In this paper, two main improvements are achieved in AdaBoost according to practical requirements of face detection. One is that a self-adaptive cost-sensitive coefficient related with cascade classifier is introduced to treat the classification status of positive and negative examples differently. The other is that the weights are normalized separately for positive and negative examples after their weights updating steps in each boosting circulation. Experiments demonstrate that in face detection, the self-adaptive cost-sensitive AdaBoost shows higher detection rates and lower false positive rates. Moreover, the training time is less than that of the naive one.
展开▼
