A Stroke-Based RNN for Writer-Independent Online Signature Verification

In the field of online handwritten signature verification, it is challenging to verify handwritten signature in a writer-independent scenario. In recent years, many researchers have been applying deep neural network methods to the signature verification task. However, these methods have not outperformed traditional methods, especially when the training samples are limited. In this paper, we propose a novel stroke-based bidirectional RNN architecture. The main idea is to split the signature into multiple patches using strokes. Concatenation of query and reference signature pairs are used as input. The proposed method uses two LSTM RNN networks to extract different features. The first one extracts the features of the strokes and the latter extracts the global features of the whole signatures. The results on the BiosecureID dataset demonstrate that our proposed method can reduce the EER by 33.05%, from 5.6% to 3.75% with fewer features and less training samples. Besides, we find that the proposed stroke based RNN network is 5x faster in training and testing time than Non stroke-based RNN network.