CAJ | 학술논문

连续AdaBoost算法要求对样本空间进行划分,传统的等距划分无法体现正负样本各自的分布规律.对基于连续AdaBoost算法的人脸检测方法进行了改进,结合离散AdaBoost中弱分类器的阈值选取策略,通过多重最优阈值选择方法实现了样本空间的合理划分.在MIT-CBCL数据库上的实验结果表明,改进后的方法比等距划分和连续AdaBoost算法检测率提高0.5%和2%,错误率降低0.15%和0.27%,算法收敛速度更快.
련속AdaBoost산법요구대양본공간진행화분,전통적등거화분무법체현정부양본각자적분포규률.대기우련속AdaBoost산법적인검검측방법진행료개진,결합리산AdaBoost중약분류기적역치선취책략,통과다중최우역치선택방법실현료양본공간적합리화분.재MIT-CBCL수거고상적실험결과표명,개진후적방법비등거화분화련속AdaBoost산법검측솔제고0.5%화2%,착오솔강저0.15%화0.27%,산법수렴속도경쾌.
Real AdaBoost algorithm demands division of the sample space. The traditional finite division can not reflect the distribution of positive and negative samples. In this paper, a new real AdaBoost algorithm based on multi-threshold method was developed. Through the selection method of multi-optimization threshold and combining the strategy of weak classifier threshold selection in discrete AdaBoost algorithm, the rational division of sample space was implemented. The experimental results on MIT-CBCL database prove the improved real AdaBoost algorithm increases the detection rate by 0.5% and 2% than the traditional finite division algorithm and real AdaBoost altorithm, and decrease the error rate by 0.15% and 0.27%, and its convergence is faster.