摘要
针对传统Ada Boost算法在人脸图片训练过程中可能会出现退化现象和训练目标类权重分布过适应的问题,提出一种基于调整正负误差比和设定阈值的改进Ada Boost算法。该算法首先把设定的阈值和当前分类错误样本的权值比较来更新样本的权值,其次通过调整正误差和负误差之间的偏重关系来控制训练样本的偏重。经过实验表明,不同人脸图像库和不同正负样本比不影响该算法的有效性,在LFW非受限人脸图像库正负样本比例为1∶1情况下,检测率为86.7%,高于传统Ada Boost算法;弱分类器数目为116,比传统Ada Boost算法多15个。实验结果可以看出所提算法抑制了退化和训练目标类权重过适应现象,有效地提高了人脸图片检测率。
Considering the degradation and problem that the weight distribution of training targets is wider than average in the traditional AdaBoost algorithm in the process of human face image training, an improved AdaBoost algorithm was proposed based on adjusting margin of error and setting the threshold value. First, the weight values of the samples were updated according to the comparative result between the threshold value and the weight value of the matching errors of the current samples. Then, the emphasis of the training samples was controlled by adjusting the emphasis relation between positive error and negative error. The experimental results showed that different human face image databases and different ratios of positive and negative errors had little effects on the validness of the improved AdaBoost algorithm. Under the positive and negative error ratio of 1：1 in unrestricted face database LFW, the detection rate was 86.7%, which was higher than that of the traditional AdaBoost algorithm; the number of weak classifiers was 116, which was 15 more than that of the traditional AdaBoost algorithm. The results prove that the proposed algorithm suppresses the degradation and the problem that the weight distribution of training targets is wider than average, and effectively improves the detection rate of human face images.
出处
《计算机应用》
CSCD
北大核心
2015年第8期2261-2265,共5页
journal of Computer Applications
基金
四川省教育厅成果转化重大培育项目(13CZ0012)
四川省科技厅项目(2012zz001)
四川省教育厅项目(12ZB201)
