Learning features for offline handwritten signature verification using deep convolutional neural networks

Abstract

Verifying the identity of a person using handwritten signatures is challenging in the presence of skilled forgeries, where a forger has access to a person's signature and deliberately attempt to imitate it. In offline (static) signature verification, the dynamic information of the signature writing process is lost, and it is difficult to design good feature extractors that can distinguish genuine signatures and skilled forgeries. This reflects in a relatively poor performance, with verification errors around 7% in the best systems in the literature. To address both the difficulty of obtaining good features, as well as improve system performance, we propose learning the representations from signature images, in a Writer-Independent format, using Convolutional Neural Networks.In particular, we propose a novel formulation of the problem that includes knowledge of skilled forgeries from a subset of users in the feature learning process, that aims to capture visual cues that distinguish genuine signatures and forgeries regardless of the user. Extensive experiments were conducted on four datasets: GPDS, MCYT, CEDAR and Brazilian PUC-PR datasets. On GPDS-160, we obtained a large improvement in state-of-the-art performance, achieving 1.72% Equal Error Rate, compared to 6.97% in the literature. We also verified that the features generalize beyond the GPDS dataset, surpassing the state-of-the-art performance in the other datasets, without requiring the representation to be fine-tuned to each particular dataset.

A white-box analysis on the writer-independent dichotomy transformation applied to offline handwritten signature verification

“…The SigNet, proposed by Hafemann et al (2017a), uses Deep Convolutional Neural Networks (DCNN) for learning the signature representations in a writerindependent way and, nowadays, represents a state of the art approach in this research area. This approach tries to build a new representation space in which different writers are clustered in separate regions, based on the most representative properties of the handwritten signatures.…”

“…In their study, Hafemann et al (2017a) analysed the local structure of the learned feature space, by using the t-SNE algorithm in a subset containing 50 writers from the development set of the GPDS-300 dataset (referred to as the validation set for verification V v ). Figure 4 represents this analysis (Hafemann et al, 2017a).…”

“…In its turn, in the transfer learning scenario, the whole set of writers are used to obtain the development sets, but we keep the number of genuine signatures and random forgeries as in (Souza et al, 2019b). (37 or 38) · 9 · 5 samples (37 or 38) · 10 · 9/2 samples Considering that each dataset has a different number of writers and signature per writers and to be able to compare the results with the state-of-the-art the testing set is acquired using a methodology similar to that described in (Hafemann et al, 2017a). For the MCYT and the CEDAR datasets all the writers were used and differently from (Hafemann et al, 2017a), random forgeries were included.…”

“…(37 or 38) · 9 · 5 samples (37 or 38) · 10 · 9/2 samples Considering that each dataset has a different number of writers and signature per writers and to be able to compare the results with the state-of-the-art the testing set is acquired using a methodology similar to that described in (Hafemann et al, 2017a). For the MCYT and the CEDAR datasets all the writers were used and differently from (Hafemann et al, 2017a), random forgeries were included. Table 6 summarizes the used Exploitation set ε for each dataset.…”

“…• Skilled forgeries: the forger has information about both the name and the genuine signature pattern of the original writer, resulting in forgeries that are more similar to genuine signatures. (Hafemann et al, 2017a). Each column shows two genuine signatures from the same writer and a skilled forgery, from the GPDS dataset.…”