Apppublic

nilotpaldhar2004/fingerprint-blood-group

sourceHugging Faceupdated 13d agoView on Hugging Face

0likes

App README

🩸 FingerPrint2BloodGroup

Research Demo — Predicting blood group from fingerprint images using Transfer Learning (ResNet-50).

![HuggingFace Space](https://huggingface.co/spaces/nilotpaldhar2004/fingerprint-blood-group) ![Python](https://www.python.org/) ![FastAPI](https://fastapi.tiangolo.com/) ![PyTorch](https://pytorch.org/)

⚠️ Disclaimer

This project is a research and portfolio demonstration only. Current validation accuracy is ~87.22% (test: 87.00%) — while this is strong performance, this model is not suitable for clinical or diagnostic use. Do not make any medical decisions based on its output.

🧠 What It Does

Upload a fingerprint image → the model predicts the corresponding blood group along with:

—Predicted class with confidence score
—Class probability bar chart (all 8 blood groups)
—Grad-CAM heatmap showing which regions of the fingerprint the model focused on

🏗️ Model Architecture

Component	Detail
Backbone	ResNet-50 (ImageNet pretrained)
Backbone	Frozen (feature extractor only)
FC Head	2048 → 1024 → 512 → 128 → 8
Activations	ReLU + Dropout (0.3, 0.2)
Loss	CrossEntropyLoss
Optimizer	Adam (lr=0.0001)
Input size	448 × 448 × 3
FC Head	2048 → 1024 → BN → 512 → 128 → 64 → 8
Dropout	0.4, 0.2, 0.1

📊 Training Results

Epoch	Val Accuracy
1	34.67%
10	67.11%
22	84.00%
31	85.78%
74	86.11%
85	87.22% ⭐

Training set: 4200 samples | Val: 900 | Test: 900 | Batch: 32 | Early stop: epoch 86/100

🗂️ Project Structure

fingerprint-blood-group/
│
├── main.py                        # FastAPI backend + Grad-CAM logic
├── static/
│   └── index.html                 # Frontend UI
├── blood_group_resnet50_best.pth  # Trained model weights
├── blood_group_classes.npy        # Label encoder classes
├── requirements.txt
├── Dockerfile
├── .gitignore
└── .github/
    └── workflows/
        └── deploy_to_hf.yml       # GitHub Actions → HF Space auto-deploy

🚀 Local Setup

bash

# Clone
git clone https://github.com/YOUR_GITHUB_USERNAME/fingerprint-blood-group.git
cd fingerprint-blood-group

# Install dependencies
pip install -r requirements.txt

# Place model files in root directory:
#   blood_group_resnet50_best.pth
#   blood_group_classes.npy

# Run
uvicorn main:app --reload --port 8000

# Open browser
# http://localhost:8000

🐳 Docker

bash

docker build -t fingerprint-blood-group .
docker run -p 7860:7860 fingerprint-blood-group
# Open: http://localhost:7860

☁️ Deploy to HuggingFace Space

1.Create a new Space on HuggingFace: fingerprint-blood-group with Docker SDK
2.Add these GitHub Secrets to your repository:
3.HF_TOKEN → your HuggingFace write token
4.HF_USERNAME → your HuggingFace username
5.Push to main branch — GitHub Actions handles the rest automatically

Important: Model weight files (.pth, .npy) are in .gitignore. Upload them manually to your HF Space via the web UI or use Git LFS.

🛠️ Tech Stack

—Backend: FastAPI, PyTorch, torchvision, OpenCV
—Frontend: Vanilla HTML/CSS/JS (no framework)
—Explainability: Grad-CAM on layer4[-1] of ResNet-50
—Deployment: Docker · HuggingFace Spaces · GitHub Actions

👤 Author

Nilotpal — CS | AI/ML HuggingFace · GitHub

📄 License

MIT License — free to use for educational and research purposes.