This paper present an analysis of the performance of supportvector machines (SVMs) for automatic detection of human faces in staticcolor images of complex scenes. Skin color-based image segmentationsinitially performed for several different chrominance spaces by use ofthe single Gaussian chrominance model and a Gaussian mixture densitymodel. Feature extraction in the segmented images is then implemented byuse of invariant orthogonal Fourier-Mellin moments. For all chrominancespaces, the application of SVMs to the invariant moments obtained from aset of 100 test images yields a higher face detection performance thanwhen applying a 3-layer perceptron neural network (NN), depending on asuitable selection of the kernel function used to train the SVM and ofthe value of its associated parameter(s). The training of SVMs is easierand faster than that of a NN, always finds a global minimum, and SVMshave a better generalization ability
展开▼
