Non-destructive discrimination of black gel pen inks via physics-guided band-gate attention network and hyperspectral...

The non-destructive identification of black gel pen inks and writing-to-alteration intervals is important in forensic ink analysis, particularly for legal and financial documents. This study proposes an intelligent framework integrating hyperspectral imaging (HSI) with a novel deep learning model, PG-BCANet, for objective ink characterization. Hyperspectral data were collected from black gel pen samples on standard A4 paper, representing different pen models and specified writing-to-alteration intervals. After noise suppression and baseline drift correction using Savitzky-Golay (SG) smoothing and Standard Normal Variate (SNV) normalization, three feature selection algorithms, including Principal Component Analysis (PCA), Interval Random Forest (iRF), and Successive Projections Algorithm (SPA), were comparatively evaluated. For classification, the proposed PG-BCANet, constructed on a 1-D Convolutional Neural Network (1DCNN), incorporates a Band-Gate mechanism, Efficient Channel Attention (ECA), and residual connections to optimize feature extraction from high-dimensional spectral data. PG-BCANet achieved accuracies of 96.6% for model identification and up to 89.2% for discrimination among the investigated temporal intervals. These results exceeded the corresponding KNN and SVM baselines under the adopted experimental protocol. This research provides a non-destructive approach for forensic ink analysis and supports further investigation of deep learning-enhanced HSI in forensic science.