Multiple color space channel fusion for skin detection

Skin detection is used in applications ranging from face detection, tracking body parts and hand gesture analysis, to retrieval and blocking of objectionable content. We investigate color based skin detection. We linearly merge different color space channels representing it as a fusion process. The aim of fusing different color space channels is to achieve invariance against varying imaging and illumination conditions. The non-perfect correlation between the color spaces is exploited by learning weights based on an optimization for a particular color space channel using the mathematical financial model of Markowitz. The weight learning process develops a color weighted model using positive training data only. Experiments on a database of 8991 images with annotated pixel-level ground truth show that the fusion of color space channels approach is well suited to stable and robust skin detection. In terms of precision and recall, the fusion approach provides a competitive performance to other state-of-the-art approaches which require negative and positive training data with the exception of the decision tree based classifier (J48). As a real-time application, we show that the weight based color channel fusion approach can be used for learning of weights for skin detection based on detected faces in image sequences.