TergeuPro 11 / 25
Signature examination
Compares a reference and a questioned signature and drafts an expert-style conclusion.
Problem
Handwriting examination is slow and subjective; investigators need a fast first opinion.
What it does
- Signature verification with similarity score and confidence
- Stroke analysis: pressure, tremor, stops
- Forgery-method detection and writer profile (confidence, speed)
- Generates a conclusion text with scientific argumentation; keeps a history
How it works
- Two images → Siamese CNN (ResNet-50 backbone, 512-d embeddings) on a GPU server
- Stroke and forgery analyzers run in the same pipeline
- Results → conclusion generator → stored in history
- R&D: distillation to a MobileNetV3-Small student (128-d) targeting <10 ms on edge
Stack & models
PyTorchSiamese CNNResNet-50MobileNetV3 (distillation)Yii2 / PHP
Numbers
- Distillation goal: <10 ms on edge with 95% accuracy retention (target, not a result)
Figures come from project docs/README; no other benchmarks are claimed.