基于拟合型弱分类器的AdaBoost算法

doi:10.11959/j.issn.1000-0801.2019219

电信科学 ›› 2019, Vol. 35 ›› Issue (11): 27-35.doi: 10.11959/j.issn.1000-0801.2019219

基于拟合型弱分类器的AdaBoost算法

宋鹏峰,叶庆卫,陆志华,周宇

宁波大学信息科学与工程学院，浙江宁波 315211

修回日期:2019-09-12 出版日期:2019-11-01 发布日期:2019-12-23
作者简介:宋鹏峰（1995- ），男，宁波大学信息科学与工程学院硕士生，主要研究方向为计算机视觉与图像处理|叶庆卫（1970- ），男，博士，宁波大学信息科学与工程学院教授、硕士生导师，主要研究方向为信号检测、最优化搜索、视频识别与跟踪等|陆志华（1983- ），男，博士，宁波大学信息科学与工程学院讲师，主要研究方向信号处理、多运动目标的实时跟踪、统计信号处理算法研究和应用等|周宇（1960- ），男，宁波大学信息科学与工程学院教授、硕士生导师，主要研究方向为信号处理、网络与信息安全、物联网技术等
基金资助:
国家自然科学基金资助项目(51675286);国家自然科学基金资助项目(61071198)

AdaBoost algorithm based on fitted weak classifier

Pengfeng SONG,Qingwei YE,Zhihua LU,Yu ZHOU

College of Information Science and Engineering,Ningbo University,Ningbo 315211,China

Revised:2019-09-12 Online:2019-11-01 Published:2019-12-23
Supported by:
The National Natural Science Foundation of China(51675286);The National Natural Science Foundation of China(61071198)

摘要/Abstract

摘要：

针对AdaBoost算法通过最小化训练错误率来选择弱分类器造成的精度不佳问题以及单阈值作为弱分类器训练过程较慢难以收敛问题，提出了一种基于拟合型弱分类器的AdaBoost算法。首先针对每个特征，在特征值与标记值之间建立映射关系，引入最小二乘法求解拟合多项式函数，并转换成离散分类值，从而获得弱分类器。其次从获得的众多弱分类器中，选择分类误差最小的弱分类器作为本轮迭代的最佳弱分类器，构成新的 AdaBoost 强分类器。与传统训练算法相比，极大地减少了待选弱分类器的个数。选取 UCI 数据集和MIT人脸图像数据库进行实验验证，相较于传统Discrete-AdaBoost算法，改进算法的训练速度提升了一个数量级，人脸检测率可达96.59%。

关键词: AdaBoost, 拟合型, 最小二乘法, 弱分类器

Abstract:

AdaBoost algorithm was proposed to minimize the accuracy caused by weak classifiers by minimizing the training error rate,and the single threshold was weaker and difficult to converge.The AdaBoost algorithm based on the fitted weak classifier was proposed.Firstly,the mapping relationship between eigenvalues and marker values was established.The least squares method was introduced to solve the fitting polynomial function,and the continuous fitting values were converted into discrete categorical values,thereby obtaining a weak classifier.From the many classifiers obtained,the classifier with smaller fitting error was selected as the weak classifier to form a new AdaBoost strong classifier.The UCI dataset and the MIT face image database were selected for experimental verification.Compared with the traditional Discrete-AdaBoost algorithm,the training speed of the improved algorithm was increased by an order of magnitude.And the face detection rate can reach 96.59%.

Key words: AdaBoost, fitting type, least squares, weak classifier

中图分类号:

TP391.4

宋鹏峰,叶庆卫,陆志华,周宇. 基于拟合型弱分类器的AdaBoost算法[J]. 电信科学, 2019, 35(11): 27-35.

Pengfeng SONG,Qingwei YE,Zhihua LU,Yu ZHOU. AdaBoost algorithm based on fitted weak classifier[J]. Telecommunications Science, 2019, 35(11): 27-35.

