YOLOS for Handwritten Signature Detection

This repository hosts the code used to finetune YOLOS models to detect handwritten signatures in document images. The model is trained on the tech4humans/signature-detection dataset, built from real-world signature data across diverse document types.

• Model: mdefrance/yolos-base-signature-detection
• Based on: hustvl/yolos-base
• Dataset: tech4humans/signature-detection
• License: Apache 2.0

---

Quick Usage

from datasets import load_dataset
from transformers import pipeline

# Load test image from dataset
dataset = load_dataset("samuellimabraz/signature-detection")
image = dataset["test"][0]["image"]

# Load the finetuned pipeline
yolos = pipeline(
    task="object-detection",
    model="mdefrance/yolos-base-signature-detection",
    device_map="auto"
)

# Run inference
prediction = yolos(image)

Performances

Metric	yolos-base-signature-detection	yolos-small-signature-detection	yolos-tiny-signature-detection
Inference Time - CPU (s)	2.250	0.787	0.262
Inference Time - GPU (s)	1.464	0.023	0.014
Parameters	127.73M	30.65M	6.47M
mAP50	0.887	0.859	0.856
mAP50-95	0.495	0.421	0.395

Inference times are computed on a laptop with following specs:

• CPU: Intel Core i7-9750H
• GPU: NVIDIA GeForce GTX 1650

Setup & Usage (Local)

Requirements

• Python 3.12.1 (not tested for other versions)
• Poetry 2.1.3 for dependency management

Installation

1. Clone the repo:

git clone https://github.com/mdefrance/signature-detection.git
cd signature-detection

2. Install dependencies with Poetry:

pip install poetry==2.1.3
poetry install

Fine-Tuning Notebook

The notebook used to fine-tune YOLOS on the signature detection dataset is available here. It includes:

• Dataset loading & preprocessing
• YOLOS training configuration
• Training with Pytorch Lightning Trainer
• Evaluation & visualization of predictions

Training Data

• Training: 1,980 images (70%)
• Validation: 420 images (15%)
• Testing: 419 images (15%)
• Format: COCO JSON
• Resolution: 640x640 pixels
• Sources: Tobacco800 + signatures-xc8up
• Preprocessed by Samuel Lima Braz via Roboflow

Out-of-Scope Use

• Forgery or Fraudulent Use: This model is for detection only, not signature generation or spoofing.
• Non-Signature Detection: Not suitable for detecting other objects or text.
• High-Stakes Automation: Use human verification in legal or financial applications.

Limitations & Ethical Use

• Bias Risk: Limited performance on unseen signature styles or document types.
• False Positives/Negatives: Manual review is recommended in production.
• Image Quality: Model struggles with noisy, low-res, or poorly lit inputs.
• Ethics: Respect privacy and regulatory compliance when deploying.

Citation

If you use this model, consider citing the original YOLOS paper:

BibTeX:

@article{DBLP:journals/corr/abs-2106-00666,
  author    = {Yuxin Fang and
               Bencheng Liao and
               Xinggang Wang and
               Jiemin Fang and
               Jiyang Qi and
               Rui Wu and
               Jianwei Niu and
               Wenyu Liu},
  title     = {You Only Look at One Sequence: Rethinking Transformer in Vision through
               Object Detection},
  journal   = {CoRR},
  volume    = {abs/2106.00666},
  year      = {2021},
  url       = {https://arxiv.org/abs/2106.00666},
  eprinttype = {arXiv},
  eprint    = {2106.00666},
  timestamp = {Fri, 29 Apr 2022 19:49:16 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2106-00666.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Additional Resources

• Blog post of comparison of Signature Detection Models: Hugging Face Blog
• Blog post associated Finetuning Notebook: Google Colab Notebook
• Finetuning of YOLOS Notebook Example: Google Colab Notebook

Author

Mario DEFRANCE Data Scientist / AI Engineer

Responsibilities in this Project 🔬 Model development and training ⚙️ Performance evaluation 📝 Technical documentation and model card