Cross-dataset and Cross-script Evaluation of Siamese Networks for Offline Signature Verification in Banking Documents
Year: 2026
Type: article
Abstract: Banks continue to depend on handwritten signatures, yet a verifier that scores well on its own benchmark can act very differently once it meets documents it has never seen. This paper examines how a Siamese convolutional network for writer-independent offline signature verification carries over between datasets, scripts and scan conditions. We train on writer-disjoint splits of CEDAR (English) and BHSig260 (Hindi and Bengali), using contrastive and triplet loss, and then test on unseen writers, on the other datasets, and on images corrupted with noise, blur, ruling lines and stamp marks. The decision threshold is set once on validation data from the training dataset and is never adjusted afterwards. Under contrastive loss the equal error rate (EER) averages 7.8% within a dataset but 19.3% across datasets. The most telling case is a CEDAR-trained model that still ranks Hindi and Bengali pairs usefully (AUC 0.868 and 0.890) while accepting 98.6% and 97.7% of forged pairs at its fixed threshold. Accuracy quoted at a best-case threshold would conceal this failure. Degraded scans raise the EER to between 30.1% and 50.1%. All results are means over 3 random seeds. Training on degraded images lowers the average EER on degraded scans from 41.8% to 20.9% (one seed), at a small cost on clean images. We propose no new architecture. Instead, we offer a controlled and reproducible account of when such a verifier can be trusted on new data and when it needs recalibration.
Author: Sanchita Jain,Dr. Ankur Jain
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236600738
