提出利用聚类方法对光谱数据进行特征筛选。通过分析类内参数γ1和类间参数γ2对筛选结果的影响，选择较好的γ1和γ2进行有用光谱特征筛选。利用烤烟叶的近红外反射光谱（1500～2400 nm间隔2 nm），选用SVM方法进行部位和颜色分组识别，训练样本的识别率为100%，测试样本的识别率分别是96.22%和92.79%。然后利用聚类方法对初始光谱进行特征筛选，选用相同的 SVM 方法及相同的学习样本和测试样本进行部位和颜色分组识别。在删减部分不相干光谱后，识别率分别提高到97.23%和95.52%；继续删除相关度不高的光谱，在识别率略有下降时，光谱特征数可减少到200个以下。结果表明：利用聚类方法进行特征筛选，不仅可提高识别率，且可大大减少光谱数据，因而极大地减少了数据采集时间，简化了分组模型，提高了系统的实时和快速处理能力。%The clustering method is applied to select the effective features from the original spectra. The effective features used for betterγ1 andγ2 is chose by analyzing the influence of intra-class parametersγ1 and inter-class parametersγ2. Part and color of tobacco leaves are classified by SVM method based on the near infrared reflecting spectra(1500 nm-2400 nm interval of 2 nm)of flue-cured tobacco leaves. The recognition rates of part and color are 100% for train sample, 96.22% and 92.79% respectively for test sample. After some irrelevant spectra are removed by clustering algorithm, the recognition rate can be improved to 97.23%and 95.52%respectively. Continue cutting spectra having low correlation with classification. The recognition rate will declined significantly when too many spectra are removed. The number of spectra can be reduced to about 200 from 451 with slightly low recognition rate. The experiment results show that the clustering method can not only improve the recognition rate but also greatly reduce the number of spectral data. This greatly lessens the time of collecting data and significantly improves the real-time and fast processing ability of the system.
展开▼
