计算机工程 ›› 2007, Vol. 33 ›› Issue (21): 172-174.doi: 10.3969/j.issn.1000-3428.2007.21.061

• 人工智能及识别技术 • 上一篇    下一篇

基于双阈值的增强型AdaBoost快速算法

严云洋1,2,郭志波1,杨静宇1   

  1. (1. 南京理工大学计算机科学与技术学院,南京 210094;2. 淮阴工学院计算机工程系,淮安 223001)
  • 收稿日期:1900-01-01 修回日期:1900-01-01 出版日期:2007-11-05 发布日期:2007-11-05

Fast Enhanced AdaBoost Algorithm Based on Dual-threshold

YAN Yun-yang1,2, GUO Zhi-bo1, YANG Jing-yu1   

  1. (1. School of Computer Science and Technology, Nanjing University of Science and Technology, Nanjing 210094; 2. Department of Computer Engineering, Huaiyin Institute of Technology, Huai’an 223001)
  • Received:1900-01-01 Revised:1900-01-01 Online:2007-11-05 Published:2007-11-05

摘要: 在应用AdaBoost算法的人脸检测中,针对训练时间太长及权重调整过适应等问题,提出一种基于特征值等分和双阈值的增强型AdaBoost快速训练算法,给出了双阈值的快速搜索方法。在MIT-CBCL人脸和非人脸训练库上对算法进行了实现。实验结果显示,改进后的双阈值增强型AdaBoost算法简化了训练过程,训练速度提高50倍,收敛速度也更快。使用训练得到的检测器对MIT+CMU人脸测试库进行了测试,结果表明,该方法在检测精度和速度等方面都优于单阈值方法。

关键词: dual-AdaBoost, 双阈值, 人脸检测

Abstract: Aiming at some problems of too much training time and weight adjustment over fitting in face detection by using AdaBoost. This paper proposes enhanced dual-AdaBoost algorithm based on feature-value-division and dual-threshold, which makes training faster and better. It gives a method to get dual-threshold, and presents the improved mode of weight adjustment. Experimental results on MIT-CBCL training data set illustrate that the dual-AdaBoost makes training process converge quickly, and the training time is as 1/50 as before. Experimental results on MIT+CMU with the detectors show that the detection speed and precision under the dual-threshold are better than single-threshold method.

Key words: dual-AdaBoost, dual-threshold, face detection

中图分类号: