Raw ripeness discrimination is a pivotal technology of machine vision system of cotton picking robot. The dimensionality curse exists in spatial and frequency feature sets describing morphology structure and boundary contour of raw cotton, and their feature selection problem is an NP hard problem. A solution algorithm is proposed using floating search by filter and stopping search by wrapper based on cross validation. The optimal l feature subset (l= 1, 2, 3, ??? ) is selected on training set using filter with an assessing function of class-separability measured value and heuristic criterion including optimal scalar feature combination and floating search. A model is established on training set using the optimal l feature subset by wrapper with an assessing function of the error rate of Bayes-classifier. The optimal feature subset is with a capacity of 6 at the minimum value of average error rate of the model on validation set, and the average classification rate of which on prediction sets is 87. 61%. The algorithm has been validated on 40 databases involved in correlative research work, and experimental results show that the algorithm has good classification performance and higher execution efficiency on 29 datasets.%描述田间籽棉成熟度的形态结构和边界轮廓特征集存在维数灾难,其特征选择问题属于NP难题.基于交叉验证,提出了一种过滤器下浮动搜索并基于封装器停止搜索的求解算法.在训练集上以最大类可分性测量值为过滤器的评估函数启发式搜索最优l维特征子集(l=1,2,3,…),启发式规则包括最优特征组合和浮动搜索;在训练集上以Bayes分类器的误分率为封装器的评估函数对最优l维特征子集建模,模型在验证集上的平均误分率极小处产生的最优特征子集的容量为6,它们在预测集上的平均识别率为87.61％.在相关研究工作所涉及的40个数据集上验证算法的有效性,结果表明,在29个数据集上算法的分类性能好,执行效率高.
展开▼
