrast when dealing with an image. It can also tell you how similar two images are to each other, based on distance between pixels (usually a Manhattan distance). 

Content-based Image Retrieval systems are fairly common and easy to implement, although not always the best for comparing images that are similar, as it only tells you information about how similar the photos are in color and texture, not in actual content.

https://en.wikipedia.org/wiki/Content-based_image_retrieval

I agree with Jeffrey Holcomb, you need to look at what problem you are trying to find a solution for to see if RGB is right for you. You may also look at contrast or object recognition to find a better result, depending on the problem.

Dear Jeffrey and Blair,

Thank you so much for your explanations. 

To be more specific, I'm hoping that by using RGB images or any other kind of images, taken by a digital camera, i would be able to judge/understand scattering and absorption properties of tissue (I'm imaging human tissue in vivo using digital cameras). But still not sure how cand I get a rough idea about scattering and absorption properties of tissue by analayizing these kinds of images?

I could get my required data form spectrum coming out of the tissue but how about photo? How can I judge optical properties using that?

Thanks again

That is a definite yes.  But before I just start answering I should probably ask if you are asking a spectroscopic question or a image correction question.  It sounds like you are asking a spectroscopic question, in which case I am about half way through the implementation of a method that could compute a scattering/absorption signature for the tissue.  Not sure when I will finish the other half though; that part is not really my focus at the moment, and I am looking at things that are more generic than tissue.

My guess is that the key word that you want to look for here is image segmentation.  I'm not sure if the attached publications are the best examples of color-based image segmentation, but they were the quickest that I could find while flipping through my library of saved papers.  I think all of these were from WACV; which tends to have a lot of plant and ant image segmentation for some reason.

After you segment the image into parts, then you will need to perform various computations related to optics to get detailed refraction/absorption information.  Even cooler, if you have the resources to co-locate the transmitter with respect to the various sensors then you ought to be able to reconstruct a fairly detailed image of the tissue that you are passing light through (which is probably what you were thinking).

If you are asking an image correction question then you probably want to look into things like glare correction, or removing "red eyes", or anything of that nature.

Conference Paper Segmenting color images into surface patches by exploiting s...

Article ClassifiÞcation of plant structures from uncalibrated image sequences

Conference Paper Image segmentation for large-scale subcategory flower recognition

Conference Paper Automatic region-of-interest detection and prioritisation fo...

Depending on your problem, other color spaces such as HSV can be more informative than RGB. This is usually the case.

Continuing on alternative color spaces, HSV and other spaces result from linear transformations of the RGB space with fixed coefficients. This means that the representation obtained does not depend on the image content. Various strategies of exploratory data analysis can provide transformations tailored for the specific application/content, possibly exploiting known properties of the data at hand. This paper reports an example (a little bit far from human tissue classification).

http://www.alslab.net/WPMH/wp-content/uploads/2014/09/4.Analysis-of-Multispectral-Ima ges-in-Cultural-Heritage-and-Archaeology.pdf

Other examples in these papers. Look at the bibliographies for some theory behind these techniques.

http://dx.doi.org/10.1016/j.sigpro.2007.07.011

http://dx.doi.org/10.1007/s10032-004-0121-8

http://dx.doi.org/10.1186/2050-7445-1-33

http://dx.doi.org/10.1016/j.culher.2012.11.013