摘要:
针对目标检测中的非对称分类问题,在分析现有的由离散AdaBoost算法扩展得到的代价敏感(即非对称)学习算法的基础上,提出了以三个不同的非对称错误率上界为核心的推导非对称AdaBoost算法的统一框架. 在该框架下, 不仅现有离散型非对称AdaBoost算法之间的关系非常清晰, 而且其中不符合理论推导的部分可以很容易得到修正. 同时, 利用不同的优化方法, 最小化这三个不同上界, 推出了连续型AdaBoost算法的非对称扩展(用Asym-Real AdaBoost和Asym-Gentle AdaBoost 表示). 新的算法不仅在弱分类器组合系数的计算上比现有离散型算法更加方便, 而且实验证明, 在人脸检测和行人检测两方面都获得了比传统对称AdaBoost算法和离散型非对称AdaBoost算法更好的性能.
Abstract:
Asymmetry is inherent in tasks of object detection where rare positive targets need to be distinguished from enormous negative patterns. That is, to achieve a higher detection rate, the cost of missing a target should be higher than that of a false positive. Cost-sensitive learning is a suitable way for solving such problems. However, most cost-sensitive extensions of AdaBoost are realized by heuristically modifying the weights and confidence parameters of the discrete AdaBoost. It remains unclear whether there is a unified framework to interpret these methods as AdaBoost, clarify their relationships, and further derive the superior real-valued cost-sensitive boosting algorithms. In this paper, according to the three different upper bounds of the asymmetric training error, we not only give a detailed discussion about the various discrete asymmetric AdaBoost algorithms and their relationships, but also derive the real-valued asymmetric boosting algorithms in the form of additive logistic regression with analytical solutions, which are denoted by Asym-Real AdaBoost and Asym-Gentle AdaBoost. Experiments on both face detection and pedestrian detection demonstrate that the proposed approaches are efficient and achieve better performance than the previous AdaBoost methods and discrete asymmetric extensions.