|
|
|
| Fast Training of Boosting Cascade Based on Information Sharing |
| Yan Sheng-ye |
| School of Information and Control, Nanjing University of Information Science and Technology, Nanjing 210044, China |
|
|
|
|
Abstract It always takes a very high computational cost to build a Boosting cascade classifier. To speed up the training procedure of the Boosting cascade, this paper proposes an extended training method which utilizes the sharing information among different stage classifiers. The proposed method takes advantage of the sharing information among different stage classifiers at two levels. First, at the classifier level, the last stage classifier is taken as the first feature, and is re-used to learn a new weak classifier to adapt the newly collected training samples of the current stage. Secondly, at the feature level, all the selected features from all the previous stage classifiers are re-used to learn new weak classifiers which adapt to the newly collected training samples of the current stage. Finally, the newly learned weak classifiers and newly selected features are added in the current stage classifier. The experimental results on frontal face detection show that the proposed method improves the training speed of the Boosting cascade classifier largely. The training speed of the proposed method is about 10 times faster than that of the traditional method. To be exact, the training time of a frontal face detector is reduced from about 3 days to 8 hours.
|
|
Received: 02 December 2013
|
|
|
|
Corresponding Authors:
Yan Sheng-ye
E-mail: shengye.yan@gmail.com
|
|
|
|
|
|
|
2014, Vol. 36 
