激光打印墨粉红外光谱的小样本种类判别  

Few-Shot Classification of Laser-Printing Toner Using Infrared Spectroscopy

作　　者：申思[1] 刘猛[1] Shen Si;Liu Meng(Department of Forensic Science and Technology,Zhejiang Police College,Hangzhou 310053,Zhejiang,China)

机构地区：[1]浙江警察学院刑事科学技术系,浙江杭州310053

出　　处：《激光与光电子学进展》2022年第19期373-380,共8页Laser & Optoelectronics Progress

基　　金：浙江省高等教育“十三五”第一批教学改革研究项目(jg20180381)。

摘　　要：激光打印墨粉的品牌型号判别是法庭科学专家认定涉案打印机具和鉴别添加打印文件的关键步骤,已有方法通常需要大量训练样本才能获得可靠的分类模型,但实际案例中往往无法获取足够的样本。针对这一问题提出了基于红外光谱和化学计量法的小样本墨粉种类判别方法。实验采集了8种不同品牌型号的激光打印墨粉红外光谱数据,根据光谱信号的特点以及遍历法选出效果最佳的光谱数据预处理方法,并对处理后的数据建立偏最小二乘判别分析(PLS-DA)模型,以随机森林(RF)和支持向量机(SVM)作为对比算法。实验结果表明:二阶导数与Savitzky-Golay平滑组合使用可以获得实验数据的最佳预处理效果;PLS-DA模型的准确率始终高于RF和SVM,训练集样本数大于90时PLS-DA模型准确率为100%,训练集样本数为60时该模型准确率为95%。提出的基于红外光谱和PLS-DA的激光打印墨粉种类判别模型准确率高、可解释性强且所需训练样本数较少,在物证检验领域具备较强的实用性。Classifying laserprinter toner is an essential step for identifying printers and dictating forged documents.However,the existing methods require numerous training samples,which is unrealistic in document examination cases.A fewshot classification method based on infrared spectroscopy and chemometrics is proposed.The infrared spectrum of eight types of toner was collected,and optimal spectroscopic data preprocessing methods were selected according to the data’s characteristics and traverse comparison experiment.Using the processed data,a partial least squarediscriminant analysis(PLSDA)model was established.Random forest(RF)and support vector machine(SVM)were used as the comparison methods.Experimental results show that the second derivative and SavitzkyGolay smoothing are the best preprocessing methods for the collected spectrum.In all conditions,PLSDA outperform RF and SVM.When the number of the training set is larger than 90,the accuracy of the PLSDA model is 100%,and when the number of the training set is 60,it reduces to 95%.The discriminant model of laserprinter toner based on infrared spectroscopy and PLSDA exhibits high accuracy and strong interpretability,requires less training samples,and can be applied in forensic science.

关 键 词：光谱学 小样本分类 红外光谱 化学计量法 墨粉分类 

分 类 号：O433.4[机械工程—光学工程]