摘要
针对传统随机森林算法中,由于数据集内部差异程度随着节点分裂迅速降低,导致模型过早收敛的问题,提出允许上下特征不一致的改进随机森林算法.该算法根据聚类评估指标S_Dbw对每次二分k-means的结果进行评估,在节点分裂的过程中选择最合适的特征算子计算数据特征以维持节点数据分裂的可靠性和判别性,提高随机森林的判别能力.在此基础上将该算法的输出形式进行结构化,实现从二维向量调整为一个二维概率分布,构造成一个结构化随机森林.实验证明,针对背景中存在类肤色干扰和光照变化的手部检测,该算法可以有效地提高模型的判别能力.
With the aim to avoid the prematurity of the traditional random forest，while the inner difference of data set would decrease rapidly with partition，an improved random forest with variable feature extraction operators is proposed.The method would choose a theoretically best feat in the process of splitting of each node，according to the S_Dbw validity index，so as to maintain the reliability and discriminant of the splitting of nodes.Then the output form would be adjusted to construct a structured random forest，and it is applied in the work of hand detection with skin-like color and variable illumination.Experiments show that the proposed algorithm could improve the discriminative ability of the random forest.
出处
《小型微型计算机系统》
CSCD
北大核心
2018年第2期393-395,共3页
Journal of Chinese Computer Systems
基金
国家自然科学基金面上项目(61672461)资助
图像与视频的不变性局部结构特征描述及应用研究(61672463)资助
关键词
S_Dbw
特征算子不一致
改进随机森林
结构化
S_Dbw variable feature extraction operators an improved random forest structured random forest
