Bean Leaf Disease Classification Project

Overview

This project focuses on the classification of bean leaf diseases into three categories:

1. Angular Leaf Spot
2. Bean Rust
3. Healthy Leaf

We utilized the dataset provided on Kaggle to train and evaluate our model. The project leverages transfer learning using the VGG16 architecture with fine-tuning to improve classification accuracy.

The repo doesn't contain saved model and the dataset folder because of size constraint from github. Instead, we use Google Drive to store it.

---

Dataset

The dataset consists of images of bean leaves categorized into three classes. The data is split as follows:

• Training and Validation Split: 80% training and 20% validation (from the train folder).
• Testing Split: All images from the val folder are used for evaluation.

---

Data Augmentation

To enhance the robustness of the model and prevent overfitting, we applied the following augmentations to the training data:

• Random Flip: Random horizontal and vertical flipping of images.
• Random Rotation: Rotations applied to images by up to 20%.
• Random Zoom: Zooming in and out on the image, with a height and width factor range of -20% to +20%.
• Random Contrast: Adjusting the contrast of the images by up to 20%.
• Random Brightness: Modifying the brightness of the images by up to 20%.

These augmentations simulate variability in the dataset and help the model generalize better to unseen data.

---

Model Architecture

We employed the VGG16 architecture, pre-trained on the ImageNet dataset, as our base model with the top layers removed (include_top=False).

Fine-Tuning

To adapt the model to our specific task, we added custom fully connected layers:

1. Fully connected (Dense) layers with ReLU activation.
2. Dropout layers for regularization.
3. A final Dense layer with softmax activation for multi-class classification.

---

Metrics

The following metrics were monitored during training and evaluation:

• Accuracy
• Precision
• Recall
• F1 Score

---

Results

Validation Data

• Accuracy: 97.09%
• Precision: 97.09%
• Recall: 97.09%
• F1 Score: 96.99%

Test Data

• Accuracy: 97.74%
• Precision: 97.74%
• Recall: 97.74%
• F1 Score: 97.77%

---

Model Export

The trained model was exported in the following formats for deployment:

1. SavedModel
2. TensorFlow.js

These formats enable flexible integration into different platforms, including mobile and web applications.