PaperlayerView exact revision on arXiv

Source-linked AI summary

Offline Handwritten Signature Verification - Literature Review

Luiz G. Hafemann, Robert Sabourin, Luiz S. Oliveira

arXiv:1507.07909v4cs.CVstat.ML

TL;DR

Offline handwritten signature verification must distinguish genuine signatures from forgeries despite limited samples, partial training knowledge, and variability in authentic signatures. This paper reviews how the problem has been addressed over recent decades, including preprocessing, feature representations, classification, datasets, and data augmentation. It reports that error rates have dropped significantly, largely alongside deep-learning advances, while identifying limited-sample classification and one-class models as future research areas.

  • Problem

    Offline signature verification is challenging because training has only genuine enrolled-user signatures, users provide few samples, and authentic signatures vary substantially.

  • Method

    The paper reviews decades of offline verification research, covering preprocessing, feature extraction, model training, datasets, and methods for limited samples.

  • Results

    Error rates have dropped significantly in recent years, mostly due to advances in applying deep learning to offline signature verification.

  • Takeaways & Limitations

    Learned feature representations can generalize to new users and users from other datasets, while dissimilarity-based and metric-learning methods address limited samples.

  • Takeaways & Limitations

    One-class classification remains insufficiently addressed, especially for systems that must work with few samples per user.

Abstract

from arXiv · show

The area of Handwritten Signature Verification has been broadly researched in the last decades, but remains an open research problem. The objective of signature verification systems is to discriminate if a given signature is genuine (produced by the claimed individual), or a forgery (produced by an impostor). This has demonstrated to be a challenging task, in particular in the offline (static) scenario, that uses images of scanned signatures, where the dynamic information about the signing process is not available. Many advancements have been proposed in the literature in the last 5-10 years, most notably the application of Deep Learning methods to learn feature representations from signature images. In this paper, we present how the problem has been handled in the past few decades, analyze the recent advancements in the field, and the potential directions for future research.

I. INTRODUCTION

Handwritten signature verification uses biometric signatures to determine whether a claimed identity’s sample is genuine or forged. Offline systems analyze static signature images, and this review organizes prior work, datasets, pipeline techniques, recent progress, and future directions.

  • Biometric setting: Biometric systems distinguish verification, which checks a claimed identity, from identification, which searches for a sample among enrolled users.The review frames signature verification as the former scenario.
  • Motivation: Handwritten signatures are behavioral biometric traits widely used for identity verification in legal, financial, and administrative settings.Their collection is non-invasive and familiar in daily life.
  • Verification task: Verification systems classify a query signature as genuine or as one of three common forgery types: random, simple, or skilled.Random forgeries use the forger’s own signature, while simple forgeries use knowledge of the user’s name; the supplied passage truncates the skilled-forgery description.
  • Offline setting: Offline verification represents a completed signature as a digital image, unlike online verification, which records time-ordered pen data and may include pressure or inclination.The acquisition method determines whether the system is categorized as online or offline.
  • Review scope: The paper surveys decades of research, formalizes the problem, reviews datasets and training-pipeline techniques, and summarizes recent progress and future research areas.The described pipeline includes preprocessing, feature extraction, and model training.

II. PROBLEM STATEMENT

Offline signature verification must distinguish genuine signatures from forgeries despite substantial within-user variation, skilled-forgery similarity, limited training knowledge, and few samples per enrolled user.

  • Problem formulation: The verification model is trained on a learning set and classifies a claimed user’s query signature as genuine or forgery using a test set of genuine signatures and forgeries.The paper describes enrollment and subsequent operations or classification phases.
  • System settings: Writer-independent systems use one model for any user, whereas writer-dependent systems train one model per user.Writer-independent evaluation separates development users from exploitation users enrolled in the system.
  • Scope: The review excludes methods relying on skilled forgeries for the enrolled users’ training set.The restriction targets methods matching the stated practical training scenario.
  • Challenges: High intra-class variability makes signatures from the same user differ substantially, while skilled forgeries create low inter-class variability by resembling genuine signatures.Figure 1 illustrates the within-user variability.
  • Challenges: Training commonly has partial knowledge because genuine signatures are available for enrolled users, but their forgeries are not.The classifier therefore lacks direct information about what distinguishes genuine signatures from enrolled-user forgeries.
  • Challenges: Real applications often provide only a few signature samples per user, requiring good performance for new users with small enrollment sets.This limitation persists even when many users are enrolled.

III. DATASETS

The review emphasizes public datasets as a basis for comparison and describes preprocessing steps that standardize and represent scanned signatures before feature extraction.

