A Study for Automating Signature Authentication Systems

Signature verification is a critical task in biometric authentication, offering widespread applications in banking, legal, and security domains. This research presents a comprehensive comparison between Convolutional Neural Networks (CNN) and Siamese Neural Networks (SNN) in the context of signature verification. We design and implement a CNN architecture employing RELU and Sigmoid activation functions across different models, incorporating layers such as Conv2D, MaxPooling2D, Flatten, Dense, and Dropout to process and classify signatures as genuine or forged. In parallel, we explore a Siamese Neural Network architecture, leveraging the unique property of learning from pairs of inputs to discern subtle differences in signatures. This network utilises Euclidean distance as a metric to compare feature vectors derived from signature images, encapsulating the essence of signature verification as a problem of similarity measurement. Our experimental setup involves a detailed training regime, where both networks are subjected to a dataset comprising genuine and forged signatures, aiming to learn distinguishing features. The performance of each model is evaluated based on accuracy, loss, and validation metrics across epochs, providing insights into their respective capabilities and limitations. The results indicate nuanced differences in how CNN and SNN approach the signature verification challenge, offering valuable perspectives for future research and application in automated signature authentication systems. Through this comparative study, we aim to contribute to the evolving landscape of biometric verification, providing a robust analysis that aids in selecting appropriate neural network architectures for signature verification tasks.