YOLOS (base-sized) Model For Handwritten Signature Detection#
YOLOS model finetuned to detect handwritten signatures in document images using [tech4humans/signature-detection ](https://huggingface.co/datasets/tech4humans/signature-detection ) dataset.
Original YOLOS was introduced in the paper You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection by Fang et al. and first released in this repository .
Model description#
YOLOS is a Vision Transformer (ViT) trained using the DETR loss. Despite its simplicity, a base-sized YOLOS model is able to achieve 42 AP on COCO validation 2017 (similar to DETR and more complex frameworks such as Faster R-CNN).
Uses#
This model is designed for detecting handwritten signatures in scanned documents, contracts, or forms.
You can try it instantly in your browser here: HF Space
Direct Use#
Here is how to use this model:
from datasets import load_dataset
from transformers import pipeline
# Load the tech4humans signature dataset
dataset = load_dataset("samuellimabraz/signature-detection")
# Load the finetuned model
yolos = pipeline(
task="object-detection",
model="mdefrance/yolos-base-signature-detection",
device_map="auto",
)
# Inference on test sample
prediction = yolos(dataset["test"][0].get("image"))Currently, both the image processor and model support PyTorch.
Out-of-Scope Use#
Fraudulent Use: This model must not be used for forging signatures or any illegal activity. It’s meant for legitimate signature detection in documents.Other Objects: Not suitable for detecting non-signature elements in documents.Critical Decisions: Should not be solely relied on for high-stakes decisions (e.g., legal or financial) without human validation.Bias, Risks, and Limitations#
Bias: May not generalize well if training data lacks diversity in signature styles or cultural context.Risks: False positives/negatives can occur, impacting document validation.Limitations: Performance may degrade on poor-quality images or in challenging visual conditions (e.g., noise, lighting).Recommendations#
Improve Training Data: Fine-tune with diverse and representative samples to reduce bias.Human Oversight: Always include a human review step for critical use cases.Image Quality: Use clean, high-resolution images; apply preprocessing if needed.Ethical Use: Follow legal and ethical standards, ensuring privacy and responsible deployment.Training Details#
Training Data#
| Dataset on HF
The training utilized a dataset built from two public datasets: Tobacco800 and signatures-xc8up , unified and processed in Roboflow . The processed dataset was created by Samuel Lima Braz , and all credit for the dataset preparation goes to him.
Dataset Summary:
Training: 1,980 images (70%) Validation: 420 images (15%) Testing: 419 images (15%) Format: COCO JSON Resolution: 640x640 pixels Training Procedure#
See mdefrance/signature-detection for details on training procedure.
Metrics#
Performances computed on the testing set:
Inference times are computed on a laptop with following specs:
CPU: Intel Core i7-9750H GPU: NVIDIA GeForce GTX 1650 License Comparison#
GNU Affero General Public License v3.0 (AGPL-3.0)#
AGPL-3.0 is a strong copyleft license designed to keep software and its modifications open-source, especially for web apps and network services.
Strong Copyleft : Modified versions must also be AGPL-licensed.Network Use : Users must get the source code, even if they only interact with the software over a network.Commercial Use : Allowed, but any changes must be shared under AGPL-3.0.Patent Protection : Includes safeguards against patent and trademark claims.Apache License 2.0#
Apache 2.0 is a permissive license that offers flexibility for both open-source and proprietary use.
Permissive : Modifications and derivatives don’t need to be open-source.Commercial Use : Fully allowed with no requirement to share changes.Patent Protection : Includes strong patent clauses.Compatibility : Easy to combine with other licenses and projects.Summary: Why Apache 2.0 Offers More Flexibility#
While AGPL-3.0 ensures openness, Apache 2.0 is better suited for businesses and closed-source use:
No obligation to disclose modified code. Easier to integrate into proprietary systems. More flexible for commercial applications. For full license texts, see:
Citation#
This model is a finetuned version of the YOLOS model introduced in the following paper. If you use this model, please cite the original work:
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#
Acknowledgements#
This model is finetuned for handwritten signature detection using the [tech4humans/signature-detection ](https://huggingface.co/datasets/tech4humans/signature-detection ) Dataset. The finetuning process and additional resources can be found in the GitHub Repository mdefrance/signature-detection .
Author # | Mario DEFRANCE Mario DEFRANCE Data Scientist / AI Engineer
|
Responsibilities in this Project#
🔬 Model development and training
⚙️ Performance evaluation
📝 Technical documentation and model card
