应用权重直方图的DW-AdaBoost快速算法

Fast DW-AdaBoost Algorithm Using Histogram of Weight

提出一种改进的AdaBoost算法,提高人脸检测的训练速度,以及检测速度和精度。先将每个Haar?Like特征下所有样本的特征值量化,然后据此分别计算出人脸和非人脸样本,再快速计算出简单分类器的阈值和偏置。分析样本特征值的分布特性,进一步提出了双阈值快速算法。在MIT-CBCL训练库上对算法进行了验证,结果显示基于权重直方图的双阈值AdaBoost算法—DW-AdaBoost的训练速度提高150多倍,收敛速度更快。在MIT+CMU人脸测试库上进行了测试,结果表明该方法在检测精度和速度等方面都优于相应的单阈值方法。

An improved AdaBoost algorithm named DW-AdaBoost is developed in face detection. Firstly, the feature values of all training samples with each Haar-like feature are quantified into [1,100]. Then the weight histograms of face and nonface are computed by using these quantified feature values. The threshold of simple classifier with each feature is computed faster based on the weight histograms. According to the distribution of feature values of face and nonface, dual-threshold is defined and computed. Experimental results on MIT-CBCL face and nonface training data set illustrate that the improved algorithm DW-AdaBoost could make training process convergence quickly and the training time is only one of 150 like before. Furthermore, experimental results on MIT＋CMU using the detectors also show that the detection speed and detection precision under dual-threshold all exceed the corresponding method.