A Hybrid System for Offline Handwritten Signature Identification and Verification Using OC-SDA Classifier and Dissimilarity Features

Handwritten signature recognition is typically addressed through separate systems for identification and verification. This paper proposes a novel hybrid system that combines both tasks using a One-Class Symbolic Data Analysis (OC-SDA) classifier. The system leverages dissimilarities extracted from Convolutional Neural Networks (CNNs) to build symbolic representation models for each writer. A new fuzzy similarity measure based on a tuned Gaussian weighted membership function is introduced. The system is evaluated on four benchmark datasets (GPDS, CEDAR, MCYT, and PUC-PR Brazilian), achieving a Hybrid Identification Rate (HIR) of 96.72% on the GPDS dataset with only five reference signatures.