Ink classification in historical documents using hyperspectral imaging and machine learning methods.

Area of Science:

• Analytical Chemistry
• Imaging Science
• Computational Science

Background:

• Ink identification is challenging due to degradation and spectral overlap.
• Hyperspectral imaging offers a non-invasive approach to spectral analysis.
• Machine learning can classify complex spectral data.

Purpose of the Study:

• To classify metallo-gallate, carbon-containing, and non-carbon-containing inks using hyperspectral imaging.
• To evaluate traditional and deep learning models for ink classification accuracy.
• To assess the feasibility of non-invasive ink analysis for historical documents.

Main Methods:

• Utilized hyperspectral imaging across multiple systems.
• Applied six supervised machine learning models: SVM, KNN, LDA, RF, PLSDA, and a DL model.
• Integrated data fusion, sample extraction, ground truth creation, and post-processing.

Main Results:

• All models achieved >90% micro-averaged accuracy on mock-up samples.
• The deep learning model yielded the highest F1-score (98%).
• Traditional models performed better on historical case study documents.

Conclusions:

• Hyperspectral imaging combined with machine learning is effective for non-invasive ink identification.
• The approach is robust even with degraded materials and spectral overlap.
• This technology supports the conservation and analysis of historical manuscripts.