signature-similarity

Model Card: Siamese Signature Verification Network

Model Details

• Developed by: Toni Ihab
• Model Type: Siamese-style Convolutional Neural Network (CNN) configured for metric learning and signature verification.
• Framework: TensorFlow / Keras 3.x
• Repository: Hugging Face Model - signature-similarity
• Dataset: Hugging Face Dataset - signatures-dataset
• Web App: Streamlit Live App

Intended Use

This model is designed to verify the authenticity of handwriting signatures by comparing a query signature against a reference signature. It calculates a dissimilarity score using Euclidean distance metrics over extracted feature embeddings to assist in forgery detection.

Training Data

The model is trained on the Sig-DS dataset, which consists of:

• 1,320 Genuine Signatures: 24 genuine samples collected across 55 distinct writers.
• 1,320 Forgery Signatures: 24 forgery samples across the same 55 writers.
• Pair Generation: The dataset was dynamically organized into positive pairs (same writer) and negative pairs (different writers) split using a 25% test validation ratio.

Preprocessing Pipeline

• Conversion of input signatures to single-channel grayscale.
• Spatial normalization and resizing to a standardized dimension of $128 \times 128$ pixels.
• Intensity normalization scaling pixel values to the $[0, 1]$ range.
• Integrated digital canvas preprocessing (auto-cropping bounding boxes, aspect-ratio padding, and Gaussian anti-aliasing to bridge domain shifts between scanned physical ink and digital user inputs).

Evaluation & Performance

• Loss Function: Contrastive loss optimizing feature space distances.
• Distance Metric: Euclidean distance via a custom Lambda layer. Identical signatures approach a distance of 0.0, while distinct writers or forgeries scale higher towards the contrastive margin.
• Live Deployment: Integrated into a fully interactive web application available via Streamlit Cloud.