The task of face detection is extracting human face data information from image and video. The most widely used algorithm is AdaBoost. Considering the traditional AdaBoost algorithm has a too long training time problem, this paper has proposed an improved algorithm. By using specific trimming percentage, the algorithm trims the features with large error in classification in each round and adds the features which were not included in the last round into this round of training. When the error rate is over 0.5, Adaboost decreases trimming percentage dynamically. The experiment shows that this algorithm improves the training time and application scope of the algorithm, in contrast with Adaboost algorithm based on feature trimming. At the same time, from the view of reducing the characteristic number to participate in the training, the improved algorithm greatly reduced the cost of training time under the requirement of accuracy.