Leaffliction

The project is an introduction to Deep Learning/Computer Vision

This project implements a deep learning solution using ResNet18 to classify plant diseases from leaf images, with comprehensive data preprocessing and augmentation pipelines.

Project Structure

./
├── Distribution.py          # Dataset distribution analysis
├── Augmentation.py          # Image augmentation implementation
├── Transformation.py        # PlantCV-based image transformations
├── PlantDiseaseDataset.py   # Custom PyTorch dataset
├── train.py                 # Model training script
├── predict.py               # Disease prediction script
└── Leaves.zip               # Leave image dataset information

Leaves Directory Structure

dataset/
├── Plant_Type1/
│   ├── Disease1/
│   │   ├── original_images/
│   │   └── Augmented/
│   │   └── Transformed/
│   └── Disease2/
│       ├── original_images/
│       └── Augmented/
│   │   └── Transformed/
└── Plant_Type2/
    └── ...

Components

1. Data Analysis

Our Distribution.py script analyzes and visualizes the dataset structure:

python Distribution.py ./path/to/dataset

Features:

• Pie chart visualization of disease distribution
• Bar chart showing image count per disease
• Handles both original and augmented images
• Excludes augmented directories from double-counting

2. Data Augmentation

The Augmentation.py script implements various augmentation techniques using torchvision:

python Augmentation.py ./path/to/image.jpg
# or
python Augmentation.py ./path/to/directory

Implemented transformations:

• Gaussian Blur (kernel_size=(3, 7), sigma=(1.5, 6))
• Random Rotation (-80° to 80°)
• Color Jitter (brightness, contrast, saturation)
• Random Resized Crop (200x200)
• Random Affine Shear (45°)
• Random Horizontal/Vertical Flip

3. Image Transformation

Transformation.py uses PlantCV for advanced image processing:

python Transformation.py ./path/to/image.jpg
# or
python Transformation.py -src source_dir -dst destination_dir

Pipeline includes:

• LAB colorspace conversion
• Binary thresholding
• Gaussian blur
• Mask application
• ROI extraction
• Shape analysis
• X/Y axis pseudolandmarks

4. Custom Dataset

PlantDiseaseDataset.py implements a PyTorch Dataset class with:

• Automatic data balancing
• Support for multiple image formats
• Caching mechanism for faster loading
• Minimum image count validation (500 images)
• Even distribution across classes

5. Model Architecture and Training

We use a modified ResNet18 architecture with transfer learning:

model = models.resnet18(weights=ResNet18_Weights.DEFAULT)
num_features = model.fc.in_features
model.fc = torch.nn.Linear(num_features, num_classes)

Training pipeline features:

• 80/20 train-validation split
• Batch size of 32
• Adam optimizer with learning rate 0.0001 and weight decay 1e-5
• CrossEntropyLoss criterion
• Training progress monitoring with loss and accuracy metrics
• Validation accuracy evaluation after each epoch

Image preprocessing pipeline:

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.GaussianBlur(kernel_size=(3, 3), sigma=(0.1, 2.0)),
    transforms.RandomGrayscale(0.1),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
])

To train the model:

python train.py --data_train /path/to/dataset --total_images 1000

The trained model is saved as 'plant_disease_model.pth' with both the model state dictionary and class labels.

6. Prediction

The predict.py script provides disease classification with visualization:

python predict.py path/to/image1.jpg path/to/image2.jpg ...

Features:

• Side-by-side visualization of original and transformed images
• Clear prediction display
• Batch prediction support
• CPU/GPU compatibility