基于改进Hodrick-Prescott分解模型的近红外自适应降噪方法  

An Improved Hodrick-Prescott Decomposition Based Near-Infrared Adaptive Denoising Method

作　　者：谢德红 李俊锋 刘菂 万晓霞[4] 叶艺 XIE De-hong;LI Jun-feng;LIU Di;WAN Xiao-xia;YE Yi(School of Light Industry and Food,Nanjing Forestry University,Nanjing 210037,China;School of Packaging and Printing Engineering,Henan University of Animal Husbandry and Economy,Zhengzhou 450046,China;Beijing Key Laboratory of Big Data Technology for Food Safety,Beijing Technology and Business University,Beijing 100048,China;Hubei Province Engineering Technical Center for Digitization and Virtual Reproduction of Color Information of Culture Relics,Wuhan University,Wuhan 430079,China)

机构地区：[1]南京林业大学轻工与食品学院,江苏南京210037 [2]河南牧业经济学院包装与印刷工程学院,河南郑州450046 [3]北京工商大学食品安全大数据技术北京市重点实验室,北京100048 [4]武汉大学湖北省文物颜色信息数字化与虚拟再现工程研究中心,湖北武汉430079

出　　处：《光谱学与光谱分析》2020年第5期1650-1655,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(61275172,61575147);江苏省高校优势学科建设工程项目(164030934);江苏省制浆造纸科学与技术重点实验室开放基金项目(201526);南京林业大学青年科技创新基金项目(CX2018024);南京林业大学大学生创新创业训练计划项(2018NFUSPITP680);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)资助。

摘　　要：在检测果蔬农药残留的近红外光谱采集过程中,往往会受噪声干扰获得低信噪比近红外光谱,且近红外光谱中表征农药和果蔬化学组分的谱峰微弱且重叠度高,因而此近红外光谱降噪普遍存在易平滑微弱的农药组分谱峰、或增加非测量物化学组分谱峰的危险,导致在后续以仅挖掘红外光谱谱峰特征为前提的分类和化学组分分析中,恶化近红外光谱的分类精度、影响农药残留成分的正确分析。针对抑制近红外光谱噪声与保持近红外光谱谱峰的矛盾,提出一种改进Hodrick-Prescott分解模型的自适应降噪方法。在该方法的Hodrick-Prescott分解模型中,以染噪光谱与复原光谱之间残差的L2范数为残差项,描述高斯噪声结构,以复原光谱信号二阶差分的L2范数为正则化项,惩罚复原光谱、迫使从染噪光谱中复原的光谱倾向于梯度减少的方向,以平滑噪声、保持原始谱峰信息。该方法同时结合L-曲线方法,自适应地获取染噪光谱在Hodrick-Prescott优化方程中的正则化参数,并通过求解该曲线最大曲率点对应的参数获得最优正则化参数,确保能平衡Hodrick-Prescott分解模型中正则化项和残差项,以得到较为理想的光谱复原结果。实验以农药残留和未残留的上海青近红外光谱为基本数据、通过降噪前后信噪比、以及支持向量机分类模型的识别率,对比分析bior6.8小波分解方法、sym8小波分解方法、互补集合模态分解方法的降噪效果。实验结果显示,该方法在处理18.79 dB信噪比染噪近红外光谱时获得了33.35 dB信噪比;在实施上海青农药残留检测中,处理训练集与测试集近红外光谱数据后,训练所得支持向量机分类模型的训练集识别率达93.58%、测试集识别率达71.18%,此识别率明显高于上述三种方法降噪后的结果,接近于原始未染噪声光谱数据。该方法在近红外光谱降噪方面具有明显的优势,能应用于�During the rapid detection of pesticide residue in fruits and vegetables by near-infrared(NIR)spectroscopy,NIR spectroscopy is often contaminated by noises.Meanwhile,the peaks in NIR spectroscopy of chemical components of pesticides and fruits and vegetables are weak and highly overlapped,so denoising the NIR spectroscopy has risks of smoothing weak peaks of pesticide components or generating peaks of non-chemical components.In the subsequent classification or chemical composition analysis,the above problems deteriorate the accuracy of classification of the NIR spectroscopy and influence analysis of chemical components of pesticide residue.In order to solve the conflict between noise suppression and peak maintenance of the NIR spectroscopy,an adaptive denoising method is proposed based on an improved Hodrick-Prescott decomposition model.In the model,L2 norm of the residual between the noisy near-infrared spectroscopy and its restored spectroscopy is used as the residual term to describe the Gaussian noise structure,and L2 norm of the second-order difference of the restored spectroscopy is used as the regularization term to penalize the restored spectroscopy.The penalty can force the restored spectroscopy to reduce its gradient,resulting in smoothing noises and keeping the original peaks.In order to acquire the regularization parameter in the optimization equation of the Hodrick-Prescott decomposition model adaptively,an L-curve method is combined into the method.So in the method,the optimal regularization parameters are obtained by solving the parameters corresponding to the maximum curvature point of the L-curve,which can balance the regularization term and the residual term in the Hodrick-Prescott decomposition model and finally obtain ideal restored spectroscopy.In order to compare wavelet decomposition methods with bior6.8 basis and sym8 basis and complete ensemble empirical mode decomposition method,signal-to-noise ratio(SNR)is computed,and a support vector machine(SVM)classification model is established by

关 键 词：Hodrick-Prescott分解 L曲线 自适应 降噪 近红外光谱 

分 类 号：O657.3[理学—分析化学]