Writer Identification by Processing Hyperspectral Images Using Singular Value Decomposition and Linear Discriminant Analysis

Writer identification is a critical task in forensic document analysis, requiring precise methods to differentiate individual handwriting samples. This study presents a framework leveraging hyperspectral imaging (HSI) for writer identification, employing Singular Value Decomposition (SVD) to reconstruct denoised hyperspectral cubes, which are then classified using Linear Discriminant Analysis (LDA) and Convolutional Neural Networks (CNN). Optimal SVD components are selected via analysis of variance (ANOVA), and pen-based blind testing is conducted under multiple training–testing ink mixtures. The framework is evaluated on the UWA Writing Ink Hyperspectral Image (WIHSI) dataset (33 bands, 400–720 nm, seven writers, multiple blue and black inks) and the iVision Handwritten Hyperspectral Image Dataset (HHID) (149 bands, 512 × 650 pixels, 54 writers, with five selected blue pens and one sample per pen) across four ink mixing configurations: 4:1, 3:2, 2:3, and 1:4. On the iVision HHID dataset, SVD-LDA achieves near-perfect accuracy of 99.99% for the 4:1, 3:2, and 2:3 configurations, and 97%for the 1:4 configuration, outperforming SVD-CNN. On the UWA WIHSI dataset, with limited training and black ink excluded in the 1:4 configuration, SVD-LDA attains 97% accuracy, while SVD-CNN drops to 86.12% with higher training data, SVD-CNN slightly outperforms SVD-LDA. Overall, SVD-LDA demonstrates stability under limited training, SVD-CNN excels with abundant data, and SVD reconstruction enhances spectral discriminability, especially for noisy datasets like iVision HHID. These results confirm the robustness, effectiveness, and practical applicability of the proposed framework for forensic writer identification.