Mels22/Signature-Detection-Verification

Signature Detection and Verification

This repository provides two models as part of a full signature authentication pipeline:

1. Detection Model (YOLOv11s-based): A lightweight object detection model fine-tuned to detect signature regions in scanned documents. The model takes full document images as input and returns bounding boxes of all detected signatures.

2. Verification Model (Siamese CNN): A Siamese network trained to determine whether two given signatures (a query signature cropped from a document and a reference signature) belong to the same person. It outputs a binary prediction: 0 = genuine, 1 = forged.

These models are designed to work together in a real-world flow: → detect signature regions from documents → crop a specific query signature → compare it to a reference sample using the verification model.

Developed by: @Mels22 and @JoeCao

Model Architecture

Detection Model

• Base architecture: YOLO11s
• Trained on: SignverOD: A Dataset Signature Object Detection
• Fine-tuned on: Mels22/SigDetectVerifyFlow (1 class: 'signature')

Verification Model

• Architecture: Convolution Siamese Network
• Loss function: Contrastive Loss
• Trained on: Mels22/SigDetectVerifyFlow

For more details on the training process and architecture, please visit our repo Github at Signature-Detect-To-Verify.

Training hyperparameters and Results

Detection Model

The following hyperparameters were used during training:

• Epochs: 50
• Optimizer: AdamW
• Batch size: 16
• Image size: 768

Results:

• Precision: 0.9025
• Recall: 0.7934
• mAP@50: 0.8222
• mAP@50-95: 0.4771

Verification Model

The following hyperparameters were used during training:

• Epochs: 15
• Optimizer: AdamW
• Batch size: 32
• Image size: 105x105
• Learning rate: 1e-4
• Embedding size: 256

Results:

• Accuracy: 100%

Testing the Full Pipeline

We evaluated the end-to-end performance by integrating both the detection and verification models into a complete flow.

• Detection metrics remain consistent with individual evaluation.
• End-to-end accuracy (detection + verification): 0.5743