Guava Disease Classification

Group 6 Members

• Muhammad Junaid Ali Asif Raja (M11217073)
• Muhammad Aown Ali (M11217077)

---

Project Overview

This project focuses on classifying guava diseases into three categories:

• Anthracnose
• Fruit Fly
• Healthy Guava

The dataset used for this project was downloaded from Kaggle: Guava Disease Dataset, and class folders were renamed accordingly for better organization.

---

Methodology

1. Model Architecture

The project leverages DenseNet Reloaded, a state-of-the-art architecture recently accepted in ECCV 2024. This architecture was implemented with the following enhancements:

• RDNet_base code was adapted from the popular timm library.
• Pretrained weights for rdnet_base were utilized for improved performance.

2. Optimization

The model was optimized using the ADOPT optimizer, a cutting-edge optimization algorithm accepted in NeurIPS 2024. This optimizer was adapted from the official repository: ADOPT GitHub.

---

Implementation Details

Dataset Processing

• Dataset: Kaggle: Guava Disease Dataset
• Class folders renamed to:
  • Anthracnose
  • Fruit Fly
  • Healthy Guava

Model Initialization

The model was initialized using pretrained weights:

model = timm.create_model('rdnet_base', pretrained=True)

Key References

• RDNet: GitHub Repository
  • RDNet Paper
• ADOPT Optimizer: GitHub Repository
  • ADOPT Paper

---

Results

Classification Report

               precision    recall  f1-score   support

  Anthracnose       1.00      1.00      1.00       156
    Fruit Fly       0.99      1.00      1.00       132
Healthy Guava       1.00      0.99      0.99        94

     accuracy                           1.00       382
    macro avg       1.00      1.00      1.00       382
 weighted avg       1.00      1.00      1.00       382

Test Metrics

• Test Top-1 Accuracy: 99.7382%

Confusion Matrix

The fine-tuned weights for this are available at this link.

---

Requirements

We also provide the requirements file in the forms of guava.yml and requirements.txt.

conda env create -f guava.yml

or

pip install -r requirements.txt

---

Acknowledgements

• RDNet_base code was exported from the timm library.
• ADOPT optimizer code was adapted from iShohei220/adopt.
• Dataset courtesy of Asadullah Galib.

---

License

This project is for academic purposes. Please refer to the original repositories for RDNet and ADOPT for licensing details.