Gaussian$-$based Color Segmentation

Although Gaussian$-$based color segmentation is compact in terms of color model, it is computationally more intensive than the histogram$-$based model. Still, you can try this approach and compare with the previous one. You need to

$*$ Collecting flesh tone training data. A couple of images are provided. It is very easy to find on the Web some color images contain skin color $($people$).$ You can crawl such images from the Web. Then, you should design a tool to collect skin pixels. Of course, a simple interface will do$.$ E$.$g$.,$ you can use imcrop or ginput in Matlab. If you are using C$++,$ you have to select pixels or small regions using a mouse.

$*$ Selecting a good color space. You$$d better try a couple of color spaces, e$.$g$.,$ RGB$,$ N$-$RGB$,$ HSI, etc, before you make your decision.

$*$ Train a Gaussian color model. You just need to estimate the mean and covariance of your training data as the parameters of the Gaussian distribution.

$*$ Find skin tone based on the Gaussian distribution. It is quite straightforward. Note: you need to set a threshold.