Towards independent color space selection for human skin detection

Skin color detection plays an important role in video based applications. Without considering the selection of suitable color space, a novel skin color detection method is proposed based on the flexible neural tree, which can identify the important components of color spaces automatically. With large training data sets, our method builds a flexible neural tree structure and optimizes its parameters using Genetic Programming and Particle Swarm Optimization algorithms. In experiments, features comprised of all channels extracted from RGB, YCbCr and HSV color spaces are used for the constructing and evaluating of the novel skin color model, in which six most important components, i.e., R, G, B, Y, Cr and S are selected for testing. Furthermore, our method achieves higher accuracy and lower false positive rate than state of the art methods on Compaq and ECU data set.