HYPERSPECTRAL DOCUMENT IMAGE ANALYSIS FOR WRITER IDENTIFICATION AND INK MISMATCH DETECTION USING TRANSFORMERS AND ADVANCED AI TECHNIQUES

Abstract

This work introduces a Spectral-Spatial Transformer Network (SSTN) that is designed to optimize hyperspectral document image processing. The objective is to enhance the detection of ink mismatches and the identification of writers in forensic applications. Normal methods like Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs) often miss the complex relationship between spectral and spatial factors in hyperspectral data. The suggested transformer design gets around this problem by simulating both handwriting structures and spectral signatures at the same time using a multi-head focus method. Hyperspectral document samples were collected between 400 and 1000 nm using standardised image settings to make sure the reflectivity was correct. The SSTN did better than the CNN and SVM baselines, identifying writers 92% of the time and finding ink mismatches 95% of the time. Confusion matrices and ROC studies proved that the system was very good at telling the difference between things, and heatmaps and spectrum plots made it easy to see where the incorrect ink spots were. Up to the ideal configurations of eight layers, four attention heads, and one hundred bands, ablation tests showed that performance increases with transformer depth and spectral resolution. The results demonstrate how well the suggested model captures morphological and chemical information, allowing for very dependable and non-destructive document verification. This study establishes transformers as a strong basis for future hyperspectral forensic and archive processing systems, despite obstacles relating to light fluctuation, sensor noise, and limited dataset variety.