In this paper, a new classifier training method and human face detection solution are proposed for multiface and multiview face detection problems which are prevalent in real life. Firstly, NPD (Normalized Pixel Difference) feature is used to describe the facial feature, the NPD feature describes the human face by judging the relative difference between the two pixel values, its eigenvalues can be obtained directly from the twodimensional table, which can greatly save the training time. At the same time, a kind of depth binary function tree structure is proposed to train the classifier, which can effectively combine the correlation between the features and combine the training with the skin color algorithm to improve the detection speed. Through the test of the proposed algorithm in this paper in CMU human face database, simulation results indicate that the algorithm improves the detection rate by 8.7%, error detection rate reduced 4.1%, detection speed improved 27.7% under multiface and multiview background.