高光谱喷墨打印墨水数据的非线性降维及分类建模方法研究  

Research on Nonlinear Dimensionality Reduction and Classification Modeling Methods of Hyperspectral Inkjet Printing Ink Data

作　　者：李硕 崔岚[1] 付沛 LI Shuo;CUI Lan;FU Pei(College of Criminal Science and Technology,Criminal Investigation Police University of China,Shenyang 110035,China)

机构地区：[1]中国刑事警察学院刑事科学技术学院,辽宁沈阳110035

出　　处：《分析测试学报》2024年第4期523-531,共9页Journal of Instrumental Analysis

基　　金：公安部科技强警基础工作计划(2022JC03);“十三五”国家重点研发计划项目资助(2016YFC0800705);中国刑事警察学院研究生创新能力提升项目(2023YCZD07)。

摘　　要：在法庭科学实践中,往往需要通过对文件中字迹墨水的成分分析来精确判定检材和样本文件的同一性。该文利用高光谱成像技术结合机器学习对喷墨打印墨水的种类进行区分,分别采集14套不同品牌、型号的4色(黑、青、品红和黄色)喷墨打印墨水打印的文件在400~1000 nm范围的高光谱图像,共提取56种样品墨迹的光谱数据。使用均匀流形逼近与投影技术(UMAP)和T分布随机近邻嵌入技术(t-SNE)两种算法对高光谱喷墨打印墨水数据进行降维处理,然后建立极致梯度提升(XGBoost)、轻量级梯度提升机器学习(LightGBM)和支持向量机(SVM)3种分类模型,以1∶4的比例确定测试集和训练集,分别对原始数据和降维后的数据进行分类。实验结果显示,UMAP降维算法结合SVM模型对喷墨打印墨水分类的效果最优,黑色墨水样品的分类精度为90%左右,其余颜色墨水样品的分类精度均为100%。该研究为喷墨打印文件的检验鉴定提供了一种新的、无损、准确的鉴别方法。In the practice of forensic science,it is often necessary to accurately determine the identi⁃ty of the test material and the sample document by analyzing the composition of the ink in the docu⁃ment.Hyperspectral imaging technology combined with machine learning was used to distinguish the types of inkjet printing inks.Hyperspectral images of documents printed with 4 colors(black,blue,magenta and yellow)of 14 sets of different brands and models were collected in the range of 400-1000 nm,and spectral data of 56 samples were extracted.Use the uniform manifold approximation and pro⁃jection(UMAP)and T-distributed stochastic neighbor embedding(t-SNE)two algorithms for hyper⁃spectral data dimension reduction processing inkjet printing ink,and then establish extreme gradient boosting(XGBoost),light gradient boosting machine(LightGBM)and support vector machine(SVM),determine the test set and training set in the ratio of 1∶4,and classify the original data and the data after dimensionality reduction respectively.The experimental results show that UMAP dimension re⁃duction algorithm combined with SVM model has the best effect on the classification of inkjet printing inks.The classification accuracy of black ink samples is about 90%,and the classification accuracy of other color ink samples is 100%.This study provides a new,non-destructive and accurate identifi⁃cation method for inkjet printing documents.

关 键 词：高光谱成像技术 喷墨打印墨水 降维算法 分类模型 

分 类 号：O657.3[理学—分析化学] TB9[理学—化学]