DEEP LEARNING FOR OFFLINE SIGNATURE VERIFICATION: A NOVEL MULTI-CHANNEL FEATURE FUSION NETWORK

Abstract

This paper addresses the critical challenge of offline signature verification, a task crucial for authenticating documents and identities. Existing deep learning approaches, primarily deep metric learning with Siamese networks and two-channel discriminative methods, face limitations. While Siamese networks excel at feature extraction, their reliance on Euclidean distance can overlook subtle directional and scaling information, hindering the capture of intricate feature relationships. Conversely, two-channel discriminative methods, though effective in initial dissimilarity assessment, often suffer from significant feature loss due to early image fusion. To overcome these challenges, we propose the Multi-channel Feature Fusion Network , a novel writer-independent model for handwritten signature verification. The proposed framework leverages a quadruple Siamese network and a dual inverse discriminative attention mechanism for robust feature extraction and enhancement from both original and inverse grayscale images. These rich, multi-dimensional features are then integrated through an innovative channel fusion process. Finally, an ACMix-based discriminative module is employed to determine image similarity with high precision. Comprehensive experiments on four diverse language signature demonstrate the superior efficacy and promising potential of the framework, affirming its advantages over current methodologies.