Deep Learning Based Stress Prediction From Offline Signatures

Abstract

Soft-Biometric measurements are now increasingly adopted as a robust means of determining individual’s non-unique characteristics with the emerging models that are widely used in the deep learning domain. This approach is clearly valuable in a variety of scenarios, specially those relating to forensics. In this study, we specifically focus on stress emotion, and propose automatic stress prediction technique from offline signature biometrics using well-known deep learning architectures such as AlexNet, ResNet and DenseNet. Due to the limited number of research that study emotion prediction from offline handwritten signatures with deep learning methods, best to our knowledge this is the first experimental study that presents empirical achievable prediction accuracy around 77%.