We describe a component based face detection system trained only on positive examples. On the first layer, SVM classifiers detect predetermined rectangular portions of faces in gray scale images. On the second level, histogram based classifiers judge the pattern using only the positions of maximization of the first level classifiers. Novel aspects of our approach are: a) using selected parts of the positive pattern as negative training for component classifiers, b) The use of pair wise correlation between facial component positions to bias classifier outputs and achieve superior component localization.