Minimizing Spatial Deformation Method for Online Signature Matching

Bin Li, Kuanquan Wang, David ZhangLecture Notes in Computer Science 2007 · DOI · OpenAlex

Abstract

Online signature verification is a critical behavioral biometric. Traditional methods like Dynamic Time Warping (DTW) often fail to capture global geometric relationships in spatial-temporal data. This paper proposes a (3+1)D matching method using Thin-Plate Spline (TPS) modeling to treat signature verification as a non-rigid spatial deformation problem. By calculating "Warping Energy" and employing a deterministic annealing framework, the method achieves an Equal Error Rate (EER) of 1.9%, significantly outperforming standard DTW approaches.

Introduction

Online signature verification presents a unique challenge in behavioral biometrics: proving identity based on the dynamic way an individual signs their name on a digital tablet. For years, the industry standard has been Dynamic Time Warping (DTW), which treats a signature as a 1D string of data points. However, signatures are inherently spatial. DTW often ignores global geometric relationships—the "shape" of the signature in 3D space (x, y, and time). If a signature is slightly stretched or deformed, DTW struggles to capture structural consistency. This paper addresses this gap by treating signature matching as a non-rigid spatial deformation problem.

Methodology

The authors propose a multi-step framework that combines geometric modeling with optimization:

  1. (3+1)D Representation: Signatures are resampled to have equidistant spacing. Each point is described by its 3D coordinates $(x, y, \text{time})$ plus a 1D feature vector containing five local attributes: horizontal displacement ($\delta x$), vertical displacement ($\delta y$), the sine and cosine of the writing angle, and absolute speed.
  2. Thin-Plate Splines (TPS): Instead of simple linear matching, the authors use TPS—a non-rigid mapping tool—to model how one signature "warps" into another. This allows the system to calculate a "Warping Energy" (WE), which quantifies how much effort is required to deform a template to match an input.
  3. Soft-Assign and Deterministic Annealing: Since point correspondences are unknown, the authors use a "soft-assign" matrix within a deterministic annealing framework. This process starts with a "fuzzy" global match and gradually "cools" into a strict, one-to-one binary mapping.
  4. Structural Constraints: To ensure physical consistency, the algorithm enforces strict monotonicity and one-to-one mapping.
3D spatial deformation and warping energy comparison between genuine and forged signatures.

Experimental Evaluation

The study utilized a custom-built database of 1,530 signatures from 102 Chinese volunteers. This included 1,020 genuine signatures and 510 "skilled forgeries" created by forgers who were allowed to observe both the final shape and the dynamic writing process of the genuine signers.

Performance Analysis

The proposed (3+1)D algorithm was compared directly against a well-known DTW-based algorithm. Performance was measured using the Equal Error Rate (EER)—the point where the rate of false acceptances equals the rate of false rejections (lower is better).

ROC curves comparing the (3+1)D method against DTW and other variants.
Method EER (3 References) EER (5 References)
Jain’s DTW-based 5.3% 3.0%
Proposed (3+1)D 3.4% 1.9%

The (3+1)D method achieved a 1.9% EER with 5 reference signatures, a significant improvement over the 3.0% achieved by the traditional DTW approach.

Conclusion

By treating time as a third spatial dimension, the (3+1)D algorithm naturally captures the rhythm and flow of writing, not just the static ink on the page. While the iterative nature of deterministic annealing and TPS modeling is computationally heavier than standard DTW, the accuracy gains demonstrate its robustness. Future work could explore more complex feature dimensions or optimize the annealing process for faster convergence.

Summary updated

BibTeX
@inproceedings{li2007minimizing,
  author = {Bin Li and Kuanquan Wang and David Zhang},
  title = {{Minimizing Spatial Deformation Method for Online Signature Matching}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2007},
  doi = {10.1007/978-3-540-74549-5\_68},
}

Loading related content...

Figure