Abstract: Focusing on the disadvantages of classical AdaBoost algorithm, this paper mainly analyzes the issue that the training time for classifiers is time-consuming and in training process the training results severely constringe the training samples (excessive training) and a new method is advanced to avoid the problems. The new method is to buffer the computational results of sorted feature values and update the training samples in time. As a result, using the method to train a cascade license plate, the experimental results show that the new method does not lead to the issue of excessive training and time consuming like classical AdaBoost often does, and moreover, the training time is shortened to 50 percent with a high detection rate and a low false alarm rate.