SigVerify: A Deep Learning Framework for Signature Authentication and Computerised Forgery Detection

Verification of offline signatures is still a difficult problem to match because of the lack of dynamic writing of information that is clubbed together like pen pressure and stroke order. Because of this, authentication would be wholly based on visual properties, and thus there would be a hard time to tell an authentic signature and one that has been well made forgery. The work introduces a deep-layered approach to enhancement of reliability of the potential application of the methods of the authentication of the statical signature. There are two levels of structure in the system. The initial step revolves around detection of digitally produced or distorted signatures in the form of a convolutional model trained to detect irregular texture and structural features. The second phase conducts specific comparison between signatures based on the use of a Siamese network, which considers the similarity based on the encoded feature representations. In order to enhance interpretability, a visual comparison mechanism is integrated to accentuate the discrepancy of reference and test signature to enable better comprehension of the mismatch areas. The system is also trained and tested on several class signature data where the system performs consistently when differentiating real and fake samples with different writing patterns.