Offline Signature Detection and Multiple Signature Verification on Financial Documents

Offline signature verification is a well studied problem due to its critical applications in areas such as banking and legal documentation. The task still remains challenging as signatures often coexist with various other elements like stamps, handwritten or typewritten text, and ruling lines, complicating detection and verification processes. In this paper, a novel solution for multiple signature verification on financial contract documents is presented using Siamese Networks. Our solution is based on a signature detection and a signature verification model, both trained on real scanned documents. These documents often contain signatures cluttered with stamps, handwritten text, and other markings such as occasionally ruling lines, adding complexity to the detection and verification processes. A key aspect of the solution is its capability to compare multiple signatures from documents with those from signers’ signature declaration documents, addressing a critical requirement for effective verification in practice. To enhance the robustness of the model, we experimented with various data augmentations, including the addition of random ruling lines to the dataset. The proposed evaluation framework is both comprehensive and unique as it incorporates both traditional performance metrics and practical usability assessments, reflecting the real-world application in banking operations. The results demonstrate the efficacy and practicality of the approach, highlighting its potential for deployment in financial institutions to improve the accuracy and reliability of signature verification processes. The proposed solution works at 95% accuracy in practice.