Massive Offline Signature Forgery Detection with Extreme Learning Machines

In this work, we present the results of different machine learning models for detecting offline signature forgeries trained from the features obtained from a massive dataset of ten thousand users. Features for training are obtained from the last layer of two different convolutional neural networks: Inception21k and Signet. Optimisation of the number of neurons and activation functions of Extreme Learning Machine (ELM) models are obtained using Equal Error Rate (EER) as a metric. Our results align with the recent results of other machine learning models. Furthermore, we found that a general-purpose network: Inception21k performs better for the writer-independent models created.