Multimodal Biometric Recognition System

A PyTorch-based multimodal biometric recognition system that uses fingerprints and iris images for person identification.

System Design

Quick Start

# Install
poetry install

# Train
poetry run train

# Serve API
poetry run serve

# Run with Docker Compose (MLflow + Airflow)
docker-compose up -d

Model Architecture

Component	Architecture	Output Dim
Fingerprint Branch	MobileNetV2 (pretrained, frozen)	1280
Iris Branch	2x Conv2D + MaxPool + GlobalAvgPool	32 (shared)
Fusion Module	Linear(1344→128) + ReLU + Dropout	128
Classifier	Linear(128→num_classes)	num_classes

Project Structure

src/biometric_recognition/
├── api/           # FastAPI server (serve.py, schema.py)
├── data/          # BiometricDataset class
├── models/        # Model architecture (branches.py, multimodal_model.py)
├── pipeline/      # Training stages (data_prep, train, evaluate, upload)
├── utils/         # Utilities (aws, mlflow, metrics, training, etc.)
└── train.py       # Main training script

airflow/           # Airflow DAG for pipeline orchestration
k8s/               # Kubernetes deployment manifests
terraform/         # AWS infrastructure (ECR, S3)
configs/           # Hydra configuration

Training

# Local training
poetry run train

# With custom parameters
poetry run train training.epochs=20 data.batch_size=16

# Via Airflow
docker-compose exec airflow-webserver airflow dags trigger biometric_training_pipeline

Pipeline Stages

Stage	Description	Outputs
data_prep	Create stratified splits	data_splits.json
train	Train with validation	best_model.pth
evaluate	Test set evaluation	evaluation_results.json, plots
upload	Upload artifacts to S3	S3 URIs

API Server

poetry run serve

Endpoint	Method	Description
/health	GET	Health check
/predict	POST	Predict from images
/model/info	GET	Model metadata
/docs	GET	Swagger UI

Example Request

import requests

files = {
    "fingerprint": open("fingerprint.bmp", "rb"),
    "left_iris": open("left_iris.bmp", "rb"),
    "right_iris": open("right_iris.bmp", "rb"),
}
response = requests.post("http://localhost:8000/predict", files=files)
print(response.json())
# {"predicted_class": 12, "confidence": 0.95, "top_k_predictions": [...]}

Deployment

Kubernetes

kubectl create secret generic aws-credentials \
  --from-literal=AWS_ACCESS_KEY_ID=<key> \
  --from-literal=AWS_SECRET_ACCESS_KEY=<secret>

kubectl apply -f k8s/

Terraform (AWS Infrastructure)

cd terraform && terraform init && terraform apply

Creates: ECR repository, S3 bucket (versioning + encryption enabled)

Configuration

Key settings in configs/config.yaml:

data:
  path: "s3://biometric-recognition-artifacts/biometric_data/"
  batch_size: 8
  num_workers: 8

model:
  backbone_name: "mobilenetv2_100"
  dropout: 0.5

training:
  epochs: 50
  learning_rate: 0.0001
  device: "auto"  # auto, cpu, cuda, mps

Environment Variables

Variable	Description	Default
MODEL_PATH	Model checkpoint path	checkpoints/best_model.pth
NUM_CLASSES	Number of classes	45
MLFLOW_TRACKING_URI	MLflow server URL	http://localhost:5000

Development

# Code quality
poetry run black src/
poetry run isort src/
poetry run flake8 src/
poetry run mypy src/

# Tests
poetry run pytest

Service URLs

Service	URL	Credentials
Airflow UI	http://localhost:8080	admin / admin
MLflow UI	http://localhost:5000	-
API	http://localhost:8000	-

License

MIT License