Source-linked AI summary
SigNet: Convolutional Siamese Network for Writer Independent Offline Signature Verification
Sounak Dey, Anjan Dutta, J. Ignacio Toledo, Suman K. Ghosh, Josep Llados, Umapada Pal
TL;DR
Offline signature verification must distinguish genuine signatures from subtly deformed skilled forgeries using only static images, and writer-independent deployment must generalize across signers. The paper proposes SigNet, a convolutional Siamese network that learns distances between paired signature embeddings. Across cross-domain benchmarks, SigNet models forgeries across scripts and handwriting styles and surpasses state-of-the-art results on most benchmark datasets.
Problem
Offline verification must distinguish genuine signatures from skilled forgeries with subtle deformations, while writer-independent systems must generalize across signers without per-writer retraining.
Method
SigNet uses twin convolutional networks with shared weights and a distance-based loss to bring similar signature pairs together and separate dissimilar pairs.
Results
SigNet surpasses state-of-the-art results on most benchmark signature datasets and models forgery across different scripts and handwriting styles.
Takeaways & Limitations
The results support writer-independent feature learning for offline verification and a generic forgery prototype learned from data rather than hand-crafted features.
Abstract
from arXiv · showhide
Offline signature verification is one of the most challenging tasks in biometrics and document forensics. Unlike other verification problems, it needs to model minute but critical details between genuine and forged signatures, because a skilled falsification might often resembles the real signature with small deformation. This verification task is even harder in writer independent scenarios which is undeniably fiscal for realistic cases. In this paper, we model an offline writer independent signature verification task with a convolutional Siamese network. Siamese networks are twin networks with shared weights, which can be trained to learn a feature space where similar observations are placed in proximity. This is achieved by exposing the network to a pair of similar and dissimilar observations and minimizing the Euclidean distance between similar pairs while simultaneously maximizing it between dissimilar pairs. Experiments conducted on cross-domain datasets emphasize the capability of our network to model forgery in different languages (scripts) and handwriting styles. Moreover, our designed Siamese network, named SigNet, exceeds the state-of-the-art results on most of the benchmark signature datasets, which paves the way for further research in this direction.
1. Introduction
Offline signature verification is important but difficult because static images omit dynamic writing information and skilled forgeries can differ only subtly. SigNet addresses writer-independent verification by learning generic distinctions between genuine signatures and forgeries.
- Problem: Offline verification uses two-dimensional signature images, whereas online systems additionally capture stroke order, speed, and pressure.Offline capture generally relies on scanners or imaging devices.
- Problem: Online systems generally perform better, but require specialized hardware that raises cost and limits deployment scenarios.Offline verification remains necessary for applications such as check transactions and document verification.
- Problem: Writer-independent systems avoid retraining for every new signer by learning generic discrepancies between genuine and forged signatures.Writer-dependent systems must be updated whenever a new signer is added, creating substantial consumer-system costs.
- Contribution: The paper proposes SigNet, a convolutional Siamese network for offline signature verification that models generic forgery techniques without relying on hand-crafted features.The approach uses paired signatures and is intended for writer-independent learning.
- Problem: Signature verification is especially complex because subtle writing-style variations can overlap with degrees of forgery.SigNet is designed to mine these ultra-fine inconsistencies from training data.
2. SigNet: Siamese Network for Signature Verification
SigNet uses twin convolutional branches with shared parameters to embed signature pairs and compare them through a distance-based loss. Its architecture combines convolution, normalization, pooling, dropout, and fully connected layers to learn discriminative signature features.
- CNN Architecture: The CNN uses convolutional kernels from 11×11 to 3×3, interleaved with pooling, ReLU nonlinearities, local response normalization, and dropout.The architecture is inspired by the CNN design of Krizhevsky et al.
- Siamese Network: SigNet accepts two signature images through identical CNN branches whose shared weights produce comparable feature representations.The branches are joined by a top-level loss function that computes a similarity metric.
- CNN Architecture: The network maps each signature to a high-level feature vector before the two branches are joined by the loss function.The figure identifies this high-level output as the final orange block from the constituting CNNs.
- Siamese Network: The contrastive loss brings embeddings closer for similar pairs and pushes dissimilar-pair embeddings apart by a margin.The Euclidean distance is computed between the CNN-derived feature representations.
- Training: The model is trained with RMSprop for 20 epochs using momentum 0.9 and mini-batches of 128.Training begins with a learning rate of 1e-4 and uses ρ = 0.9 and ϵ = 1e-8.
- Interpretation: Response maps show that SigNet correctly classifies a genuine–forged pair as dissimilar using filter activations from the last convolutional layer.Corresponding activation zones indicate signature features learned for distinguishing the pair.
