拉曼光谱结合深度学习对司法物证中印文色料的分类鉴别  

Raman spectroscopy combined with deep learning for classification and identification of Chinese pigments in judicial evidence

作　　者：谢尚志 陈维娜[1] 李开开[1] 张涛[1] XIE Shangzhi;CHEN Weina;LI Kaikai;ZHANG Tao(Investigation College,People’s Public Security University of China,Beijing 102623,China)

机构地区：[1]中国人民公安大学侦查学院,北京102623

出　　处：《应用化工》2025年第2期521-526,共6页Applied Chemical Industry

基　　金：国家自然科学基金(22203104)。

摘　　要：在法庭科学中,印文色料的检验是鉴别印迹真伪的重要依据。印文色料主要分为普通印油、光敏印油及印泥三类,适用于不同的印章盖印印文。为实现这三类印文色料的无损高精度快速鉴别,实验收集了市面上不同品牌、型号的40种印文色料样本,采集600条拉曼光谱数据并扩增至6000条;实验结合全波段(All Band)、主成分分析(PCA)、连续投影(SPA)、随机蛙跳变量选择(RF)四种特征提取方法和Lenet-5、VGG16、Resnet50三种深度学习算法,建立了12种分类预测模型并比较和讨论了其识别效果。结果表明:12种分类模型的测试准确率均达到80%以上,其中PCA+VGG16模型性能最佳,泛化性最好,模型测试准确率达到了96%,各评价指标为全部模型中最优。在模拟仿真实验中,该模型对7个未知样本的类型预测全部正确,预测准确率为100%,且每个样本的置信度高达99.9%以上。这为印章印文类物证的快速无损检验提供了一个解决思路,为拉曼光谱技术和深度学习算法在法庭科学领域的应用提供一定参考。In forensic science,the inspection of printing materials is an important basis for distinguishing the authenticity of imprints.Printing materials are mainly divided into three types:ordinary ink,photosensitive ink,and ink paste,which are suitable for different types of seals and seals.To achieve non-destructive,high-precision,and rapid identification of these three types of printing materials,40 samples of different brands and models of printing materials on the market were collected in the experiment,and 600 Raman spectroscopy data were collected and amplified to 6000;the experiment combines four feature extraction methods,namely All Band,Principal Component Analysis(PCA),Continuous Projection(SPA),Random Frog Jump Variable Selection(RF),and three deep learning algorithms,Lenet-5,VGG16,and Resnet50,to establish 12 classification prediction models and compare and discuss their recognition effects.The results showed that the testing accuracy of the 12 classification models reached over 80%,among which the PCA+VGG16 model had the best performance and generalization.The model testing accuracy reached 96%,and all evaluation indicators were the best among all models.In the simulation experiment,the model correctly predicted the types of 7 unknown samples,with a prediction accuracy of 100%and a confidence level of over 99.9%for each sample.This provides a solution for the rapidnon-destructive testing of physical evidence in seal printing,and provides a certain reference for the application of Raman spectroscopy technology and deep learning algorithms in the field of forensic science.

关 键 词：印文色料 拉曼光谱 深度学习 法庭科学 特征提取 

分 类 号：TQ628.5[化学工程—精细化工] O657.37[理学—分析化学]