# Image features (spectral bands) other than RGB for image analysis

What features can I derive from the image, other than its RGB, to help me analyze an image?

I know I can use some features like its lightness, etc. But I don't want a single value as lightness, I want a 2d array just like R or G or B.

• "I want a 2d array" -> I think the word you're looking for is parametric map. Many filters exist that can produce such maps, e.g. Gabor, textures, colorspaces. Commented Jun 3, 2019 at 10:22

I think what you are looking for is called a colour space. There are multiple available, and each of them was created to solve a certain problem.

The RGB colour space is the most common I guess, since that is close to the way the computer displays show information (although they do apply some transformations before lighting the pixels). It is also similar to the way we see light (we have red, green and blue receptors).

Light is additive. If you keep adding colours, you get to white. Paint is subtractive. That leads to some differences into how colours mix and lead to the creation of the CMYK standard for printing. It represents colours in terms of cyan, magenta, yellow and black.

Perhaps more important for you is the HSB color space. It is rather nicely shaped, as a 3D cylinder. H stands for hue, the colour nuance, S stands for saturation, how pronounced is the colour and B stands for brightness. HSL is a similar color space, where the vertical axis is lightness instead of brightness. I find these two color spaces quite useful in analysis.

The RG cromacity space is used to analyse the color itself, and discard the brightness information. You basically represent the amounts of red and green from a normalized color (a color where the sum of r g b is 1). This way, you operate in a simpler, 2D space, and analyze the color information. See [1], [2].

The CIE colorspace is based on measurements of our color perception. It tries to mimic the way our photoreceptors react to color. The original CIE colorspace, known as CIEXYZ, was published in 1931. It has been revised in 1976 to form the CIELab colorspace, which is perceptually uniform. This means that it accounts for our non-liniar perception of color, making sure that the same increase produces the same perceptual difference across the color space. See [3], [4], [5].

The TSL colorspace (tint, saturation, lightness) was proposed in [6] for face detection.

There are many other colorspaces, but these are the ones I found most useful. There are usually conversions across different color spaces.

A publication by Logvinenko, 2014 might help with understanding the geometry of a color space and choosing a colorspace that helps your analysis [7]. Also, have a look at [8].

• Thank you so much. Your information was really useful. Do you know of any relevant papers? Commented Jun 3, 2019 at 0:38
• I've added a bunch of publications. Hope it helps. Commented Jun 3, 2019 at 10:01