Deep Learning applied to Offline Signature Verification

Dr. Ángel Sánchez Calle, (Universidad Rey Juan Carlos)
Resumen: Handwritten signatures are a non-intrusive widely-accepted biometric modality which has been commonly used to verify the authorship of documents. Signature verification consists in determining the similarity degree between a test signature and a model signature to decide whether or not this test signature is authentic or a forgery. One of the main difficulties in both tasks is the intrapersonal and interpersonal variabilities when signing, as the signatures from same writer presents variations due to the available space for signing, the type of pen used, the condition of the signer, and other aspects. Off-line signatures are images obtained after scanning a signed document and there is not any dynamic informationfrom the act of signing (e.g. pressure on the paper, ordering of produced strokes and so on). The off-line verification of handwritten signatures still remains a challenging open problem. This talk outlines some of the problems related with the automatic verification of random and skilled signature forgeries, and it also presents several Deep Learning-based solutions for the case of using random forgeries.