模糊协方差学习矢量量化的茶叶品种分类研究

Classification of Tea Varieties Using Fuzzy Covariance Learning Vector Quantization

茶叶是全球最受欢迎饮品之一,且具有丰富的营养价值,但目前市面上的茶叶鱼龙混杂,难以辨别。因此,快速准确的分类方法对茶叶进行鉴别具有重要的研究意义。由于大多数化合物基频吸收带均出现在波长为2500~25000 nm的中红外区域,茶叶的中红外光谱中含有大量关于茶叶品种的特征鉴别信息,利用这一显著特点可以对其进行分类。提出模糊协方差学习矢量量化(FCLVQ),该算法在GK(Gustafson-Kessel)聚类的基础上,引入学习向量量化(LVQ)中学习速率的概念,用以控制模糊类中心的更新速率。FCLVQ结合中红外光谱,通过不断迭代计算样本模糊隶属度值和模糊聚类中心,实现对茶叶的快速精准分类。选取市场上的峨眉山茶叶、优质竹叶青茶叶、劣质竹叶青茶叶作为实验对象。将实验对象分为3组(每个品种各1组),每组32个,共计96个样本。利用FTIR-7600型傅里叶红外光谱分析仪分别采集每组样本的中红外光谱数据,每组样本采集三次,取其平均值作为样本的红外光谱数据。首先,由于原始光谱含有噪声数据,故使用多元散射校正(MSC)作降噪预处理;其次,由于光谱数据维数高达1868维,采用主成分分析(PCA)将光谱数据降至14维,其14个主成分的累计贡献率为99.74%;然后将降维后的光谱数据使用线性判别分析(LDA)进一步降至2维,同时提取数据中的鉴别信息;最后运行模糊C均值聚类算法(FCM),将其运算得到的聚类中心作为FCLVQ的初始聚类中心参与迭代,设置模糊隶属度的权重指数m=2,最终分类准确率高达95.25%。将FCM算法、GK算法、模糊Kohonen聚类网络(FKCN)算法与FCLVQ算法的运行结果进行对比,FCM,GK和FKCN的分类准确率分别为90.91%,92.41%和90.91%。结果表明,与其他三个算法相比较,FCLVQ在m=2,主成分个数为14时有着更好的分类效果,可以用来实现对茶叶品种的准确分类。

Tea,with much nutrition,is one of the most popular drinks in the world.Good and bad tea are mixed at the market,so it is difficult to make a classification among them.Therefore,using a fast and accurate method to identify tea varieties is meaningful.Most chemical compound’s fundamental frequency absorption bands are within the wavelength range of 2500~25000 nm(Mid-infrared region).Large amounts of feature discriminant information in the mid-infrared spectra of tea can be applied to classify tea varieties.This paper proposed a fuzzy covariance the learning vector quantization(FCLVQ)based on the Gustafson-Kessel(GK)clustering.It introduces learning rate of learning vector quantization(LVQ)to control the update rate of cluster centers.Combined with mid-infrared spectroscopy,FCLVQ realizes fast and accurate identification of tea varieties by iteratively calculating the fuzzy membership values and fuzzy clustering centers of samples.Three different kinds of tea(i.e.Emeishan tea,high-quality bamboo-leaf-green tea and low-quality bamboo-leaf-green tea)were selected as 96 samples in total at the market.Each variety corresponds to one group,which consists of 32 samples.The Fourier mid-infrared spectra were collected using an FTIR-7600 spectrometer,and the average spectral data were computed as the final experimental spectra.Firstly,the original spectral data contained noise data,so they were pretreated with multiplicative scattering correction(MSC)to reduce noise.Secondly,principal component analysis(PCA)was employed to reduce the dimensionality of data from 1868 to 14,and the cumulative contribution of the 14 principal components was 99.74%.Thirdly,the dimensionality of the processed data was reduced to 2,and the discriminant information was extracted by linear discriminant analysis(LDA).Finally,fuzzy C-means clustering(FCM)was run to get initial cluster centers for FCLVQ.The experimental results showed that when the weight index m=2,the accuracy rate of FCLVQ was 95.25%.On the condition of m=2,for the same spectra,the classification accuracy rates of FCM,GK and fuzzy Kohonen clustering(FKCN)were 90.91%,92.41%and 90.91%respectively.The experimental results showed that compared with the other three algorithms,FCLVQ had a better classification accuracy when m=2 and the number of principal components were 14.Thus,it can be used to classify different tea varieties.