A comparative study of color spaces in skin-based face segmentation

This paper presents a comparative study of five color spaces commonly used for detecting human skin. The evaluated models were: normalized RGB, HSV, YCbCr, CIE Lab, and CIE Luv. These color spaces attempt to separate the luminance from chrominance components, which is useful to make the face skin detection illumination independent. We used the Microsoft Kinect®sensor for acquiring 705 RGB images from 47 subjects in the age range from 18 to 45 years and distinct skin tones. Besides, each image was segmented manually to define true skin pixels. A probabilistic classifier was built for each tested colorspace to classify a pixel color into skin class or non-skin class. The Matthews correlation coefficient (MCC) was used to evaluate the quality of the computerized skin classification. The results pointed out that the CIE Lab colorspace reached the best MCC performance with median value equal to 0.779 and Q_n estimator equal to 0.074. The worst performance was attached by normalized RGB with with median value equal to 0.606 and Q_n estimator equal to 0.143.