An Optimal Chrominance Plane in the RGB Color Space for Skin Color Segmentation
Cheol-Hwon Kim, Juneho Yi, Dongan-gu Anyang · 2006
In the case of illumination or pose changes, the distribution of skin colors in color planes is very variant. This makes it hard to employ color histogram based approaches to skin color segmentation. We propose online PCA analysis of skin colors directly in the RGB color space. Our idea is simple but works well. Based on the observation that skin colors of a person in the RGB color space are approximately distributed in a linear fashion, we apply PCA techniques to RGB values of skin colors from a set of training images collected online. The axis along which the data is spread most corresponds to the axis of illumination intensity. We use the remaining two axes to form a chrominance plan in which the distribution of skin colors is invariant. We have preprocessed the input images using a retinex algorithm in order to compare the color constancy property of the proposed plane with the HS and IQ planes. As we expected, the retinex algorithm significantly changed the distribution of skin colors in the HS and IQ planes. The experimental results show that the performance of skin color segmentation using the proposed plane is much more robust to illumination and pose changes compared to those using HS or IQ planes.