An efficient training framework for gray-image face detection was presented. Our system includes two stages.In the first stage, the pattern rejection theory is used for features selection. The local Haar-like wavelet features used as rejection features to reject those patterns are not faces obviously. In the second stage, the Kullback-Leibler divergence in information theory is applied to choose more effective features further and to construct hierarchical classifier. The probability functions of two classes are estimated by joint-histograms. Final decisions are made according to the likelihood ratios between two classes. The experimental results show that our system is the same robust and efficient as the best reported methods, while the training efficiency is higher than others.