💵 Fraud Currency Detection Using ResNet50

A deep learning-based web application for detecting genuine and counterfeit Bangladeshi currency notes using a fine-tuned ResNet-50 model.

The application currently supports 500 BDT and 1000 BDT banknotes and provides the predicted class along with the model's confidence score through an interactive Streamlit interface.

---

Features

• Detects genuine and counterfeit Bangladeshi banknotes
• Supports:
  • 500 BDT
  • 1000 BDT
• Fine-tuned ResNet-50 classifier
• Displays prediction confidence
• Handles low-confidence predictions
• Clean and responsive Streamlit interface
• GPU support (CUDA) when available

---

Demo

Main Interface

---

Project Structure

Fraud-Currency-Detection/
│
├── app.py
├── best_model9740.pth
├── label_encoder.pkl
├── requirements.txt
├── README.md
│
├── assets/
│   └── app.png
│
└── notebooks/

---

Model

The classifier is built using ResNet-50 from TorchVision.

The original classification head has been replaced with a custom fully connected classifier.

Linear(2048 → 256)
ReLU
Dropout(0.5)
Linear(256 → 4)

The four output classes are:

• Genuine 500 BDT
• Counterfeit 500 BDT
• Genuine 1000 BDT
• Counterfeit 1000 BDT

---

Technologies Used

• Python
• PyTorch
• TorchVision
• Streamlit
• Pillow
• scikit-learn
• Joblib

---

Image Preprocessing

Each uploaded image undergoes the following preprocessing steps before inference:

• Resize to 224 × 224
• Convert to Tensor
• Normalize using ImageNet mean and standard deviation

Resize(224×224)
ToTensor()
Normalize(mean=[0.485,0.456,0.406],
          std=[0.229,0.224,0.225])

---

Installation

Clone the repository

git clone https://github.com/mahady13/Fraud-Currency-Detection.git

Move into the project directory

cd Fraud-Currency-Detection

Install dependencies

pip install -r requirements.txt

Run the application

streamlit run app.py

---

Usage

1. Launch the Streamlit application.
2. Upload a 500 BDT or 1000 BDT banknote image.
3. Click Predict.
4. View:
   • Predicted class
   • Confidence score

If the confidence score is very low, the application warns that the uploaded image may not belong to the supported classes.

---

Dataset

This project was trained using the Bangladeshi Counterfeit Currency Image Dataset.

Citation

Afif, A., & Rashid, M. (2026). Bangladeshi Counterfeit Currency Image Dataset. Mendeley Data, V1.

DOI:

https://doi.org/10.17632/gzzz5nrvbn.1

---

Limitations

This model has been trained only on:

• 500 BDT
• 1000 BDT

Uploading other denominations (100, 200, 50 BDT, etc.) or unrelated images may produce unreliable predictions.

---

Future Improvements

• Support additional Bangladeshi banknote denominations
• Deploy the model on the cloud
• Real-time webcam detection
• Grad-CAM visualization for model interpretability
• Mobile-friendly interface
• Model optimization for faster inference

---

Author

Mohiuddin Mahady

B.Sc. in Computer Science & Engineering

Mymensingh Engineering College
(Affiliated with the University of Dhaka)

GitHub: https://github.com/mahady13

LinkedIn: https://www.linkedin.com/in/mohiuddin-mahady/

---

License

This project is released under the MIT License.