Mels22/SigDetectVerifyFlow

Signature Detection and Verification Dataset

A comprehensive dataset designed for building and evaluating end-to-end signature analysis pipelines, including signature detection in document images and signature verification using genuine/forged pair classification.

Developed by: @Mels22 and @JoeCao

Pipeline Overview

This dataset supports a complete signature detection and verification pipeline. The process involves identifying the signature in a document and comparing it with a reference to determine if it is genuine or forged.

Figure 1: Detection and Verification Pipeline.

• The Detection Model locates the signature in the document.
• The cropped signature (to_verify_signature) is passed along with a sample signature (sample_signature) to the Verification Model.
• The model then classifies the signature as either Genuine or Forged.

Dataset Summary

Split	Samples
Train	23,206
Test	6,195
Total	29,401

This dataset supports two key tasks:

• Detection: Identifying the bounding boxes of signatures in scanned document images.
• Verification: Comparing a signature within the document to a reference (sample) signature to determine whether it's genuine (label = 0) or forged (label = 1).

Features

Each sample in the dataset contains the following fields:

• document (Image): The full document image that contains one or more handwritten signatures.
• bbox (List of Bounding Boxes): The coordinates of the signature(s) detected in the document. Format: [x_min, y_min, x_max, y_max].
• to_verify_signature (Image): A cropped signature from the document image that needs to be verified.
• sample_signature (Image): A standard reference signature used for comparison.
• label (int): Indicates if the to_verify_signature is genuine (0) or forged (1) when compared to the sample_signature.

Data Sources & Construction

This dataset is constructed by combining and modifying two publicly available datasets:

• Signature Images were sourced from the Kaggle Signature Verification Dataset, which provides genuine and forged signatures from multiple individuals for verification tasks.

• Document Images with Signature Bounding Boxes were taken from the Signature Detection Dataset by NanoNets, which contains scanned documents with annotated signature regions.

How This Dataset Was Created

To create a seamless, unified pipeline dataset for detection + verification, the following modifications were made:

• Synthetic Placement: Signature images were programmatically inserted into real documents at their correct signing regions (e.g., bottom of the page or designated signature lines).
• Blending with Background: Signatures were rendered with varying opacities, filters, and transformations to match the document background, mimicking real-world signature scans.
• Labeling and BBoxes: The new locations of the inserted signatures were used to generate accurate bounding boxes for detection tasks.
• Pairing for Verification: Each inserted signature (to_verify_signature) was paired with a reference (sample_signature) and assigned a label: 0 for genuine or 1 for forged.

This process enables researchers to train and evaluate models for both signature localization and signature verification in a realistic, document-centric setting.

Sample Code

from datasets import load_dataset
data = load_dataset("Mels22/SigDetectVerifyFlow")

for i, example in enumerate(data['train']):
    example['document'].show()
    example['to_verify_signature'].show()
    example['sample_signature'].show()
    print(f"Bbox: {example['bbox']}")
    print(f"Label: {example['label']}")
    break