This project utilizes advanced deep learning techniques to identify diseases in apple trees from images of their leaves. The dataset comprises annotated photos categorized into four classes: "healthy," "multiple_diseases," "rust," and "scab." The goal is to develop a robust model for early disease detection in apple orchards.
- Dataset: Annotated photos organized into training and test sets. Training labels are extracted from a CSV file with an additional 'label' column for further organization.
- Preprocessing:
- Standardized image resolution to 256x256 pixels.
- Applied augmentation techniques such as shear, zoom, horizontal flip, and vertical flip.
- Utilized
ImageDataGeneratorfor splitting the dataset into training and validation subsets.
-
CNN (Convolutional Neural Network):
- Architecture: Convolutional layers with pooling followed by dense layers for feature learning and classification.
- Training: 27 epochs with early stopping callbacks and model checkpointing.
-
ResNet:
- Architecture: ResNet50 with pre-trained weights. Added a dense layer for final classification.
- Training: 30 epochs with early stopping callbacks and model checkpointing.
-
InceptionNet:
- Architecture: Built on InceptionV3 with pre-trained weights. Added a dense layer for classification.
- Training: 30 epochs with callbacks for performance optimization.
| Algorithm | Accuracy | F1 Score |
|---|---|---|
| CNN | 0.9173 | 0.9022 |
| InceptionNet_v3 | 0.9476 | 0.9016 |
| ResNet50 | 0.4407 | 0.4127 |
- Prediction Plots:
- Accuracy and F1 Score comparisons.
- Confusion matrices for each model.