A New Computer Vision Prototype for Document Cross Strokes Analysis in a Portable and Non-Destructive Way

Documentoscopy is a specialized field in criminalistics that verifies the authenticity of suspect documents during fraud investigations, providing crucial forensic evidence. Traditional documentoscopy relies on costly scanning electron microscopy (SEM), subject to bias, with an accuracy of about 63% in determining stroke sequences. Recent advancements led to cost-effective alternatives. Using Computer Vision (CV) and Convo-lutional Neural Networks (CNN), a prototype was developed to identify document fraud by analyzing stroke overlap. This electronic eye captures document images inexpensively, automatically, and non-destructively, followed by CNN-based analysis to classify stroke overlap due to different pens and inks. The prototype used with 2052 images of intersecting lines from various pens and inks, achieved over 94% accuracy for training data and over 65% for testing data, surpassing traditional documen-toscopy accuracy. The prototype is simple, affordable, user-friendly without specialized technicians, portable for fieldwork, and non-destructive, preserving samples for forensic counterproof in future investigations.