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

In the practice of forensic science，it is often necessary to accurately determine the identity of the test material and the sample document by analyzing the composition of the ink in the document.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 400-1 000 nm，and spectral data of 56 samples were extracted.Use the uniform manifold approximation and projection（UMAP） and T-distributed stochastic neighbor embedding（t-SNE） two algorithms for hyperspectral 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 reduction 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 identification method for inkjet printing documents.