T0KII/signatures-dataset

Sig-DS: Signature Dataset

This dataset contains a collection of genuine and forged handwritten signatures designed for training image similarity and forgery detection algorithms.

Dataset Details

• —Developed by: Toni Ihab
• —Type: Image Dataset
• —Language(s): English
• —License: MIT

Intended Use

• —Primary use cases: Training deep learning models for handwriting authentication, biometric similarity scoring, and forgery detection. It is highly suitable for creating positive and negative pairs to train Siamese Neural Networks.
• —Out-of-scope uses: This dataset does not contain temporal or pressure data (such as stroke speed or depth) and should not be used to train models requiring dynamic hardware capture inputs.

How to Get Started

To download the images directly into your Python environment or Jupyter Notebook, you can use the huggingface_hub library:

# Install the library if you haven't already
# !pip install huggingface_hub

from huggingface_hub import snapshot_download

# Download the dataset repository to your local machine or Colab environment
dataset_path = snapshot_download(repo_id="T0KII/signatures-dataset", repo_type="dataset")

print(f"Dataset downloaded to: {dataset_path}")

Dataset Structure

The dataset is structured to provide an equal balance of authentic and forged samples across multiple writers. It consists of two primary categories:

• —Genuine Signatures (`full_org`): Contains 1320 genuine signatures, providing 24 genuine samples for each of the 55 distinct writers[cite: 2].
• —Forgery Signatures (`full_forg`): Contains 1320 forgery signatures, providing 24 forgeries for each of the 55 writers[cite: 2].

Performance & Limitations

• —Limitations: The dataset focuses entirely on static, offline signature images. It is limited to the stylistic variations of the 55 included writers. Models trained solely on this data may require further fine-tuning or domain-adaptive pretraining to generalize to highly erratic signatures or different writing instruments not represented in the collection.