In order to recognize human face by machine, it is necessary to firstly detect human face regions from images. It is not an easy problem because human face is nonrigid. We propose a novel face detection algorithm by selecting useful features from high dimensional features and generating new ones. Our algorithm is composed of the following steps; 1. Many kernel features are generated based on Kullback-Leiblber Divergence: 2. A boosting algorithm selects some useful features for face detection: 3. Steps 1 and 2 are performed iteratively. Our algorithm achieves almost equal or better detection rate than that of a Support Vector Machine (SVM). It also achieves nearly one-tenth calculation cost of the SVM.