RGB ratios based skin detection

Many different applications like face/people detection, image content interpretation, de-identification for privacy protection in multimedia content, etc. requires skin detection as a pre-processing step. There is no a perfect solution for skin detection, since this process is a compromise on speed, simplicity and precision (detection quality). There are many different techniques for skin detection modeling ranging from simple models based on one or several thresholds to advanced models based on neural network, Bayesian classifier, maximum entropy, k-means clustering, etc. This paper proposes a simple model, based on ratios of red, green and blue components of the RGB color model. It describes how to make a compromise in a skin detection modeling by using three levels of rules. Data analysis that supports conclusions is performed on the dataset from Universidad de Chile (UChile, dbskin2 - complete set) that contains 103 images and their annotations.