基于1DCNN和PLSDA酸枣仁真伪高光谱图像鉴别中的关键特征分析  

Key Feature Analysis in Identification and Authenticity of Ziziphi Spinosae Semen by Using Hyperspectral Images Based on 1DCNN and PLSDA

作　　者：赵昕 石玉娜 刘怡彤 姜洪喆 褚璇 赵志磊[1,4] 王宝军 陈晗[1] ZHAO Xin;SHI Yu-na;LIU Yi-tong;JIANG Hong-zhe;CHU Xuan;ZHAO Zhi-lei;WANG Bao-jun;CHEN Han(School of Quality and Technical Supervision,Hebei University,Baoding 071002,China;College of Mechanical and Electronic Engineering,Nanjing Forestry University,Nanjing 210037,China;College of Mechanical and Electrical Engineering,Zhongkai University of Agriculture and Engineering,Guangzhou 510225,China;National&Local Joint Engineering Research Center of Metrology Instrument and System,Hebei University,Baoding 071002,China)

机构地区：[1]河北大学质量技术监督学院,河北保定071002 [2]南京林业大学机械电子工程学院,江苏南京210037 [3]仲恺农业工程学院机电工程学院,广东广州510225 [4]计量仪器与系统国家地方联合工程研究中心,河北保定071002

出　　处：《光谱学与光谱分析》2025年第3期869-877,共9页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(32102087);河北省省级科技计划项目(21344801D);河北省专业学位研究生教学案例建设项目(KGJSZ2022005);2024年省级大学生创新训练计划项目(S202410075089)资助。

摘　　要：酸枣仁因其养心益肝的功效,是安神助眠类保健品和中药制剂的重要原料。目前市售酸枣仁掺假现象严重,极大损害了消费者利益,扰乱了市场秩序。传统人工检测或基于实验室的高效液相色谱方法存在效率低,推广难的问题。本研究基于卷积神经网络和偏最小二乘判别提出了一种高光谱成像酸枣仁真伪鉴别方法,并对两类模型中的关键光谱特征进行了讨论研究,为后续多光谱系统和便携式仪器开发提供借鉴。提取酸枣仁及其常见伪品(理枣仁、兵豆和枳椇子)高光谱图像(400~1000 nm)中所有单籽粒的平均光谱。基于平均光谱分别建立偏最小二乘判别分析(PLSDA)模型和一维卷积神经网络(1DCNN)模型。PLSDA建模前采用竞争性自适应重加权算法(CARS)挑选特征波长。在1DCNN模型中添加了自定义波长选择层,并对卷积层和全连接层输出结果应用t分布随机邻域嵌入(t-SNE)进行可视化分析。为了与CARS-PLSDA模型进行有效对比,构建了基于五个波长的5W-1DCNN模型。结果表明CARS-PLSDA和1DCNN模型都能获得理想的预测效果,校正集和预测集分类正确率均在99%以上。对比CARS与自定义层挑选的特征波长,670、721和850 nm附近的波长在两种模型中均具有重要作用。研究结果为酸枣仁真伪快速鉴别的多光谱和便携式检测设备提供参考。Ziziphi spinosae semen is an important raw material of health care products and traditional Chinese medicine preparations because it nourishes the heart and the liver,making it ideal for calming the nerves and helping sleep.At present,the adulteration of ziziphi spinosadsemen in the market is serious,which greatly damages the interests of consumers and disrupts the market order.Traditional manual detection or laboratory-based high-performance liquid chromatography methods have problems of low efficiency and difficult promotion.In this study,a hyperspectral imaging method for ziziphi spinosadsemen authenticity identification was proposed based on convolutional neural network and partial least squares discrimination,and the key spectral features in the two types of models were discussed and studied.The study will reference the subsequent development of multispectral systems and portable instruments.The average spectra of all single kernels in the hyperspectral images(400~1000 nm)of ziziphi spinosae semen and its common counterfeits(Ziziphus mauritiana lam,Hovenia dulcis Thunb.and Lens culinaris)were extracted.The partialleast squares discriminant analysis(PLSDA)model and the one-dimensional convolutional neural network(1DCNN)model were respectively established based on the average spectra.The competitive adaptive reweighting algorithm(CARS)selects characteristic wavelengths before PLSDA modeling.A custom wavelength selection layer was added to the 1DCNN model.T-distributed stochastic neighborhood embedding(t-SNE)was applied to the outputs of convolutional and fully connected layers for visual analysis.To effectively compare with the CARS-PLSDA model,a 5W-1DCNN model based on five wavelengths was constructed.The results showed that both the CARS-PLSDA and1DCNN models could achieve precision prediction results,and the classification accuracies of both the calibration set and the prediction set are above 0.99.Comparing the feature wavelengths selected by CARS and custom layers,wavelengths near 670,721,and 850 nm play

关 键 词：高光谱成像 一维卷积神经网络 t分布随机邻域嵌入 偏最小二乘法判别分析 

分 类 号：TS255.7[轻工技术与工程—农产品加工及贮藏工程]