摘要
该文提出了高光谱成像技术结合机器学习快速无损鉴别黑色签字笔墨水种类的新方法。采集36支不同品牌型号的黑色签字笔笔迹的高光谱图像,对每支签字笔笔迹的高光谱图像选取18个感兴趣区域,共提取648个平均光谱作为样本集。对450~950 nm的原始光谱进行Savitzky-Golay平滑、Z-Score标准化和两种组合方法光谱预处理,使用线性判别分析(LDA)和随机子空间-线性判别分析(RSM-LDA)分别构建黑色签字笔墨水种类鉴别模型。实验结果表明:不同预处理方法对RSM-LDA模型的鉴别准确率影响较小,而对于LDA模型,组合预处理具有更优的鉴别准确率;相比LDA模型,RSM-LDA模型分类效果更佳,训练集的平均分类准确率达100%,交叉验证平均分类准确率达99.09%,测试集的平均分类准确率达90.70%,每类样本的准确率、精准率、召回率均高于LDA模型分类结果,模型的接受者操作特征曲线下方面积(AUC值)达0.9983,模型性能良好。因此,采用高光谱成像技术结合RSM-LDA可实现不同品牌型号黑色签字笔墨水的快速无损鉴别。
Signing-pen inkblok is an important evidence in the forgery cases which comprise docu⁃ments,certificates,checks etc.By analyzing the ink types of suspicious handwriting,forensic sci⁃ence experts could recognize suspect's writing behaviors and infer whether the writing utensils are identical.In this paper,a new method for the rapid and nondestructive identification on types of black signing-pen ink by combining hyperspectral imaging(HSI)technique and machine learning was proposed.Primarily,thirty-six collected black signing-pens were numbered in turn through different brands and models,then each of them were used to write their own number three times repeatedly on the same specification of white A4 printing paper as the handwriting for the test.After that,the hyper⁃spectral imager was used to collect the hyperspectral images of the prepared handwritings,and black and white calibration was performed for all the images in order to reduce the effects of dark current of camera and changes of light intensity on the image signal.And then,ENVI was used to read the hy⁃perspectral image information,and eighteen representative region of interest(ROI)were selected manually on the hyperspectral image of each pen's handwriting.In term of the average spectra calcula⁃tion of the extracted region of interest,a total of 648 average spectra were finally obtained as the sam⁃ple set.For investigating the effect of different preprocessing methods,Savitzky-Golay smoothing,Z-Score standardization and their combined spectral pre-processing methods were respectively used to preprocess the handwriting original spectra data of 450-950 nm.Furthermore,linear discrimi⁃nant analysis(LDA)and random subspace method-linear discriminant analysis(RSM-LDA)were respectively adopted to establish two identification models of black signature-pen ink types,and com⁃paring their merits and drawbacks.The experimental results showed that:(1)The hyperspectral im⁃ages of black signing-pen ink were too consistent to identify,so it was necessary to process by ma⁃chine learning algorithm;(2)Different pre-processing methods had little influence on the identifica⁃tion accuracy of RSM-LDA model,while LDA model had better identification accuracy after it was combined with spectral pre-processing method;(3)The average classification accuracy rates of LDA model for training set,cross validation set and testing set were 99.54%,98.16%and 84.50%re⁃spectively;(4)Compared with LDA model,RSM-LDA model had better classification effect.The average classification accuracy rates for training set,cross validation set and testing set could reach 100%,99.09%and 90.70%,respectively.And the accuracy rates,precision rates and recall rates of each type of samples were all higher than those of LDA model.The AUC value of RSM-LDA model was 0.9983,which indicated that the RSM-LDA model's performance was remarkably good.To sum up,RSM-LDA model was performed better than LDA model in processing redundant data,possessing anti-interfered,solving over fitting problem,getting better classification accuracy etc.,which exhibited better classification effect and robustness.Therefore,hyperspectral imaging technique combined with RSM-LDA model could be used to achieve the rapid and nondestructive classification and discrimination of different brands and models of black signing-pen ink.
作者
王书越
杨玉柱
何伟文
李润康
WANG Shu-yue;YANG Yu-zhu;HE Wei-wen;LI Run-kang(School of Investigation,People's Public Security University of China,Beijing 100038,China)
出处
《分析测试学报》
CAS
CSCD
北大核心
2021年第10期1489-1496,共8页
Journal of Instrumental Analysis
基金
国家重点研发计划(2017YFC0822506-3,2019YFF0303405)。