In the area of biometrics and document forensics, handwritten signatures are one of the most commonly accepted symbols. Thus, financial and commercial institutions usually use them to verify the identity of an individual. However, offline signature verification is still a challenging task due to the difficulties in discriminating the minute but significant details between genuine and skilled forged signatures. To tackle this issue, we propose a novel writer-independent offline signature verification approach using attention-based multiple siamese networks with primary representation guiding. The proposed multiple siamese networks regard the reference signature images, query signature images, and their corresponding inverse images as inputs. These images are fed to four weight-shared parallel branches, respectively. We present an efficient and reliable mutual attention module to discover prominent stroke information from both original and inverse branches. In each branch, feature maps of the first convolution are utilized to guide the combination with deeper features, named as primary representation guiding, which guides the model into concerning the shallow stroke information. The four branches are concatenated in an ordered way and are put into four contrastive pairs, which is helpful to obtain useful representations by comparing reference and query samples. Four contrastive pairs generate four preliminary decisions independently. Then, the eventual verification result is created based on the four preliminary decisions using a voting mechanism. In order to assess the performance of the proposed method, extensive experiments on four widely used public datasets are conducted. The experimental results demonstrate that the proposed method outperforms existing approaches in most cases and can be applied to various language scenarios.

在生物识别和文件取证领域，手写签名是最普遍接受的符号之一。因此，金融和商业机构通常使用它们来验证个人的身份。然而，离线签名验证仍然是一项具有挑战性的任务，因为很难区分真实签名和熟练伪造签名之间的微小但重要的细节。为了解决这个问题，我们提出了一种新颖的独立于作者的离线签名验证方法，使用基于注意力的多个暹罗网络和主要表示指导。所提出的多个孪生网络将参考签名图像、查询签名图像及其相应的逆图像作为输入。这些图像分别馈送到四个权重共享的并行分支。我们提出了一种高效可靠的相互关注模块，可以从原始分支和逆分支中发现突出的笔划信息。在每个分支中，利用第一个卷积的特征图来引导与更深层次特征的组合，称为初级表示引导，引导模型关注浅层笔划信息。这四个分支以有序的方式连接起来，并放入四个对比对中，这有助于通过比较参考样本和查询样本来获得有用的表示。四个对比对独立地产生四个初步决策。然后，使用投票机制根据四个初步决定创建最终的验证结果。为了评估所提出方法的性能，对四个广泛使用的公共数据集进行了广泛的实验。实验结果表明，所提出的方法在大多数情况下优于现有方法，并且可以应用于各种语言场景。