图/表 10

图1

表1

图2

表2

表3

图3

表4

图4

表5

图5

参考文献 13

[1]	唐彪, 金炜, 符冉迪 ,等. 多稀疏表示分类器决策融合的人的脸识别[J]. 电信科学, 2018,34(4): 31-40.
	TANG B , JIN W , FU R D ,et al. Face recognition using decision fusion of multiple sparse representation-based classifier[J]. Telecommunications Science, 2018,34(3): 31-40.
[2]	LI X , WANG L , SUNG E . AdaBoost with SVM-based component classifiers[J]. Engineering Applications of Artificial Intelligence, 2008,21(5): 785-795. doi: 10.1016/j.engappai.2007.07.001
[3]	YI X , YING W , PENG J . An improved AdaBoost face detection algorithm based on the weighting parameters of weak classifier[C]// IEEE International Conference on Cognitive Informatics＆ Cognitive Computing,July 16-18,2013,New York,NY,USA. Piscataway:IEEE Press, 2013.
[4]	徐凯, 陈平华, 刘双印 . 基于 AdaBoost-Bayes 算法的中文文本分类系统[J]. 微电子学与计算机, 2016,33(6): 63-67.
	XU K , CHEN P H , LIU S Y . A chinese text classification system based on AdaBoost-Bayes algorithm[J]. Microelectronics ＆Computer, 2016,33(6): 63-67.
[5]	HUANG Z , . Vehicle pedestrian detection algorithm based on AdaBoost[C]// International Conference on Intelligent Transportation,December 19-20,2015,Halong Bay,Vietnam. Piscataway:IEEE Press, 2015.
[6]	WU S , NAGAHASHI H . Parameterized AdaBoost:introducing a parameter to speed up the training of real AdaBoost[J]. IEEE Signal Processing Letters, 2014,21(6): 687-691. doi: 10.1109/LSP.2014.2313570
[7]	ZHU J Q , CAI C H . Real-time face detection using gentle AdaBoost algorithm and nesting cascade structure[C]// International Symposium on Intelligent Signal Processing ＆ Communications Systems，Nov 4-7,2012,Taipei, China. Piscataway:IEEE Press, 2013.
[8]	KARLOS S , FAZAKIS N , KOTSIANTIS S ,et al. Self-Train LogitBoost for semi-supervised learning[C]// EANN,Sep 25-28,2015,Rhodes,Greece. Heidelberg:Springer, 2015.
[9]	GAO C , SANG N , TANG Q . On selection and combination of weak learners in AdaBoost[J]. Pattern Recognition Letters, 2010,31(9): 991-1001. doi: 10.1016/j.patrec.2009.12.019
[10]	YANG H , LIU S , LU R ,et al. Prediction of component content in rare earth extraction process based on ESNs-Adaboost[J]. IFAC-Papers on Line, 2018,51(24): 42-47.
[11]	CAO Y , MIAO Q G , LIU J C ,et al. Advance and prospects of AdaBoost algorithm[J]. Acta Automatica Sinica, 2013,39(6): 745-758. doi: 10.3724/SP.J.1004.2013.00745
[12]	CHENG W C , JHAN D M . A self-constructing cascade classifier with AdaBoost and SVM for pedestrian detection[J]. Engineering Applications of Artificial Intelligence, 2013,26(3): 1016-1028. doi: 10.1016/j.engappai.2012.08.013
[13]	田垅, 刘宗田 . 最小二乘法分段直线拟合[J]. 计算机科学, 2012,39(S1): 482-484.
	TIAN L , LIU Z T . Least-squares method piecewise linear fitting[J]. Computer Science, 2012,39(S1): 482-484.

数据集	实例数	特征维数
Sonar	208	60
CMC	962	9
Mushroom	5 644	22
Credit-g	1 000	20
Cylinder-bands	277	39

对比项	Discrete-AdaBoost	Real-AdaBoost	线性拟合	二阶拟合	三阶拟合	迭代次数
Sonar	0.116	0.113	0.121	0.105	0.079	30
	0.097	0.058	0.050	0.052	0.037	60
	0.042	0.026	0.024	0.016	0.005	100
CMC	0.413	0.391	0.386	0.381	0.364	30
	0.401	0.383	0.377	0.366	0.351	60
	0.390	0.371	0.368	0.357	0.326	100
Mushroom	0.328	0.331	0.320	0.321	0.315	30
	0.322	0.325	0.315	0.311	0.307	60
	0.315	0.313	0.309	0.302	0.298	100
Credit-g	0.211	0.205	0.186	0.188	0.186	30
	0.208	0.201	0.191	0.171	0.175	60
	0.199	0.193	0.176	0.161	0.160	100
Cylinder-bands	0.144	0.158	0.140	0.081	0.077	30
	0.112	0.117	0.094.	0.050	0.031	60
	0.103	0.108	0.090	0.036	0.027	100

对比项	Discrete-AdaBoost/s	Real-AdaBoost/s	线性拟合/s	二阶拟合/s	三阶拟合/s
Sonar	148	133	19	32	41
CMC	266	251	23	34	53
Mushroom	354	369	32	67	81
Credit-g	276	280	27	45	70
Cylinder-bands	136	127	12	21	43

对比项	Discrete-AdaBoost	Real-AdaBoost	本文算法
人脸数/张	159	159	159
检出人脸/张	133	137	139
漏检个数/张	26	25	20
误检个数/张	51	57	34
检测率	83.64%	86.16%	87.42%
误检率	24.29%	26.39%	17.62%

对比项	Discrete-AdaBoost	Real-AdaBoost	本文算法
人脸数/张	205	205	205
检出人脸/张	192	197	198
漏检个数/张	13	8	7
误检个数/张	18	13	9
检测率	93.66%	96.09%	96.59%
误检率	8.07%	5.96%	4.21%

基于拟合型弱分类器的AdaBoost算法

AdaBoost algorithm based on fitted weak classifier

在线阅读

PDF下载

可视化

摘要/Abstract

引用本文

使用本文

图/表 10

参考文献 13

相关文章 3

Metrics

推荐阅读 0

[1]	刘建,赵加奎,陈雨泽,方学民,刘玉玺,王树龙,陈雁,欧阳红. 考虑春节影响的售电量预测结果修正方法[J]. 电信科学, 2019, 35(3): 127-134.
[2]	陈雨泽,李磊,方学民,刘建,赵加奎,刘玉玺,左松林. 基于回归分析的迎峰度夏用电量分析[J]. 电信科学, 2019, 35(11): 125-131.
[3]	杨路,何萍,王珊. UFMC系统中的载波频率同步技术[J]. 电信科学, 2016, 32(11): 50-55.