稀疏集SVN惩罚校正方法及其种质评价应用研究

Research on a penalty regularization method for SVN based on sparse training set and its application to seed quality evaluating

针对支持向量学习网络(sVN)学习稀疏样本数据集时,稀疏目标和非稀疏目标的分类器错误率严重失衡而实用性大大降低的问题,在拉格朗日乘数渐近分析基础上,引入惩罚校正因子、逆向训练样本和错误训练率等概念,提出了惩罚校正支持向量网络学习算法和校正方法,并将该方法应用于以CT图像特征数据集为基础的小麦籽种品质定级。等值分析说明该学习算法能有效地等级化籽种特征数据,准确率达95%;和其他同源方法的对比试验显示:针对稀疏样本集,该算法在获得可观综合预测准确性的同时,能显著改善稀疏样本集各目标分类器的预测错误率的极性分布,并展现良好的学习性能。

Aiming at the problem when an unbalanced, sparse training set is trained by using support-vector networks （SVN） ,the output classifiers are badly imbalanced in their re^s-prediction rate, so they are badly unavailable, the novel concepts of penalty regularization coefficient, through-bound-SV, adverse training sample, mis-training rate and so on were introduced into this study based on the analysis of Lagrange multiplier, and the Regularized Penalty SVN learning algorithm was proposed;a method for regularizing penalty coefficient was designed. The algorithm was ap- plied to grading wheat seeds quality based on a feature data set of CT image of the seeds. The contour analysis dem- onstrates it can effectively grade image features data with an accuracy rate of 95 %. The results of the comparison ex- periment against some prior and congeneric algorithms suggests where a sparse training set is concerned. This meth- od can reform the polar distribution of the mis-prediction rate of concerned classifiers, achieve an amusing accuracy of prediction, and present a very good global performance of machine learning.