Offline Signature Verification USING Siamese Neural Network

Abstract

Verifying offline signatures is challenging in identity verification procedures for legal documents and financial transactions. The subtle but significant distinctions between authentic and fraudulent signatures must be carefully considered because some forgeries may merely alter a portion of the signature. This task is more difficult when the identity of the writer is unknown, which frequently occurs in real-world scenarios. In this paper, we propose a system that verifies offline, writer-independent signatures automatically and accurately using a Siamese neural network. This network consists of two or more identical sub-networks that share parameters and weights. The network calculates Euclidean distances between images by processing both similar and distinct images. When two signatures are the same, the distance is smaller; when they are different, it is larger. One-shot learning, which allows the network to learn from fewer image pairs, is its primary advantage. Two BHsig260 datasets including both authentic and fake signatures in Bengali and Hindi were used to test our technique. The model's proficiency was demonstrated by the Siamese network's strong accuracy of 82% and 84% on the Bengali and Hindi signature datasets, respectively. These findings show how well the model can differentiate between authentic and fake signatures, indicating its potential for real-world use across a range of industries.