Machine Learning-based Signature Verification using OCR and Line Sweep Technique

Abstract

Handwritten signatures remain a widely used biometric for identity verification in banking, legal and governmental domains. Manual verification is slow and error-prone; automated verification is necessary to reduce fraud and speed processing. This paper presents a hybrid machine-learning based offline signature verification system that integrates Optical Character Recognition (OCR) for signature localization, Connected Components analysis and a Line Sweep geometric feature extraction technique, combined with supervised classifiers (Support Vector Machine and Convolutional Neural Network). The system takes scanned cheque or document images, localizes signature regions using OCR, extracts stroke and geometric descriptors via connected-components and line-sweep operations, composes feature vectors, and classifies signatures as genuine or forged. Experimental evaluation on our dataset shows effective detection performance (≈91% accuracy) and robust separation between forged and genuine samples. The proposed pipeline is scalable and suitable for deployment in semi-automated banking workflows