We present a method for face detection which uses a new {SVM} structure trained in an expert manner in the eigenface space. This robust method has been introduced as a post processing step in a real-time face detection system. The principle is to train several parallel {SVMs} on subsets of some initial training set and then train a second layer {SVM} on the margins of the first layer of {SVMa}. This approach presents a number of advantages over the classical {SVM}: firstly the training time is considerably reduced and secondly the classification performance is improved, we will present some comparisions with the single {SVM} approach for the case of human face class modeling.