Fake Currency Detection Using Deep Learning

Project Overview

This project focuses on detecting fake and real Indian currency notes using deep learning. The goal was to build an end-to-end system that can accurately classify currency images as real or fake, utilizing both image data and extracted note features.

The project consists of four major notebooks:

1. 01_Base_Model_Training.ipynb — Builds and trains the base model using transfer learning.
2. 02_Two_Branch_Model_Training.ipynb — Enhances the model using a two-branch architecture combining image and feature embeddings.
3. 03_Feature_Enhancements.ipynb — Fine-tunes the base model for improved performance.
4. 04_Attention_Enhanced_Model.ipynb — Implements attention mechanisms (CBAM) for further performance improvement.

---

Notebook 1: Base Model Training (01_Base_Model_Training.ipynb)

Objective

To train a baseline convolutional neural network using MobileNetV2 as a pretrained feature extractor for classifying real vs fake currency notes.

Key Steps

• Loaded the dataset of real and fake currency images using OpenCV.
• Preprocessed images (resizing to 224×224, normalization, RGB conversion).
• Created TensorFlow datasets for efficient loading and batching.
• Used MobileNetV2 (ImageNet pretrained) as a frozen feature extractor.
• Added custom classification layers on top of the base model.
• Trained the model using Adam optimizer with callbacks like EarlyStopping, ModelCheckpoint, and ReduceLROnPlateau.

Model Summary

• Architecture: MobileNetV2 + Dense layers
• Input Shape: (224, 224, 3)
• Loss: Binary Cross-Entropy
• Optimizer: Adam
• Best Validation Accuracy: ~99.06%

Training Output (Excerpt)

Epoch 20/20
187/187 ━━━━━━━━━━━━━━━━━━━━ 10s 44ms/step - accuracy: 0.9991 - loss: 0.0065 - val_accuracy: 0.9906 - val_loss: 0.0322

Model Saved As

• Models/best_model.h5

---

Notebook 2: Two-Branch Model Training (02_Two_Branch_Model_Training.ipynb)

Objective

To improve classification accuracy by integrating both main image features and currency-specific precomputed feature embeddings into a unified two-branch neural network.

Key Steps

• Loaded the pretrained base model (best_model.h5) and froze its layers.
• Built a feature extractor from the last convolutional layer of the base model.
• Precomputed embeddings for all feature images (e.g., watermark, security thread, color patterns) and stored them as .npy files.
• Designed a custom data generator (TwoBranchEmbeddingGenerator) to load both the main image and precomputed embeddings during training.
• Created a Two-Branch Neural Network:
  • Branch 1: Image branch (MobileNetV2 + dense layers)
  • Branch 2: Feature embedding branch (fully connected layers)
  • Combined using concatenation followed by dense layers.
• Trained using similar callbacks as the base model.

Model Summary

• Architecture: Two-Branch CNN (Image + Embedding)
• Feature Extractor: MobileNetV2 (frozen)
• Loss: Binary Cross-Entropy
• Optimizer: Adam
• Best Validation Accuracy: ~98.59%

Training Output (Excerpt)

Epoch 11/15
745/745 ━━━━━━━━━━━━━━━━━━━━ 260s 350ms/step - accuracy: 0.9949 - loss: 0.0158 - val_accuracy: 0.9859 - val_loss: 0.0567

Model Saved As

• Models/best_two_branch_model.keras

---

Notebook 3: Feature Enhancements (03_Feature_Enhancements.ipynb)

Objective

Fine-tune the base MobileNetV2 model to improve accuracy and generalization.

Key Steps

• Loaded the pretrained base model (best_model.h5) and unfroze deeper layers for fine-tuning.
• Compiled the model with a very low learning rate to prevent catastrophic forgetting.
• Used EarlyStopping, ReduceLROnPlateau, and ModelCheckpoint callbacks.
• Trained the fine-tuned model on the original dataset.
• Visualized training accuracy and loss.

Model Summary

• Architecture: MobileNetV2 (fine-tuned) + Dense layers
• Input Shape: (224, 224, 3)
• Loss: Binary Cross-Entropy
• Optimizer: Adam (low learning rate)
• Best Validation Accuracy: ~99.66%

Training Output (Excerpt)

Epoch 10/10
187/187 ─ 9s 49ms/step - accuracy: 0.9984 - loss: 0.0060 - val_accuracy: 0.9832 - val_loss: 0.0504

Model Saved As

• Models/fine_tuned_base_model.keras

---

Notebook 4: Attention-Enhanced Model (04_Attention_Enhanced_Model.ipynb)

Objective

Further improve model performance using attention mechanisms (CBAM) to focus on key regions of currency images.

Key Steps

• Loaded fine-tuned base model and applied CBAM blocks to enhance feature representation.
• Added GlobalAveragePooling and Dense layers after attention.
• Compiled and trained the attention-enhanced model with callbacks.
• Evaluated the model using accuracy, confusion matrix, precision, recall, f1-score, and ROC-AUC.
• Achieved robust classification performance with minimal misclassifications.

Model Summary

• Architecture: MobileNetV2 + CBAM + Dense layers
• Loss: Binary Cross-Entropy
• Optimizer: Adam
• Evaluation Metrics: Accuracy, Precision, Recall, F1-score, ROC-AUC
• Best Validation Accuracy: ~98–99%

Sample Evaluation Output

Confusion Matrix:
 [[489  13]
 [ 18 969]]

Classification Report:
               precision    recall  f1-score   support
        Fake       0.96      0.97      0.97       502
        Real       0.99      0.98      0.98       987

ROC-AUC Score: 0.9986

Model Saved As

• Models/attention_enhanced_model.keras

---

Visualizations (To Be Added)

Below plots will be added in future updates:

• Accuracy and Loss curves for both models
• Confusion Matrix
• Sample Predictions Visualization
• ROC-AUC Curve

---

Repository Structure

├── 01_Base_Model_Training.ipynb
├── 02_Two_Branch_Model_Training.ipynb
├── 03_Feature_Enhancements.ipynb
├── 04_Attention_Enhanced_Model.ipynb
├── Models
│   ├── best_model.h5
│   ├── best_two_branch_model.keras
│   ├── fine_tuned_base_model.keras
│   └── attention_enhanced_model.keras
└── README.md

---

Dependencies

• TensorFlow / Keras
• OpenCV (cv2)
• NumPy
• scikit-learn
• matplotlib
• tqdm