2023年8月2日发(作者：)

龙源期刊网

一种改进的AdaBoost检测算法

作者：刘苹光 文成玉 杜鸿

来源：《计算机应用》2015年第08期

摘要：针对传统AdaBoost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AdaBoost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AdaBoost算法；弱分类器数目为116，比传统AdaBoost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。

关键词：AdaBoost算法；正误差；负误差；阈值；人脸图像库

中图分类号： TP391.4

文献标志码：A

Improved detection algorithm of AdaBoost

LIU Pingguang*， WEN Chengyu， DU Hong

College of Communication Engineering， Chengdu University of Information Technology，

Chengdu Sichuan 610225， china

Abstract： 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.