Signature verification approach using fusion of hybrid texture features

Ankan Kumar Bhunia, Alireza Alaei, Partha Pratim Roy

TL;DR

This paper proposes a novel offline signature verification system that fuses two distinct texture features, Local Phase Quantization (LPQ) and Wavelet Transform (DWT), using a score-level fusion technique with one-class Support Vector Machines (SVMs). This approach effectively handles the challenges of intra-class variability and limited training data by focusing solely on genuine signatures, achieving state-of-the-art performance across multiple benchmark datasets.

What Problem Does This Paper Address?

Handwritten signature verification is a critical biometric modality for person identification and authentication. However, offline signature verification, which relies solely on static images, presents significant challenges. These include high intra-class variability (the same person's signatures can look quite different) and low inter-class variability (forgeries can closely mimic genuine signatures), especially skilled forgeries. Furthermore, obtaining a sufficient number of genuine and forged signatures for training robust writer-dependent models is often difficult in real-world scenarios. Existing methods often struggle to balance accuracy with the need for writer-specific models and limited training data.

Methods

The proposed method tackles these challenges through a multi-stage approach:

  • Pre-processing: Standard image pre-processing techniques are applied, including binarization (using Otsu's algorithm), noise reduction (Gaussian filter), and cropping to achieve translation invariance.
  • Feature Extraction: Two complementary texture feature extraction methods are employed:
  • Local Phase Quantization (LPQ): This method captures spatial blurring characteristics of the image by analyzing local spectral phase information. It's known for its compact feature representation.
  • Discrete Wavelet Transform (DWT): This transform decomposes the image into different frequency sub-bands (LL, LH, HL, HH), providing both approximation and detail information that characterizes texture.
  • One-Class SVM Classification: For each writer, two separate one-class SVMs are trained – one for LPQ features and another for DWT features. One-class SVMs are chosen because they only require genuine signatures for training, aligning with the practical constraint of limited forgery data. These SVMs output an "authenticity score" for each feature set.
  • Score-Level Fusion: A novel score-level fusion rule is proposed to combine the scores from the two one-class SVMs. Specifically, the average of the LPQ and DWT scores is calculated. This fused score is then compared against an acceptance threshold to make the final genuine/forgery decision.

Datasets

The proposed method was evaluated on four publicly available offline signature datasets:

  • MCYT: Contains 2250 signatures from 75 signers, with 15 genuine and 15 forged signatures per class.
  • GPDS-300: Comprises 16,200 signatures from 300 signers, with 24 genuine signatures per signer.
  • BHSig260: Includes 6,240 genuine and 7,800 forged signatures from 260 individuals, with signatures in Bengali and Hindi.
  • CEDAR: Consists of 1,320 genuine and 1,320 forged signatures from 55 signers, with 24 genuine and 24 forged signatures per signer.

Key Insights and Ideas

  • Complementary Nature of LPQ and Wavelet Features: The paper highlights that LPQ and Wavelet features capture different aspects of signature images. LPQ focuses on spatial blurring and compact representation, while Wavelet captures multi-resolution textural information. Their fusion proves more effective than using either alone.
  • Effectiveness of One-Class SVM for Limited Data: The use of one-class SVMs is a strategic choice that directly addresses the common problem of having only genuine signatures available for training, making the system practical for real-world deployment.
  • Score-Level Fusion for Robustness: Combining scores from different feature extractors at the decision level (rather than fusing features themselves) is a robust strategy that can improve overall performance by leveraging the strengths of each individual classifier.
  • Writer-Dependent Models are Key: The approach emphasizes creating individual models for each writer, which is crucial for handling the high intra-class variability inherent in signatures.

Key Results

The proposed fusion method consistently outperformed systems using only LPQ or Wavelet features individually across all tested datasets. The average error rates (AER) were significantly reduced, with improvements ranging from 3% to 4%. The paper demonstrates that as the number of training signatures per user increases, the AER generally decreases. The system achieved state-of-the-art performance, particularly when 8 or 12 genuine signature samples were used for training. For instance, on the MCYT dataset with 12 training samples, the proposed method achieved an AER of 8.56%, outperforming other reported methods.

Limitations and Future Work

  • Parameter Tuning: The performance of the one-class SVMs and the overall system is dependent on tuning parameters like $\theta$, $\sigma$, and $k$. While the paper explores optimal parameter selection, this tuning process adds complexity.
  • Generalization to Other Languages/Scripts: While the BHSig260 dataset includes Bengali and Hindi, the primary focus and evaluation are on datasets that might be more representative of Latin scripts. Further validation on a wider range of scripts could be beneficial.
  • Online Signatures: The current work is focused on offline signatures. Extending the fusion approach to online signature verification, which includes dynamic information, could be a future direction.
  • Exploration of Other Fusion Strategies: While score-level fusion is effective, exploring other fusion techniques or more advanced machine learning models for combining the scores could potentially yield further improvements.

Summary updated

Abstract

In this paper, a writer-dependent signature verification method is proposed. Two different types of texture features, namely Wavelet and Local Quantized Patterns (LQP) features, are employed to extract two kinds of transform and statistical based information from signature images. For each writer two separate one-class support vector machines (SVMs) corresponding to each set of LQP and Wavelet features are trained to obtain two different authenticity scores for a given signature. Finally, a score level classifier fusion method is used to integrate the scores obtained from the two one-class SVMs to achieve the verification score. In the proposed method only genuine signatures are used to train the one-class SVMs. The proposed signature verification method has been tested using four different publicly available datasets and the results demonstrate the generality of the proposed method. The proposed system outperforms other existing systems in the literature.
BibTeX
@article{bhunia2019signature,
  author = {Ankan Kumar Bhunia and Alireza Alaei and Partha Pratim Roy},
  title = {{Signature verification approach using fusion of hybrid texture features}},
  journal = {Neural Computing and Applications},
  year = {2019},
  url = {https://openreview.net/forum?id=WJdKxv7eGx},
}
Figure