Performance analysis of skin classifiers in RGB and YCbCr color space

Skin detection serves as a preliminary step for number of applications like face detection, gesture recognition, internet pornographic image filtering, and surveillance system. Number of artificial neural network (ANN) based skin detection algorithms have been presented in literature which are mostly based on back propagation (BP) ANNs. This paper attempts to analyze the performance of skin classifiers using AdaBoost learning algorithm in both RGB and YC_bC_r color space. Three RGB based classifiers (i.e., red, green, and blue) and one YC_bC_r based classifier is designed in order to analyze the performance of algorithm for each case. Set of weak heuristic rules are designed for the classifiers to reduce the false positive rate (FPR) without significantly affecting the correct detection rate (CDR). The results reveal that the best performance is achieved by RGB based classifiers with heuristic rules in terms of both accuracy and processing time. Without heuristic rules the best results have been provided by Y-classifier. The classifiers are trained and tested using SFA database. The classifiers are also tested by using images of FERET and CVL database.