BharatSign AI: Leakage-Controlled Multi-Script Offline Cheque Signature Forgery Classification for Indian Banking

BharatSign AI is the first writer-level verified, script-isolated multi-engine framework for offline cheque signature forgery classification in Indian multi-script banking. Three independent engines — English (CEDAR, SUNY Buffalo), Hindi, and Bengali (BHSig260, ISI Kolkata) — are built using Scale Invariant Feature Transform keypoint extraction, Bag of Visual Words encoding, twelve geometric contour features, and Linear Support Vector Classification, each trained and evaluated exclusively on its own script dataset under a strict 70/15/15 writer-level split with no data leakage. Verified results: Bengali engine achieves 81.25% accuracy and 16.67% Equal Error Rate — the best reported on BHSig260 Bengali with classical handcrafted features; English achieves 77.55% accuracy and 22.69% EER; Hindi achieves 69.79% accuracy and 29.43% EER. Two original contributions are introduced — the Hesitation Index, a novel pixel-density spike feature capturing the neurological difference between genuine muscle-memory signing and deliberate forgery, and a filename-based label-safety mechanism preventing silent prediction inversion in production biometric classifiers. Total end-to-end inference is 32.4 milliseconds on standard CPU hardware with no GPU dependency. The system directly addresses active Reserve Bank of India mandates on automated fraud detection and delivers the first known deployable multi-script signature verification solution for Indian banking.