Learning Features for Offline Handwritten Signature Verification using Deep Convolutional Neural Networks (1705.05787v1)

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.

• The paper presents a novel writer-independent CNN framework that learns discriminative features for offline handwritten signature verification.
• It introduces a two-phase methodology with a feature learning stage on a development set followed by writer-dependent classification.
• Experiments show significant error rate reductions, notably lowering the GPDS-160 Equal Error Rate to 1.72%.

An Overview of "Learning Features for Offline Handwritten Signature Verification using Deep Convolutional Neural Networks"

The paper "Learning Features for Offline Handwritten Signature Verification using Deep Convolutional Neural Networks" by Hafemann, Sabourin, and Oliveira, explores the development of a novel framework to enhance the accuracy of offline handwritten signature verification systems using deep learning methodologies. The authors present an advanced approach to learn features directly from signature images using Convolutional Neural Networks (CNNs), addressing key challenges in distinguishing genuine signatures from skilled forgeries.

Technical Contribution and Methodology

Traditionally, offline signature verification systems face significant hurdles due to the lack of dynamic information inherent in static (offline) signatures, leading to higher error rates in distinguishing genuine signatures from skilled forgeries. To combat this, the authors propose a Writer-Independent framework for feature learning, leveraging a CNN model to explore and capture signature features that transcend individual writer characteristics.

The methodology is divided into two main phases:

1. Feature Learning Phase: A CNN is trained using a development dataset to learn a robust feature representation for signatures in a Writer-Independent format. These features aim to encapsulate the inherent properties of genuine signatures that distinguish them from forgeries.
2. Classification Phase: Writer-Dependent classifiers are subsequently trained using the learned features on an exploitation set. This two-tiered approach allows the system to adapt and verify new users without requiring forgeries for every registered user.

Key to this paper is the incorporation of skilled forgery information from a subset of users during the training of the CNN. This is a significant deviation from common practice, aiming to enable the CNN to learn visual cues indicative of forgeries across different users. The authors explore different formulations, including a multi-task learning strategy that optimizes the CNN both for distinguishing different users and detecting forgeries.

Results and Evaluation

The proposed system was tested across multiple datasets, including GPDS, MCYT, CEDAR, and Brazilian PUC-PR, demonstrating significant improvements over existing state-of-the-art methods. Notably, on the GPDS-160 dataset, the method achieved an Equal Error Rate (EER) of 1.72%, markedly improving from the previous best of 6.97%. Moreover, the paper illustrates that these learned features generalize well to other datasets, emphasizing the system's adaptability to different signature styles and conditions.

The experiments indicate a substantial reduction in error rates even when only a minimal number of genuine samples per user are available, showcasing the efficacy of the feature learning methodology. The application of CNNs in this domain highlights the potential of deep learning models to redefine and advance traditional biometric security systems, shifting from handcrafted feature extraction towards data-driven solutions.

Implications and Future Directions

The findings suggest promising potential for practical applications where security and accuracy in signature verification are critical. The interdisciplinary nature of the approach, blending computer vision and biometric verification, points towards future explorations that could incorporate multi-modal biometric systems, combining visual and dynamic information.

Future research directions may entail refining the feature learning process to further accommodate variations across different signature collection protocols or cultures, thus ensuring even broader applicability. Moreover, advancements in user-specific thresholding strategies could further enhance the real-world applicability of signature verification solutions.

This paper significantly contributes to the ongoing development of biometric verification systems by employing deep learning models, raising discussions about the intersection between technology and security in biometric applications.