Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Signature Forgery Detection: Improving Cross-Dataset Generalization

Matheus Ramos Parracho

TL;DR

The paper addresses the cross-dataset generalization gap in offline signature forgery detection by evaluating two input representations—raw signature images and a novel shell-based preprocessing pipeline—across three benchmarks (CEDAR, ICDAR, GPDS Synthetic). It leverages Siamese neural networks with contrastive and triplet losses to learn robust embeddings and assesses generalization through writer-disjoint, cross-dataset testing, using AUC-based ROC analysis. Key findings show that raw-image models achieve stronger in-domain performance, while shell-based representations regularize cross-dataset behavior and offer potential for future improvements. The work provides a scalable framework for cross-domain signature verification and highlights directions such as improved sampling, advanced losses, and refined shell features to close the generalization gap.

Abstract

Automated signature verification is a critical biometric technique used in banking, identity authentication, and legal documentation. Despite the notable progress achieved by deep learning methods, most approaches in offline signature verification still struggle to generalize across datasets, as variations in handwriting styles and acquisition protocols often degrade performance. This study investigates feature learning strategies for signature forgery detection, focusing on improving cross-dataset generalization -- that is, model robustness when trained on one dataset and tested on another. Using three public benchmarks -- CEDAR, ICDAR, and GPDS Synthetic -- two experimental pipelines were developed: one based on raw signature images and another employing a preprocessing method referred to as shell preprocessing. Several behavioral patterns were identified and analyzed; however, no definitive superiority between the two approaches was established. The results show that the raw-image model achieved higher performance across benchmarks, while the shell-based model demonstrated promising potential for future refinement toward robust, cross-domain signature verification.

Signature Forgery Detection: Improving Cross-Dataset Generalization

TL;DR

The paper addresses the cross-dataset generalization gap in offline signature forgery detection by evaluating two input representations—raw signature images and a novel shell-based preprocessing pipeline—across three benchmarks (CEDAR, ICDAR, GPDS Synthetic). It leverages Siamese neural networks with contrastive and triplet losses to learn robust embeddings and assesses generalization through writer-disjoint, cross-dataset testing, using AUC-based ROC analysis. Key findings show that raw-image models achieve stronger in-domain performance, while shell-based representations regularize cross-dataset behavior and offer potential for future improvements. The work provides a scalable framework for cross-domain signature verification and highlights directions such as improved sampling, advanced losses, and refined shell features to close the generalization gap.

Abstract

Automated signature verification is a critical biometric technique used in banking, identity authentication, and legal documentation. Despite the notable progress achieved by deep learning methods, most approaches in offline signature verification still struggle to generalize across datasets, as variations in handwriting styles and acquisition protocols often degrade performance. This study investigates feature learning strategies for signature forgery detection, focusing on improving cross-dataset generalization -- that is, model robustness when trained on one dataset and tested on another. Using three public benchmarks -- CEDAR, ICDAR, and GPDS Synthetic -- two experimental pipelines were developed: one based on raw signature images and another employing a preprocessing method referred to as shell preprocessing. Several behavioral patterns were identified and analyzed; however, no definitive superiority between the two approaches was established. The results show that the raw-image model achieved higher performance across benchmarks, while the shell-based model demonstrated promising potential for future refinement toward robust, cross-domain signature verification.
Paper Structure (102 sections, 42 equations, 91 figures, 15 tables, 6 algorithms)

This paper contains 102 sections, 42 equations, 91 figures, 15 tables, 6 algorithms.

Table of Contents

  1. Introduction
  2. Objective
  3. General Objective
  4. Specific Objectives
  5. Methodology
  6. Dataset Preparation
  7. Model Development and Training
  8. Cross-Dataset Evaluation
  9. Analysis and Visualization
  10. Develop and implement the Shell Algorithm
  11. Structure of the Work
  12. The Theory Behind Signature Recognition
  13. What is a Neural Network?
  14. The Perceptron: The first and simplest Neural Network
  15. Training the Perceptron
  16. ...and 87 more sections

Figures (91)

  • Figure 1: Similarity between neurons in the brain and neurons on neural networks
  • Figure 2: The Perceptron Schema
  • Figure 3: The Perceptron Schema
  • Figure 4: Loss landscapes without skip connections. It shows the loss function of a 56-layer and 110-layer net using the CIFAR-10 dataset, without residual connections. li2018visualizing
  • Figure 5: Data Augmentation Examples
  • ...and 86 more figures