ahmadous/anip-benin-2025-identity-ai

AI system for identity document verification — Face matching, Age estimation, OCR & Fraud detection — ANIP Benin 2025 # ANIP Benin 2025 — Identity AI System 🪪🤖 ## AI-Powered Identity Document Verification ### Finalist — ANIP Benin Challenge 2025 --- ## Three Tasks Implemented ### Task 1 — Face Matching & Verification Siamese Network for facial recognition and identity verification on official documents. **Approach:** - Contrastive loss function - Cosine similarity embeddings - TensorFlow + OpenCV **Stack:** Python | TensorFlow | OpenCV | NumPy --- ### Task 2 — Age Estimation Deep learning regression model for age estimation from facial photos. **Results:** - Best val_MAE: 9.16 years - Training: 30 epochs with ReduceLROnPlateau - Dataset: 200 test images **Stack:** Python | TensorFlow | Transfer Learning --- ### Task 3 — OCR & Document Fraud Detection Five-class fraud classification on identity documents from 4 countries. **Two versions developed:** - v1: EfficientNetB0 baseline - v2 (improved): EfficientNetV2B0 + Cosine Decay + Advanced data augmentation **Results:** - Best val_accuracy: 20.4% (difficult 5-class problem) - 200 test images | 4 countries: arizona_dl, esp, est, rus - CPU-only training **Stack:** Python | TensorFlow | EfficientNetV2 | OCR --- ## Key Learnings This challenge demonstrated the complexity of real-world biometric AI: - Face verification requires robust embedding spaces - Age estimation degrades on small/unbalanced datasets - Document fraud detection with 5 classes and limited data is genuinely hard The value lies in the methodology, not just the metrics. --- ## Repository Structure ``` anip-benin-2025-identity-ai/ ├── │ └── tache1_face_matching.ipynb ├── │ └── tache2_age_estimation.ipynb ├── │ ├── tache3_ocr_fraud_detection.ipynb (v1) │ └── tache3_improved.ipynb (v2) └── README.md ``` ## Competition Context ANIP — Agence Nationale d'Identification des Personnes Benin | November 2025 Final presentation: Palais des Congrès, Cotonou ## Tache 1 — Detail complet (V2 → V9) Le repo ci-dessous contient toutes les iterations de la tac …