基于谱变换和高阶稀疏Hodrick-Prescott分解的茶叶品种鲁棒判别方法  

Robust discrimination method for tea varieties based on spectral transformation and high-order Sparsity-aided Hodrick-Prescott decomposition

作　　者：赵秀芝[1,2] 宁井铭 谢德红[4] ZHAO Xiu-zhi;NING Jing-ming;XIE De-hong(College of Artificial Intelligence,Zhejiang Industry&Trade Vocational College,Wenzhou 325002,China;School of Computer and Artificial Intelligence,Wenzhou University,Wenzhou 325002,China;State Key Laboratory of Tea Plant Biology and Utilization,Anhui Agricultural University,Hefei 230036,China;College of Information Science and Technology,Nanjing Forestry University,Nanjing 210037,China)

机构地区：[1]浙江工贸职业技术学院人工智能学院,浙江温州325002 [2]温州大学计算机与人工智能学院,浙江温州325002 [3]安徽农业大学茶树生物学与资源利用国家重点实验室,安徽合肥230036 [4]南京林业大学信息科学技术学院,南京210037

出　　处：《吉林大学学报（工学版）》2023年第5期1465-1473,共9页Journal of Jilin University:Engineering and Technology Edition

基　　金：茶树生物学与资源利用国家重点实验室开发基金项目(SKLTOF090113);食品安全大数据技术北京市重点实验室开放基金项目(BTBD-2019KF02)。

摘　　要：为了提高可见-近红外光谱定性分析的精度,需对光谱进行降噪预处理。针对降噪易产生额外小谱峰、恶化定性分析准确度的问题,提出一种基于谱变换和高阶稀疏Hodrick-Prescott分解的降噪方法。在该方法的优化方程中,假设可见-近红外光谱由低通的基本波形光谱、带通的特征波形光谱及噪声组成,以含噪光谱与基本波形光谱、带通的特征波形光谱之间残差L2范数为残差项,保证估计值逼近真实值;依据特征波形光谱的稀疏性,以其二阶差分的L1范数为正则化项,约束估计特征波形光谱,从而分解出茶叶中重要的特征吸收峰。该方法同时利用滤波器的谱变换技术获得低通和带通零相位滤波器矩阵,协助分解基本波形光谱和特征波形光谱,并利用L-曲线方法获取优化方程中的最佳正则化参数。本实验以6种茶叶的可见-近红外光谱为基础实验数据。在实验中,以信噪比、均方根差和茶叶品种定性分析分类模型的准确性为衡量指标,与小波分解法、改进的Hodrick-Prescott法和Savitzky-Golay法进行了比较。实验结果显示:对含高斯噪声合成光谱数据和含高斯-脉冲混合噪声合成光谱数据,该方法信噪比最高;对于合成和真实两个数据集,分类模型准确率均高于上述3种方法预处理后的结果,且远高于含噪数据下的分类结果。因此,该方法在可见-近红外光谱降噪方面具有优势,能应用于基于可见-近红外光谱的茶叶品种定性检测的预处理。In order to improve the accuracy of qualitative analysis of visible near infrared spectroscopy,noise reduction pretreatment is needed.Aiming at the problem that it is easy to produce additional small spectral peaks which deteriorate the accuracy of qualitative analysis during noise reduction,a noise reduction method based on spectral transformation and high-order sparse Hodrick-Prescott decomposition is proposed.In the optimization model of this method,it is assumed that the visible near infrared spectroscopy is composed of low-pass basic waveform spectrum,band-pass characteristic waveform spectrum and noise.In this method,the L2 norm of the residual between the noisy spectroscopy,the basic waveform spectroscopy and the band-pass characteristic waveform spectroscopy is taken as the residual term to ensure that the estimated value is close to the real value.According to the sparsity of the characteristic waveform spectrum,taking the L1 norm of the second-order difference of it as the regularization term,it is constrained to estimate that the characteristic waveform spectrum has a certain sparsity to characterize absorption peaks of the important chemical components in tea.In this method,the low-pass and band-pass zero phase filter matrices are obtained by using the spectral transformation of the filter to help decompose the basic waveform spectroscopy and the characteristic waveform spectroscopy,and the optimal regularization parameters in the optimization equation are obtained by using the L-curve method.The experiment takes the visible near infrared spectroscopy of six kinds of tea as the basic experimental data.In the experiment,our method compares with the wavelet decomposition method,improved Hodrick-Prescott method and savitzky-Golay method by the signal-to-noise ratio,root mean square difference,and the accuracy of the classification model of qualitative analysis of tea varieties as the measures.The experimental results show that our method has the highest signal-to-noise ratio in synthetic spectra containi

关 键 词：可见-近红外光谱 茶叶 品种 降噪 稀疏 

分 类 号：O657.3[理学—分析化学]