In view of the problem of the long training time in dealing with large training dataset in the training process of the traditional Adaboost algorithm, the authors introduced an improved methods to these problem. Improved algorithm using PCA dimension reduction technique, extracts main ingredients for the training sample feature, removes the correlation between the input sample characteristics, and improves the classification accuracy. At the same time, from the angle of sample threshold search takes into consideration the divisions and eigenvalue space dimension, threshold fast search method is presented. Experimental results show that the algorithm to achieve better results on UCI datasets.