3. Experiments
Experiments evaluate SigNet for offline signature verification on four widely used benchmark databases. The paper compares the proposed method with existing algorithms and makes its source code available after acceptance.
- Evaluation: The evaluation uses CEDAR, GPDS300, GPDS Synthetic Signature Database, and BHSig260.These are described as widely used benchmark databases for signature verification.
3.1. Datasets
The experiments cover four datasets spanning multiple signer populations, scripts, and genuine-forged signature collections. Their construction provides benchmarks for writer-independent verification across varied signature styles.
- CEDAR: CEDAR contains 55 signers, each with 24 genuine and 24 forged signatures, totaling 1,320 genuine and 1,320 forged samples.Forgers attempted to emulate signatures from three persons, eight times each.
- GPDS300: GPDS300 contains 300 persons with 24 genuine and 30 forged signatures per person, totaling 7,200 genuine and 9,000 forged samples.The genuine signatures were collected in one day and shown to forgers for imitation.
- GPDS Synthetic: GPDS Synthetic contains 4,000 signers with 24 genuine and 30 forged signatures each, yielding 96,000 genuine and 120,000 forged signatures.The database is built using a synthetic individuals protocol.
- BHSig260: BHSig260 contains 260 persons: 100 signed in Bengali and 160 in Hindi, with 24 genuine and 30 forged signatures per signer.The resulting totals are 2,400 genuine and 3,000 forged Bengali signatures, plus 3,840 genuine and 4,800 forged Hindi signatures.
3.2. Performance Evaluation
SigNet evaluates signature pairs by thresholding learned distances and selecting the maximum balanced accuracy across thresholds.
- A threshold d classifies signature pairs as similar or dissimilar using SigNet's distance measure D(x_i, x_j).Pairs with the same identity are Psimilar; pairs with different identities are Pdissimilar.
- True positives comprise similar pairs whose SigNet distance is at most d.
- The evaluation defines true-positive and true-negative rates for each distance threshold.
- Accuracy is computed as one-half the sum of TPR(d) and TNR(d).
- The reported accuracy is the maximum over thresholds from the minimum to maximum distance, evaluated in 0.01 increments.
3.3. Experimental Protocol
The experiments use writer-independent signer splits and also test a setting where genuine signatures from other writers serve as unskilled forgeries.
- For each dataset, M of K available signers are used for training and the remaining K − M signers for testing.All original and forged signatures from the selected training signers are retained.
- Each author contributes 276 genuine-genuine pairs because the datasets contain 24 genuine signatures per author.
- Table 3 reports the K and M values used for different datasets.
- The protocol additionally evaluates training with genuine signatures paired against other signers' genuine signatures as unskilled forgeries.This setting addresses cases where forged training samples are difficult to obtain.
3.4. Results and Discussions
SigNet outperforms prior methods on three benchmark datasets, while genuine-forged training is stronger than training with genuine-unskilled forged pairs.
- SigNet outperformed state-of-the-art methods on GPDS Synthetic, Bengali, and CEDAR datasets.The comparison is reported in Table 4 across multiple signature databases.
- The GPDS300 result was lower, possibly because fewer signature samples had to represent many different signature styles.
- On GPDS Synthetic, SigNet outperformed Dutta et al.'s method, possibly because more training samples represented the available signature styles.
- Training on genuine-forged pairs outperformed training on genuine-unskilled forged pairs.The paper links this difference to the need to learn minute signature-specific forgery details.
GPDS Synthetic
Cross-dataset evaluation trains SigNet on one corpus and tests it on the others, with results summarized in Figure 3.
- Figure 3 reports cross-dataset accuracies, with training datasets in rows and testing datasets in columns.
- Each cross-dataset experiment trains SigNet on one dataset and tests it on every other corpus.The process is repeated with each dataset serving as the training corpus.
- The highest accuracy for every dataset is obtained when the model is trained and tested on the same dataset.
4. Conclusions
The paper presents SigNet, a Siamese-network framework for writer-independent offline signature verification that learns features from data rather than relying on hand-crafted features. Experiments show modelling across handwriting styles and scripts, with results surpassing state-of-the-art performance on most benchmark datasets.
- SigNet performs writer-independent feature learning for offline signature verification without relying on hand-crafted features.The framework learns features directly from data.
- Experiments on the GPDS Synthetic dataset support modelling a generic prototype for real forgeries from synthetically generated data.
- Cross-domain experiments show that the architecture models forgeries across diverse handwriting styles, signers, backgrounds, and scripts.
- SigNet surpasses state-of-the-art results on most benchmark signature datasets, motivating further research.Future work includes developing richer network models and exploring other verification frameworks.