基于特征剪裁的AdaBoost算法及在人脸检测中的应用

针对AdaBoost算法存在训练消耗大并且误检率较高的问题,提出一种基于AdaBoost的高效检测方法.它主要包含一种基于特征剪裁的AdaBoost算法（FPAdaBoost）和一种新的检测扫描方法--确认和跳过检测机制（CSDS）.FPAdaBoost算法在每一轮训练中会根据分类误差剪裁掉一部分特征,提高算法的训练速度；而CSDS检测方法在传统的检测方法基础上引入验证和确认机制,在保证检测率的条件下有效控制误检的发生.在MIT CBCL训练集和MIT+CMU检测集上对提出的方法进行验证,结果表明,FPAdaBoost算法相比原始AdaBoost算法在性能上没有明显退化,但却大大改善了训练速度,同时CSDS检测机制的引入极大地降低误检率,提高检测结果的可靠性.

Feature pruning based AdaBoost and its application in face detection

The AdaBoost algorithm is highly computational consuming and has high false positive rate. To deal with these problems, an efficient detection method based on AdaBoost,which consists of a Feature Pruning based AdaBoost (FPAdaBoost) algorithm and a confirmation and skipping detection scheme (CSDS), is presented in this paper. FPAdaBoost cuts off features at a certain cutting coefficient according to the classification error in each iteration of training process, which effectively speeds up the learning process and greatly reduces the computational cost. And CSDS employs verification and confirmation scheme in the conventional scanning process, which effectively eliminates false positive detections. The performance of proposed detection method was tested in face detection using the MIT-CBCL training set and the MIT+CMU test set. The results show that, compared with traditional Adaboost detection method, the training time of FPAdaBoost dramatically decreases without suffering a decline in classification capability, meanwhile the false positive rate is significantly reduced due to employing CSDS in the scanning process.