How goyashek/signature-forgery-verification works
This repository is an offline signature-forgery verification demo and model-development project. The main user flow is to upload 3–5 genuine reference signatures and a questioned signature in the Gradio app; shared inference deduplicates and preprocesses them, embeds them with the saved model, then returns a calibrated genuine, forged, or inconclusive verdict. Separate supporting branches prepare combined datasets, detect known leakage patterns, and document model experiments. The graph follows the sampled runtime and README; notebook internals were not inspected.
Main entry points: inference.py Image inference · app.py Gradio app · sign_data Signature datasets · analyze_band.py Threshold analysis
Components 14
Demo workflow
- Gradio app
app.py - NFI samples
sign_data_nfi
Verification runtime
- Image inference
inference.py - Image preprocessing
inference.py - Embedding model
models/siamese_bh_embedding.keras - Writer decision
inference.py - Model metadata
models/siamese_bh_meta.json
Dataset preparation
- Signature datasets
sign_data - Combined manifest
sign_data_combined/manifest.csv - Dataset builder
build_combined_dataset.py - Leak detector
check_data_leak.py
Model research
- Model experiments
notebooks - Threshold analysis
analyze_band.py
Other components
- Verifier
Connections 13
- Verifier → Gradio app: uploads images
README.md - Gradio app → Image inference: calls verification
app.py - Gradio app → NFI samples: loads samples
app.py - Image inference → Image preprocessing: preprocesses images
inference.py - Image inference → Embedding model: predicts embeddings
inference.py - Image inference → Writer decision: scores embeddings
inference.py - Image inference → Model metadata: loads settings
inference.py - Signature datasets → Dataset builder: provides images
build_combined_dataset.py - Dataset builder → Combined manifest: writes manifest
build_combined_dataset.py - Signature datasets → Leak detector: checks leakage
check_data_leak.py - Model experiments → Image inference: demonstrates shared path
README.md - Threshold analysis → Embedding model: loads model
analyze_band.py - Threshold analysis → Model metadata: uses calibration settings