Tomato Disease Classification Project

This project implements multiple deep learning models to classify tomato leaf diseases using TensorFlow/Keras. The project compares the performance of a custom CNN architecture against transfer learning approaches using ResNet50V2 and EfficientNetB0.

Setup Instructions

1. Create and Activate Virtual Environment

On Windows:

# Create virtual environment
py -m venv venv
# Activate virtual environment
.\venv\Scripts\activate

On macOS/Linux:

# Create virtual environment
python3 -m venv venv
# Activate virtual environment
source venv/bin/activate

2. Install Dependencies

pip install tensorflow pillow matplotlib numpy gdown pathlib scipy

3. Project Structure

tomato-disease-classification/
├── venv/
├── models/                 # Directory containing saved model files
├── plots/                 # Training visualizations and comparisons
├── results/              # JSON files with training metrics
├── corrupted_images/     # Any corrupted images found during verification
├── main.py              # Main script with model implementations
└── README.md

Running the Project

1. Make sure your virtual environment is activated
2. Run the main script:

python main.py

The script will:

• Download the dataset from Google Drive automatically
• Verify image integrity and handle corrupted images
• Train three different models:
  • Basic CNN
  • ResNet50V2 (transfer learning)
  • EfficientNetB0 (transfer learning)
• Generate comparison plots
• Save the trained models
• Output performance metrics

Output Files

The script generates several files with timestamps:

• plots/model_comparison_[timestamp].png: Visual comparison of model performance
• results/model_results_[timestamp].json: Detailed metrics for each model
• Saved model files in the models/ directory:
  • tomato_disease_basic_cnn_[timestamp].keras
  • tomato_disease_resnet50v2_[timestamp].keras
  • tomato_disease_efficientnetb0_[timestamp].keras

Dataset

The dataset will be automatically downloaded from Google Drive when running the script. The download process includes:

• Automatic download handling using gdown
• Image verification to detect and handle corrupted files
• Directory structure creation
• Dataset extraction

Features

• Image integrity verification
• Automatic dataset download and extraction
• Multiple model architectures:
  • Custom CNN with batch normalization
  • Transfer learning with ResNet50V2
  • Transfer learning with EfficientNetB0
• Data augmentation for training
• Early stopping and model checkpointing
• Comprehensive performance comparison
• Automated results logging

System Requirements

• Python 3.8 or higher
• At least 8GB RAM
• NVIDIA GPU recommended but not required
• Internet connection for downloading the dataset

Model Architecture Details

Basic CNN

• Three convolutional blocks with batch normalization
• Global average pooling
• Dense layers with dropout for regularization

Transfer Learning Models

• Pre-trained ResNet50V2 and EfficientNetB0 base
• Custom top layers for classification
• Frozen pre-trained weights
• Dropout for regularization

Troubleshooting

1. If you encounter memory issues:

   • Reduce batch_size (default is 32)
   • Reduce input shape (default is 128x128)
   • Free up system memory
   • Use a machine with more RAM

2. For download issues:

   • Check your internet connection
   • Verify the Google Drive file ID is still valid
   • Try downloading the dataset manually and place in the correct directory structure

3. For GPU-related issues:

   • Verify TensorFlow can access your GPU
   • Update GPU drivers
   • Check CUDA and cuDNN compatibility with your TensorFlow version

4. For corrupted images:

   • Check the corrupted_images directory for any files that were moved
   • Verify the integrity of the downloaded dataset
   • Try re-downloading the dataset

Performance Monitoring

The training process includes:

• Real-time accuracy and loss monitoring
• Automatic early stopping to prevent overfitting
• Model checkpointing to save the best weights
• Comprehensive performance comparison plots
• Detailed JSON logs of training metrics

Future Improvements

• Implement cross-validation
• Add model ensemble capabilities
• Expand to more disease classes
• Add prediction visualization tools
• Implement gradual unfreezing for transfer learning models