I once evaluated different color descriptors for object recognition [1]. One of the best performing descriptors where the "Color Names" by van de Weijer et al. [2], and their successor "Discriminative Color Descriptors" by Khan et al. [3]. Both algorithms assign the Lab values of image pixels to one of 11 basic color terms (black, red, green, yellow, blue, brown, purple, pink, orange, grey). The codebook for this assignment was learned on a large set of images retrieved by Google queries. In result, they are implicitly invariant against changes in illumination, lighting geometry [1]. If clustered using Divisive Information Theory, they further increase their discrimination ability for recognition problems [3,1].
Using such descriptors, you can easily quantify the proportion of "yellow" and "brown" in your images. BTW: the C implementations of these algorithms are available with MATLAB bindings [4,5].
TLDR:
- convert your pixel values into color names [4,5]
- count the pixels per color
Have fun!
[1] Article Plant species classification using flower images—A comparati...
[2] Article Learning Color Names for Real-World Applications
[3] Conference Paper Discriminative Color Descriptor
[4] http://lear.inrialpes.fr/people/vandeweijer/color_names.html
[5] http://www.cat.uab.cat/~joost/ColorDescriptors.html
Amin Jaberi
RGB is a correlated color space. You'll face problems defining proper RGB ranges for distinct colors. Euclidean distance metrics will just fail in RGB space. In addition, the perception and representation of color heavily depends on illumination source, lighting geometry, and the spectral response of the capturing devices.Hence, instead of RGB, you might define proper ranges of the hue in HSI colorspace, which is decorrelated, or a and b in Lab color space, which is decorrelated and Euclidean. I'd prefer Lab, as the hue in HSI is unstable along the grey axis (low saturation).
- Badges
- Science topic