11
Sources & Publications
6
Forensic Datasets
2.8k+
Training Images (Detection)
88.7%
mAP50 — Sig Detection
Publications

Peer-Reviewed Research

The following peer-reviewed papers directly underpin the algorithms, methodologies, and standards implemented in ForensicsFlow. All DOIs are verified and link to the published sources.

PNAS · Open Access

Accuracy and Reliability of Forensic Handwriting Comparisons

The largest-scale study to date measuring the accuracy, reproducibility, and repeatability of forensic handwriting comparison decisions by practising forensic document examiners. 86 examiners made 3,713 decisions across samples spanning a range of quality, quantity, and casework attributes. Results establish the empirical error rates that calibrated AI systems must be measured against — the foundation for Score-Based Likelihood Ratio validation in ForensicsFlow.

SLR Calibration Error Rates FDE Accuracy
PNAS, 119(32):e2119944119 · Aug 2022 · DOI: 10.1073/pnas.2119944119 Full Paper ↗
Statistical Analysis & Data Mining · Wiley

Handwriting Identification Using Random Forests and Score-Based Likelihood Ratios

Introduces the core methodology implemented in ForensicsFlow's Handwriting Static Analysis module: decomposing handwriting into graphical units, assigning them to 40 exemplar clusters, and using the cluster-frequency vector as a feature space for Random Forest comparison. Score-Based Likelihood Ratios are computed from the forest similarity scores against empirical same-writer and different-writer distributions. Funded by NIST/CSAFE cooperative agreement.

Random Forest Score-Based LR Graph Clusters
Stat. Analysis Data Mining, 2022 · DOI: 10.1002/sam.11566 Full Paper ↗
NeurIPS 2021 · arXiv:2106.00666

You Only Look at One Sequence: Rethinking Transformer in Vision Through Object Detection

Introduces YOLOS — the Vision Transformer (ViT) architecture fine-tuned for object detection using DETR bipartite matching loss, treating the full image as a single patch sequence. YOLOS-Base achieves 42.0 box AP on COCO val with 127M parameters. This paper underpins the Detect Signature module: fine-tuned on forensic document corpora, it achieves mAP50 = 0.887 for handwritten signature bounding box prediction without requiring region proposal networks or non-maximum suppression.

YOLOS Vision Transformer DETR Object Detection
NeurIPS 2021 · arXiv:2106.00666 · Fang et al. arXiv ↗
Nature Scientific Reports · 2023

Handwriting Identification and Verification Using Artificial Intelligence-Assisted Textural Features

Presents a Spatial Variation-dependent Verification (SVV) scheme using textural features to authenticate handwritten signatures across different writing surfaces and instruments. The study demonstrates that AI-assisted textural analysis reliably discriminates genuine from forged signatures under varying document conditions — informing the preprocessing and feature extraction robustness requirements of the ForensicsFlow static analysis pipeline.

Textural Features Signature Auth AI Verification
Scientific Reports, 2023 · DOI: 10.1038/s41598-023-48789-9 PMC ↗
Mathematics (MDPI) · 2024

Enhancing Signature Verification Using Triplet Siamese Similarity Networks in Digital Documents

Develops a triplet Siamese network architecture for offline signature verification — extending the standard contrastive Siamese approach with a three-way triplet loss that simultaneously enforces similarity to genuine specimens and dissimilarity from forgeries. The method significantly reduces the Equal Error Rate on standard benchmark corpora and underpins the forgery classification approach in ForensicsFlow's Signature Comparison module.

Siamese Network Triplet Loss Forgery Detection
Mathematics 12(17):2757 · 2024 · Tehsin et al. Full Paper ↗
Scientific Programming · Wiley · 2022

A Comparative Study Among Handwritten Signature Verification Methods Using Machine Learning Techniques

Provides a systematic benchmark comparison of machine learning approaches for signature verification — including SVM, CNN, Random Forest, and hybrid architectures — evaluated across CEDAR and other public datasets. The study's equal-error-rate analysis across methods informed the multi-algorithm ensemble strategy used in ForensicsFlow, where no single model is relied upon exclusively for the final forensic opinion.

SVM CNN Benchmark CEDAR
Scientific Programming 2022:8170424 · DOI: 10.1155/2022/8170424 Full Paper ↗
ENFSI · Best Practice Manual · 2022

Best Practice Manual for the Forensic Examination of Handwriting (ENFSI-BPM-FHX-01, Ed. 3)

The European Network of Forensic Science Institutes' authoritative best practice manual for forensic handwriting examination — covering examination methodology, evidence evaluation using Likelihood Ratios, reporting standards, and quality assurance requirements. ForensicsFlow's report engine, SLR output format, and verbal scale conform to the guidelines established in this document. All reports generated on the Laboratory and Enterprise plans are structured to satisfy this standard.

