Abstract

Human skin color distribution is relatively compact in a color space, and it changes from frame to frame following illumination variations, also it is subject to camouflage interruption in complicated background for skin segmentation. This paper mainly talks about face detection based on skin color feature in complicated background. We assume that skin pixels in each frame are closed together as a “dot cloud” in a color space, its shape evolution from frame to frame is modeled as the mixture of translation, scaling and rotation. We introduce linear combination of forecasts to predict these parameters related to the above shape evolutions, so that skin distribution of next frame to be segmented can be predicted and skin segmentation for face detection can be improved against illumination variations. Additionally human skin biological feature is then introduced to remove camouflage noise. For face gender recognition we adopt 2-D Gabor transform for feature extraction and SVM for recognition. Extensive tests prove that this algorithm is quite sensitive to human color, and more accurate for human skin segmentation with Bayes classifier; Good performance of gender classification test is also achieved on a relative large scale and low-resolution video database.