  • Datasets: Private datasets make it difficult to determine whether performance improvements reflect better methods or cleaner, simpler databases.Public signature datasets emerged in the last decade to address this comparison gap.
  • Datasets: Public-dataset collection typically gathers genuine signatures across sessions using forms with multiple cells, while forgeries are collected by asking users to imitate genuine samples.The supplied passage truncates the final description of the forgery-collection process.
  • Datasets: Table I summarizes the signature datasets most commonly used in the literature.The supplied materials provide the table title and caption but not its dataset entries.
  • Preprocessing: Preprocessing addresses variations in pen thickness, scale, and rotation that can occur even among authentic signatures from one person.The review presents preprocessing as an important part of signature verification.
  • Preprocessing: Signature extraction locates and crops a signature from its document, a difficult step for bank cheques with complex backgrounds.Most studies begin with already extracted signatures and therefore omit this step.
  • Preprocessing: Noise removal uses filters such as median filtering and morphological operations to reduce noise, fill small holes, and remove small connected regions.These operations are applied to scanned signature images.
  • Preprocessing: Size normalization and centering may use tight or narrower bounding boxes, fixed frame sizes, and image-centroid alignment.The supplied passage truncates the final centering description.
  • Preprocessing: Signature representations can include grayscale images, skeletons, outlines, ink distributions, high-pressure regions, and directional frontiers.These alternatives provide inputs beyond the raw graylevel image.

V. FEATURE EXTRACTION

Offline signature verification uses diverse static or pseudo-dynamic and global or local feature-extraction strategies. Recent work increasingly learns representations directly from signature images instead of manually designing extractors.

  • Feature representations: Feature extraction methods are broadly categorized as static or pseudo-dynamic, with pseudo-dynamic features attempting to recover execution information such as speed or pressure.They are also categorized as global or local according to the scope of the described signature information.
  • Feature representations: Recent representation-learning methods learn feature representations directly from signature images rather than designing task-specific feature extractors.The cited studies approach offline verification from a representation-learning perspective.

A. Handcrafted feature extractors

Handcrafted feature extraction for offline signatures spans global and local shape descriptors, graphometric measures, and directional representations of strokes.

  • A. Handcrafted feature extractors: The literature summarizes handcrafted descriptors as a major focus of offline signature research, with multiple feature representations proposed for the task.These passages cover geometric, graphometric, and directional alternatives.
  • A. Handcrafted feature extractors: Geometric features describe overall signature shape using dimensions, caliber, area, endpoints, closed loops, or grid-based pixel density.Caliber is defined as the image height-to-width ratio.
  • A. Handcrafted feature extractors: Graphometric descriptors adapt forensic handwriting measurements into algorithmic static features for automated signature verification.The cited work selected graphometric features that could be described algorithmically.
  • A. Handcrafted feature extractors: Directional features characterize signature images through the directions of their strokes.Examples include Directional-PDF features from gradients and PHOG histograms of edge orientations across multiple scales.

4) Mathematical transformations:

Signature verification has used mathematical transforms, projection-based descriptors, texture histograms, and local interest-point matching to extract image features.

  • 4) Mathematical transformations:: Mathematical transformations used as feature extractors include the Hadamard, Contourlet, discrete Radon, Wavelet, and Fractal transforms.The cited studies associate these transforms with spectrum analysis, smooth-contour capture, HMM observations, or signature verification.
  • 4) Mathematical transformations:: Extended Shadow Code overlays directional bars on the signature image and counts activated bins in projections as a descriptor.The grid contains horizontal, vertical, and diagonal bars, each divided into fixed bins.
  • 4) Mathematical transformations:: Local Binary Pattern variants describe local image patterns through histograms and are reported among the best handcrafted extractors for this task.GLCM is identified as another texture descriptor.
  • 4) Mathematical transformations:: SIFT-based matching extracts local interest points from query and reference signatures to build writer-dependent classification features.Classification uses the number of query keypoints and matched keypoints, with only stable interest points used for matching.

8) Pseudo-dynamic features:

The reviewed literature contrasts handcrafted representations with learned features, including early deep-learning work and two recent strategies for skilled-forgery classification.

  • 8) Pseudo-dynamic features:: Pseudo-dynamic graphometric features estimate signing-process properties such as stroke tension, speed, continuity, uniformity, slant, and form.These descriptors are derived from pixel distribution and geometric characteristics rather than directly recorded signing dynamics.
  • 8) Pseudo-dynamic features:: Curvature-based descriptors fit Bezier curves to the signature outline and use the curve parameters as features.The fitting focuses on the largest segment of the signature.
  • 8) Pseudo-dynamic features:: Deep Learning methods learn feature representations directly from raw signature pixels instead of relying on hand-engineered extractors.This represents a recent shift toward representation learning for signature images.
  • 8) Pseudo-dynamic features:: Early representation-learning studies used private datasets or limited forgery settings and reported limited evidence of broad verification success.RBM work did not report discrimination results, while CNN experiments considered only random forgeries.
  • 8) Pseudo-dynamic features:: Recent skilled-forgery research follows either learned writer-independent features with writer-dependent classifiers or joint writer-independent metric learning.Metric learning trains distances to bring genuine pairs closer and separate forgeries from references.

