基于特征分层选择和融合度相结合的近红外光谱多类识别度量算法研究

发布时间：2018-04-28 06:39

本文选题：近红外光谱 + 特征分层选择　；参考：《光谱学与光谱分析》2017年04期

【摘要】：针对高维空间下获取最优特征子集异常复杂和模型识别准确率较低的问题,提出了基于特征分层选择和融合度相结合的近红外光谱多类识别度量算法。首先引入跳跃度,构造了一种特征分层方法,将所有特征依据对样本的重要性程度划分不同的特征子集,从而避免了从原始特征数据逐个剔除无关特征构建特征子集的繁琐过程;同时又改进了样本的融合度,将其代替K最近邻分类器(KNN)中依据概率进行类别判断的方式,提高了分类器的识别精度,较好地解决了多类识别准确率较低的问题。为验证该算法的有效性,选取五类具有代表性382个烟叶样品为实验对象,构建了烟叶产地识别度量模型,并选取64个样本进行了模型测试,以预测均方根误差(RMSEP)、交互验证均方差(RMSECV)和相关系数(r)作为模型稳健性的评价指标,以产地识别准确率作为算法优劣评价标准。仿真实验结果表明,利用该算法构建的模型具有较低的RMSEP(0.117),RMSECV(0.106)和较高的r(0.973),平均识别准确率达到98.44%,性能明显优于其他算法,该算法对于高维光谱数据具有良好的识别性能。
[Abstract]:Aiming at the problem that obtaining the optimal feature subset in high dimensional space is very complicated and the accuracy of model recognition is low, a multi-class recognition metric algorithm based on the combination of feature stratification selection and fusion degree is proposed. Firstly, by introducing jump degree, we construct a feature stratification method, which divides all features into different feature subsets according to their importance to the sample. It avoids the tedious process of constructing the feature subset from the original feature data one by one, and at the same time improves the fusion degree of the sample and replaces the method of classification judgment based on probability in the K-nearest neighbor classifier (KNNN). The recognition accuracy of the classifier is improved and the problem of low accuracy of multi-class recognition is well solved. In order to verify the validity of the algorithm, a measurement model of tobacco origin recognition was constructed by selecting 382 representative tobacco samples from five categories as experimental objects, and 64 samples were selected to test the model. The root-mean-square error (RMSEPV), cross-validation mean square error (RMSECV) and correlation coefficient (r) were used as the evaluation index of model robustness, and the accuracy of origin recognition was taken as the evaluation criterion of the algorithm. The simulation results show that the model constructed by this algorithm has lower RMS EPU 0.117 and higher RMSECV 0.106), and the average recognition accuracy is 98.443.The performance of the algorithm is obviously better than that of other algorithms. The algorithm has good recognition performance for high-dimensional spectral data.
【作者单位】：中国海洋大学信息科学与工程学院;
【基金】：国家科技支撑计划课题(2015BAF12B01)资助
【分类号】：O657.33

【相似文献】