ENFSI Standard LR Reporting Best Practice
ENFSI-BPM-FHX-01 Ed.3 · 2022 · enfsi.eu PDF ↗
Expert Systems with Applications · Springer · 2023

Writer Independent Handwritten Signature Verification on Multi-Scripted Signatures Using Hybrid CNN-BiLSTM

Proposes a hybrid CNN-BiLSTM architecture for multi-script offline signature verification that operates without writer-specific enrolment — critical for forensic scenarios where no reference signatures exist. The bidirectional LSTM captures both forward and reverse temporal dependencies in stroke feature sequences. This writer-independent approach directly informs the open-set verification capability of ForensicsFlow's Signature Comparison module.

CNN-BiLSTM Multi-Script Writer-Independent
Expert Systems with Applications 214:119111 · 2023 · DOI: 10.1016/j.eswa.2022.119111 Full Paper ↗
Training & Validation Data

Built on Verified Forensic Corpora

Every model deployed in ForensicsFlow is trained and validated on publicly available, peer-reviewed forensic datasets — ensuring reproducibility, transparency, and resistance to overfitting on proprietary data.

CSAFE Handwriting Database
Funded by NIST/CSAFE cooperative agreements 70NANB15H176 & 70NANB20H019
The primary training corpus for the Random Forest Score-Based LR system. Contains handwriting samples across multiple prompts and sessions from a diverse writer population. Used with partner universities including Carnegie Mellon, Duke, UC Irvine, and the University of Virginia.
CEDAR Handwriting & Signature Database
University at Buffalo · SUNY · Public benchmark
One of the most widely cited benchmark databases for offline signature verification and handwriting authorship identification. Provides 2,640 genuine and 2,640 forged signature pairs across 55 writers. Used for SVM and Siamese network validation in the Signature Comparison module.
GPDS-960 Signature Corpus
960 subjects · 24 genuine + 30 skilled forgeries per subject
The standard large-scale benchmark for offline signature verification research. 24 genuine samples and 30 skilled forgeries per writer across 960 subjects. Used for Siamese network training and cross-validation of the signature comparison engine's EER performance.
Tobacco800 Document Corpus
800+ scanned document images · Mixed types · Real-world degradation
A real-world forensic document corpus containing scanned legal, financial, and corporate documents with natural degradation, stamps, and mixed content. Primary source for YOLOS fine-tuning data — contributing the document background diversity critical for robust signature region detection.
Forensic Signature Detection Dataset
2,819 labelled images · 70/15/15 split · COCO JSON · 640×640px
Unified COCO JSON dataset combining Tobacco800 and a dedicated signatures corpus, standardised at 640×640 pixel resolution. Used exclusively for YOLOS fine-tuning: 1,980 training / 420 validation / 419 test images. Achieves mAP50 = 0.887 and mAP50-95 = 0.495 on the test split.
IAM Handwriting Database
657 writers · 1,539 pages · English connected text
The IAM database contains complete forms of unconstrained handwritten text scanned at 300 dpi. Used as supplementary training data for the multi-style concurrent comparison engine and for evaluating the graph decomposition pipeline's robustness across diverse handwriting styles and layouts.
Scientific Partnerships

Research Collaborations

ForensicsFlow actively engages with academic institutions, government forensic science bodies, and international standards organisations to validate and advance its algorithmic foundations.

CSAFE — Center for Statistics and Applications in Forensic Evidence
Academic Research · Iowa State University
NIST/CSAFE cooperative agreement research underpins the Random Forest Score-Based Likelihood Ratio methodology implemented in the Handwriting Static Analysis module. CSAFE's work on quantifying forensic evidence strength is the scientific foundation for all SLR outputs in ForensicsFlow.
ENFSI — European Network of Forensic Science Institutes
International Standards Body
ForensicsFlow report outputs conform to ENFSI BPM-FHX-01 (Edition 3, 2022) — the Best Practice Manual for Forensic Handwriting Examination — including LR verbal scale, methodology documentation requirements, and evaluative reporting guidelines for expert witness evidence.
OSAC — Organization of Scientific Area Committees (NIST)
US Federal Standards · NIST
Alignment with OSAC Forensic Document Examination Subcommittee (FDX-1) standards and SWGDOC methodology guidelines. The score-based LR approach implemented in ForensicsFlow is consistent with OSAC-endorsed best practices for quantitative handwriting evidence evaluation.
Open Source Tools & Implementations

Software Foundations

The ForensicsFlow platform builds upon and extends open-source forensic software tools developed by academic and research institutions. The following repositories represent the algorithmic foundations implemented in our modules — adapted, extended, and validated for production forensic workflows.

