Abstract: To overcome the limitation of AdaBoost algorithm on Haar-like features,Turbo-Boost algorithm is proposed in this paper.Our proposed algorithm has a 2-iteration AdaBoost training framework.In the first iteration,An F-dimension principal feature subspace is selected.In the second iteration,a strong classifier constructed of T>F weak classifiers is trained in the F-dimension subspace.A 10 folds cross-validation on the CAS-PEAL-R1 facial expression database shows that Turbo-Boost outperforms AdaBoost significantly with a 93.6% overall precision on 5 categories of facial expressions including smiling,frowning,surprising,mouse opening,and eyes closing.Furthermore,Turbo-Boost algorithm is fast and suitable for real time applications.