基于GADF变换和多尺度CNN的哈密瓜表面农药残留可见-近红外光谱判别方法  

Vis-NIR Spectra Discriminant of Pesticide Residues on the Hami Melon Surface by GADF and Multi-Scale CNN

作　　者：喻国威 马本学[1,2] 陈金成 党富民[4,5] 李小占[1] 李聪[1] 王刚 YU Guo-wei;MA Ben-xue;CHEN Jin-cheng;DANG Fu-min;LI Xiao-zhan;LI Cong;WANG Gang(College of Mechanical and Electrical Engineering,Shihezi University,Shihezi 832003,China;Key Laboratory of Northwest Agricultural Equipment,Ministry of Agriculture and Rural Affairs,Shihezi 832003,China;Mechanical Equipment Research Institute,Xinjiang Academy of Agricultural and Reclamation Science,Shihezi832000,China;Food Quality Supervision and Testing Center(Shihezi),Ministry of Agriculture,Shihezi 832000,China;Analysis and Testing Center,Xinjiang Academy of Agricultural and Reclamation Science,Shihezi 832000,China)

机构地区：[1]石河子大学机械电气工程学院,新疆石河子832003 [2]农业农村部西北农业装备重点实验室,新疆石河子832003 [3]新疆农垦科学院机械装备研究所,新疆石河子832000 [4]农业部食品质量监督检验测试中心(石河子),新疆石河子832000 [5]新疆农垦科学院分析测试中心,新疆石河子832000

出　　处：《光谱学与光谱分析》2021年第12期3701-3707,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(32060411)资助。

摘　　要：针对哈密瓜表面农药残留化学检测方法成本高且具有破坏性等问题,探索了可见-近红外(Vis-NIR)光谱技术对农药残留定性判别的可行性。以哈密瓜为载体,百菌清和吡虫啉农药为研究对象,采集哈密瓜表面无残留、百菌清和吡虫啉残留的可见-近红外漫反射光谱,利用格拉姆角场(GAF)将一维光谱数据转换为二维彩色图像,构建GAF图像数据集。设计一种包含Inception结构的多尺度卷积神经网络模型用于哈密瓜表面农药残留种类判别,包括1层输入层、3层卷积层、1层融合层、1层平坦层、2层全连接层和1层输出层。模型测试混淆矩阵结果表明,格拉姆角差场(GADF)变换对哈密瓜表面农药残留的可见-近红外光谱表达能力较强。此外,构建AlexNet、VGG-16卷积神经网络(CNN)模型和支持向量机(SVM)、极限学习机(ELM)机器学习模型与提出的多尺度CNN模型进行性能对比。结果表明,3种CNN模型对哈密瓜表面有无农药残留的判别效果较好,综合判别准确率均高于SVM和ELM模型。对比3种CNN模型性能,多尺度CNN模型的性能最佳,训练耗时为14 s,综合判别准确率为98.33%。多尺度CNN模型结构利用多种小尺寸滤波器组合(1×1,3×3和5×5)和并行卷积模块,能够捕获不同层次和尺度的特征,通过级联融合模式进行深度特征融合,提高了模型的特征提取能力。与传统深度CNN模型相比,在保证计算复杂度不变的情况下,多尺度CNN模型的精度得到了有效提高。实验结果表明,GADF变换结合多尺度CNN模型可以有效进行光谱数据解析,利用可见-近红外光谱技术可以实现哈密瓜表面农药残留的定性判别。研究结果为大型瓜果表面农药残留的快速无损检测技术的研发提供了理论参考。Given the costly and destructive detection of pesticide residues on the Hami melon surface,the feasibility of visible/near-infrared(Vis/NIR)spectroscopy for the qualitative discriminant was assessed.In this study,Hami melon was taken as experimental samples.Two pesticides were taken as the research objects,including chlorothalonil and imidacloprid.Hami melon’s Vis/NIR spectra of Hami melon with no,chlorothalonil and imidacloprid residues were collected in the diffuse reflectance mode.Then the one-dimensional spectrum was transformed into a two-dimensional image by using gramian angular fields(GAF).The GAF image data set was constructed.A multi-scale convolutional neural network(CNN)architecture incorporatedan Inception module was developed,including aninput layer,three convolution layers,amerging layer,aflatten layer,two fully-connected layers,and an output layer.The confusion matrix result of the multi-scale CNN model suggested that the best method for expressing Vis/NIR spectral features was gramian angular difference fields(GADF)transformation.Moreover,two CNN models(AlexNet and VGG-16)and two machine learning models(support vector machine(SVM)and extreme learning machine(ELM))were established toverify the proposed model performance.With higher average accuracy than SVM and ELM models,the CNN models had a better effect to identifying whether there were pesticide residues on the Hami melon surface.Compared with AlexNet and VGG-16 models,the proposed multi-scale CNN model had the best performance with the shortest training time of 14 s and the highest test accuracy of 98.33%.The multi-scale CNN structure can capture different level and scale features by using combinations of various small-size filters(11,33 and 55)and stacking of parallel convolutions.The multi-scale deep feature fusion was carried out in the concatenation mode,which can improve the feature extraction ability of the CNN model.Compared with traditional CNN models with large depth,the model proposed in this study improved the discriminant accura

关 键 词：哈密瓜 格拉姆角场变换 可见-近红外光谱 多尺度卷积神经网络 农药残留判别 

分 类 号：O657.3[理学—分析化学]