Handwritten signature verification is a widely used biometric for person identity authentication in document forensics. Despite the tremendous efforts in past research, offline signature verification still remains a challenge, particularly in discriminating between genuine signatures and skilled forgeries, because the difference of appearance between genuine and skilled forgery may be smaller than that between genuine ones. This challenge is even more critical in writer-independent scenario, where each writer has very few samples for training. This paper proposes a region based Deep Convolutional Siamese Network using metric learning method, which is applicable to both writer-dependent (WD) and writer-independent (WI) scenario. For representing minute but discriminative details, a Mutual Signature DenseNet (MSDN) is designed to extract features and learn the similarity measure from local regions instead of whole signature images. Based on local regions comparison, the similarity scores of multiple regions are fused for final decision of verification. In experiments on public datasets CEDAR and GPDS, the proposed method achieved state-of-the-art performance of 6.74% EER and 8.24% EER in WI scenario, respectively, and 1.67% EER and 1.65% EER in WD scenario, respectively.

手写签名验证是文档取证中广泛用于个人身份认证的生物特征。尽管在过去的研究中付出了巨大的努力，但离线签名验证仍然是一个挑战，特别是在区分真签名和技术赝品方面，因为真品和技术赝品之间的外观差异可能比真品之间的差异要小。在与作者无关的情况下，每个作者只有很少的样本进行培训，这一挑战就显得尤为关键。本文提出了一种使用度量学习方法的基于区域的深度卷积连体网络，它适用于作家依赖（WD）和作家独立（WI）场景。代表细微但有区别的细节，Mutual Signature DenseNet (MSDN) 旨在从局部区域而不是整个签名图像中提取特征并学习相似性度量。在局部区域比较的基础上，融合多个区域的相似度分数，用于最终的验证决策。在公共数据集 CEDAR 和 GPDS 的实验中，所提出的方法在 WI 场景中分别实现了 6.74% EER 和 8.24% EER，在 WD 场景中分别实现了 1.67% EER 和 1.65% EER。