AlaaAlowaidi/BankSigNet-140

MSFV-Net — Multi-Script Forgery Verification Network

Pretrained checkpoint for MSFV-Net, a Siamese ResNet50-based network for offline signature verification on real bank cheque documents, covering Arabic, English, and Hindi scripts simultaneously.

Paper: MSFV-Net: Explainable Multi-Script Offline Signature Verification on Real Bank Cheque Data Under review

---

Model Description

MSFV-Net is trained on BankSigNet-140 — the first offline signature dataset collected from real Indian bank cheques with printed backgrounds and three scripts in a single benchmark.

Key components:

• Backbone: Siamese ResNet50 (shared weights)
• Loss: Forged-aware dual loss (triplet + binary cross-entropy)
• Training: Script-balanced batch sampling + synthetic background augmentation
• Explainability: Difference GradCAM with Deletion/Insertion AUC faithfulness metrics

---

Performance

BankSigNet-140 (open-set signer-disjoint, mean ± std over 3 seeds):

Metric	Value
Accuracy	73.56 ± 2.16 %
EER	26.07 ± 1.45 %
AUC	0.810 ± 0.017

Per-script accuracy:

Script	ACC (%)
Hindi	88.6
English	69.2
Arabic	66.5

BHSig260-Hindi (cross-dataset):

Method	ACC (%)
SigNet	84.64
IDN	93.04
MSFV-Net	93.15

---

Usage

import torch
from model.msfvnet import MSFVNet

model = MSFVNet()
checkpoint = torch.load('verifier_msfv.pth', map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

Full training and evaluation code: github.com/Alaa-hub964/BankSigNet-140

---

Dataset

BankSigNet-140 — 2,100 signature images (140 signers × 15 samples) from real Indian bank cheques, 300 DPI, three scripts, skilled forgeries.

Available in the GitHub repository above.

---

Citation

@article{alowaidi2026msfvnet,
  title   = {MSFV-Net: Explainable Multi-Script Offline Signature Verification
             on Real Bank Cheque Data},
  author  = { Alaa Alowaidi,  and Pushpendra Kumar Pateriya },
  year    = {2026},
  journal={Al-Khwarizmi Engineering Journal}
  note    = {Under review}
}