VI. MODEL TRAINING

Model training in signature verification includes writer-dependent and writer-independent classifiers, hybrid strategies for limited data, and sequence-based HMM approaches.

  • VI. MODEL TRAINING: Writer-dependent systems train a separate model for each user using genuine signatures and random forgeries from other users.At operation time, the model associated with the claimed identity classifies the query signature.
  • VI. MODEL TRAINING: Writer-independent systems provide an alternative classifier organization to writer-dependent models.The supplied passage introduces the two broad classifier groups but truncates the writer-independent description.
  • VI. MODEL TRAINING: Hybrid systems use writer-independent classification when few genuine samples are available, then train writer-dependent classifiers after a sample threshold is reached.This strategy is described for solutions combining both approaches.
  • VI. MODEL TRAINING: Beyond thresholding and nearest-neighbor methods, the literature studies several model strategies for signature verification.The reviewed models include Hidden Markov Models.
  • VI. MODEL TRAINING: Left-to-right HMMs model grid-column observations, using quantized cell features and observation likelihoods during verification.The topology has been studied because it matches the left-to-right hand movements described for American and European handwriting.

B. Support Vector Machines

Support Vector Machines have been widely used in both writer-dependent and writer-independent signature verification, while newer neural and ensemble approaches extend classifier design.

  • Support Vector Machines: Support Vector Machines have been extensively used for both writer-dependent and writer-independent signature verification.They have been reported as among the most effective classifiers for the task.
  • Support Vector Machines: One-Class SVMs model only genuine signatures, matching settings where enrolled users provide genuine training samples.The low number of genuine signatures remains an important challenge for this strategy.
  • Support Vector Machines: Neural Networks have been explored for writer-dependent and writer-independent systems, including multi-resolution classification and comparisons with SVMs.Huang and Yan trained multiple networks on features extracted at different resolutions, followed by a decision network.
  • Support Vector Machines: Deep Multitask Metric Learning compares signatures by learning a distance metric with shared lower layers and user-specific final layers.The shared layers use common weights across users, while the last layer specializes for each individual.
  • Support Vector Machines: Classifier ensembles combine predictions from multiple models using static or dynamic selection strategies.Examples include genetic selection from classifiers trained with different grid sizes and dynamic selection based on HMM likelihoods.

E. Data augmentation

Offline signature verification faces limited training data per user, motivating synthetic-sample generation and standardized dataset comparisons across methods.

  • E. Data augmentation: Low numbers of samples per user are a main challenge in building automated signature verification systems.Researchers address this constraint by generating additional samples from existing genuine signatures.
  • E. Data augmentation: Signature perturbations such as slant, rotation, scaling, and perspective can generate new genuine samples and distorted forgeries.Huang and Yan distinguished slight distortions for genuine samples from heavy distortions for forgeries.
  • E. Data augmentation: Synthetic signature generation has also been inspired by a neuromotor model.Ferrer and colleagues proposed this synthesis approach in recent work.
  • E. Data augmentation: Comparisons require standard datasets and similar experimental protocols, but training and testing procedures commonly differ across studies.Differences include the number of reference signatures, random-forgery selection, and reported metrics.
  • E. Data augmentation: The review consolidates state-of-the-art results on the GPDS, MCYT, and CEDAR datasets.GPDS results are grouped across releases that added increasing numbers of users.
  • E. Data augmentation: Feature learning methods showed potential and achieved the best performance on the reviewed datasets.The review identifies texture and directional descriptors as particularly important among earlier feature extractors.

VII. CONCLUSION

The conclusion identifies improved representations, limited-sample classification, dataset augmentation, and classifier ensembles as major directions in offline signature verification. It highlights deep learning and one-class methods as promising areas for continued research.

  • VII. CONCLUSION: Error rates have dropped significantly in recent years, mostly due to applying Deep Learning to offline signature verification.Distinguishing genuine signatures from skilled forgeries remains challenging.
  • VII. CONCLUSION: Feature learning can produce representations that generalize from a subset of users to new users and users from other datasets.The conclusion presents learned representations from signature images as a continuing research direction.
  • VII. CONCLUSION: Dissimilarity-based writer-independent and metric-learning solutions are promising for classification when only a small number of samples per user is available.This direction addresses severe sample constraints in practical applications.
  • VII. CONCLUSION: Synthetic signature generation can augment datasets by increasing the number of samples available for training.This direction is linked to the limited number of samples per user.
  • VII. CONCLUSION: Static and dynamic classifier ensembles have been investigated to increase classification accuracy and solution robustness.Dynamic selection methods are identified as particularly promising future directions.
  • VII. CONCLUSION: One-class classification remains insufficiently addressed despite theoretically matching the verification problem and requiring effective operation with few samples per user.The authors identify low-sample one-class systems as an interesting area for future research.
Loading 1507.07909v4…