Comparative Analysis of Deep Learning Models: CNN, MobileNetV2, and ResNet50 for Offline Signature Verification

Abstract

Off line-signature verification is important when it comes to dealing with security and authentications in different fields such as banking and legal realms. The comparative analysis of three DLA that are CNN, MobileNetV2 and ResNet50 is performed in this study with regard to offline SV. In the past, there have been methods where engineers crafted features manually, thus they did not capture subtleties of signature variations. Thus, the deep learning approaches, especially the CNN has outperformed other approaches since it learns the features from the raw data. To compare these models,
this analysis looks at each person’s structure, effectiveness, and computational speed to inform the appropriateness of their application to functional issues. CNNs are good for feature extraction while MobileNetV2 provides a small model ideal for scenarios that have fewer resources, ResNet50 has a use of residual connections to solve vanishing gradient problem and performs well in detecting features that even a human might not notice on the images. Finally, this research aims at restoring the selection of proper models that fits the particular application definitely improving the off-line signature verification systems.