Accurate felt-tip pen brands classification based on a convolutional neural network using data augmentation.

Area of Science:

• Forensic Science
• Analytical Chemistry
• Machine Learning

Background:

• Accurate ink analysis is crucial for document examination, but limited datasets hinder algorithmic classification of ink brands.
• Felt-tip pen ink identification presents challenges due to data scarcity, impacting the reliability of forensic analysis.

Purpose of the Study:

• To evaluate the effectiveness of Gaussian noise data augmentation (GNDA) and extended multiplicative signal augmentation (EMSA) for classifying felt-tip pen ink brands.
• To assess the performance of various classification models when applied to augmented spectral data.

Main Methods:

• FT-IR spectroscopy was used to analyze four brands of felt-tip pen inks.
• Two data augmentation techniques, GNDA and EMSA, were applied to the spectral data.
• Five classification models were employed: CNN, KNN, SVM, RF, and PLS-DA.

Main Results:

• Augmented datasets generated by GNDA and EMSA showed similarity to original data with added diversity.
• The EMSA method, when paired with a CNN, achieved superior classification performance.
• EMSA-CNN demonstrated high accuracy (99.86%), precision (99.87%), recall (99.86%), and F1 score (99.86%).
• In contrast, the GNDA-CNN method yielded lower results (ACC: 80.90%, PRE: 87.34%, REC: 81.62%, F1: 79.23%).

Conclusions:

• Data augmentation methods, particularly EMSA, are effective in enhancing the performance of machine learning models for ink brand identification, especially with limited spectral data.
• Combining EMSA with CNNs offers a robust solution for accurate felt-tip pen ink classification in forensic document analysis.