R Package · GPL-3.0
CSAFE-ISU · Iowa State University

Forensic Handwriting Analysis with Random Forests & Score-Based Likelihood Ratios

An R package developed by the Center for Statistics and Applications in Forensic Evidence (CSAFE) at Iowa State University. Implements the complete forensic handwriting comparison pipeline: decompose handwriting samples into graph-based graphical units, sort them into 40 exemplar clusters, compare cluster-frequency vectors using a Random Forest ensemble, and output a calibrated Score-Based Likelihood Ratio (SLR) under two competing forensic hypotheses — Hp (same writer) and Hd (different writers).

Random Forest Ensemble Score-Based LR Graph Decomposition CSAFE Database Kernel Density Calibration
View on GitHub →
How ForensicsFlow Uses This
Handwriting Static Analysis module — the Random Forest + graph cluster approach is the core engine for authorship comparison, producing the SLR output used in all court-admissible reports.
SLR calibration methodology — kernel density estimation of same-writer vs different-writer score distributions is extended with bootstrap resampling for more stable calibration across diverse document types.
CSAFE Handwriting Database integration — the pre-trained Random Forest models bundled in the package form the baseline for ForensicsFlow's static analysis classifier, extended with additional multi-script training data.
Language R
License GPL-3.0
Commits 228
ForensicsFlow Module 01 — Static Analysis
Web App · AGPL-3.0
ASP.NET Core · TypeScript · C#

Handwriting Analysis Tool — Concurrent Multi-Style Similarity Comparison

A web-based handwriting analysis tool built in ASP.NET Core and TypeScript that implements concurrent multi-style scribal hand analysis. Multiple predefined handwriting styles can be analysed simultaneously against a questioned or unknown sample, producing a quantitative similarity score for each reference style and sorting them in ranked order. Designed for manuscript research, palaeography, and historical document attribution — identifying scribes, dating manuscripts, and tracking gradual style evolution over time.

Multi-Style Concurrent Similarity Ranking Scribal Hand ID Style Drift Detection Manuscript Dating
View on GitHub →
How ForensicsFlow Uses This
Handwriting Dynamic Analysis module — the concurrent multi-style comparison architecture underpins the ranked similarity output, enabling full-population scribal hand screening in a single pass rather than sequential pairwise comparison.
Manuscript & historical document workflows — the tool's use cases (scribal hand identification, school of writing attribution, style drift dating) are directly reflected in ForensicsFlow's Dynamic Analysis use case documentation.
Similarity scoring paradigm — the relative ranked comparison output (score per style, sorted descending) is adopted directly as the primary result format for the Dynamic Analysis module's population screening mode.
Languages TypeScript (58%) · C# (24%)
Framework ASP.NET Core
License AGPL-3.0
ForensicsFlow Module 02 — Dynamic Analysis
Python Package · MIT
BDALab · PyPI Available

Handwriting Features — Kinematic, Dynamic, Spatial & Temporal Online Handwriting Analysis

A PyPI-installable Python library developed by the BDALab for the computation of conventionally used handwriting features from online handwriting data. Provides a comprehensive feature extraction pipeline across five categories — kinematic (velocity, acceleration, jerk), dynamic (azimuth, tilt, pressure), spatial (stroke dimensions, intersection counts), temporal (stroke duration, interruptions), and composite (writing tempo, stops, profile change counts). Supports input from SVC, JSON, NumPy, and Pandas. Interfaces with the Featurizer API for scalable feature extraction.

Kinematic Features Pressure Analysis Temporal Features Spatial Features Writing Tempo SVC / JSON / NumPy
How ForensicsFlow Uses This
Dynamic Analysis feature extraction — the kinematic (velocity, acceleration, jerk), dynamic (azimuth, tilt, pressure), and temporal (stroke duration, interruptions) feature categories directly map to the feature vectors computed in ForensicsFlow's Handwriting Dynamic Analysis module.
Signature Dynamic Analysis — writing tempo, stops, and the number of changes in velocity, pressure, and azimuth profiles are key discriminative features for detecting hesitation patterns and non-habitual signing execution in the Dynamic Signature module.
Input format compatibility — the SVC and JSON input support enables seamless ingestion of tablet-captured online handwriting data from Wacom, Huion, and XP-Pen devices into ForensicsFlow's dynamic processing pipeline.
Feature Categories
3
Kinematic
3
Dynamic
22
Spatial
7+14
Temporal · Composite
ForensicsFlow Modules 02 · 05

Explore Our Research in Depth

Access the full publication list, request technical documentation, or reach out to discuss research collaboration.