基于多任务学习的注意力机制双向GRU用于在线手写签名认证

Bidirectional GRU with Attention Mechanism Based on Multi-Task Learning for Online Handwritten Signature Verification

摘要: 深度学习的进步极大地提高了在线签名认证(online signature verification, OSV)系统的性能，但是如何从有限的签名样本中学习具有判别性的签名特征仍然是该领域所面临的挑战。为了缓解这个问题，本研究提出了一种基于多任务学习的注意力机制双向门循环单元(MT-A-BiGRU)模型来实现在线签名的认证。首先，通过基于注意力机制的双向门循环单元(A-BiGRU)模型来实现序列特征的有监督表征学习，并引入深度度量学习任务来充分挖掘序列特征的潜在表示。在此基础上，将稀疏自动编码器压缩后的全局特征与A-BiGRU模型所提取的特征进行融合，以实现特征信息的互补。最后，本文提出一种基于多任务学习的训练方法，进一步提高了OSV系统的准确率。所提出的方法在SVC-2004-task2数据集中达到了2.16%的等错误率和96.88%的准确率，实验结果表明所提出的方法够有效地提高OSV系统的认证精度。

Abstract: Advances in deep learning have greatly improved the performance of online signature verification (OSV) systems, but how to learn discriminative signature features from limited signature samples remains a challenge in this field. To alleviate this problem, this study proposes a multitask learning-attention mechanism-bidirectional gate recurrent unit (MT-A-BiGRU) model to achieve online signature verification. Firstly, the supervised representation learning of sequence features is realized through an attention mechanism-bidirectional gate recurrent unit (A-BiGRU) model, and a deep metric learning task is introduced to fully mine the latent representations of sequence features. On this basis, the global features compressed by sparse auto encoder are fused with the features extracted by the A-BiGRU model to achieve the complementarity of feature information. Finally, this study proposes a training method based on multi-task learning, which further improves the accuracy of the OSV system. The proposed method achieves an equal error rate of 2.16% and an accuracy of 96.88% in the SVC-2004-task2 dataset. The experimental results show that the proposed method can effectively improve the verification accuracy of the OSV system.