Offline Signature Verification in the Banking Domain

In this paper, we address a specific industrial application of offline signature verification, in a context arising in the banking domain. In this context, we need to verify signatures placed in different types of real-world documents, e.g. bank cheques. Due to the huge amount of clients the problem is framed into the writer-independent configuration. Peculiar to this industrial application is that the signatures are associated with additional information, i.e. the personal data associated to the bank account. We propose a deep learning architecture to solve the signature verification of our industrial application, in which we take advantage of the additional textual information with a pretraining of an image embedding component driven by a text-recognition task. Moreover, our model employs an attention mechanism over the image embeddings to learn more powerful features. We extensively compare our model with a state-of-the-art siamese approach, with a text recognition model adopted to specifically solve this task and using a general purpose image encoder pretrained on the ImageNet dataset instead of text recognition embeddings. Our model proves to be more effective to solve the offline signature verification in two real-world large scale datasets. Ablation studies also confirm the importance of our ideas: the pretraining of the image encoders with a text recognition task and the attention mechanism.