PERNOSPHERE is a project made for the course AI-lab Computer Vision and NLP 2024/25. In this project, I implemented a basic Convolutional Neural Network (CNN) to categorize the most common diseases in vinegrape plants.

Carlo Da Roma

• Basic PyTorch CNN
• Training with early stopping and methods to avoid overfitting
• Simple web Gradio interface
• Details and explanations about the diseases
• High accuracy test (0.95%+)

• The dataset contains images of vinegrape leaves, either diseased or healthy.
• Author: Rajarshi Mandal
• License: CC0 1.0 Universal

0. Install Gradio, OpenCV (cv2), PyTorch, Matplotlib, and Pandas

1. Download the dataset from the link below: [https://www.kaggle.com/datasets/rm1000/grape-disease-dataset-original]

2. MOVE ALL THE SUB-DIRECTORIES (test,train) OF THE DOWNLOADED ZIP INTO Archive/trainTest/leaf!

3. Create a usable dataset (CSV, images) by launching support_function_test.py and support_function_train.py

Run the training script:

python train_test.py

It will automatically stop.

Launch the Gradio GUI to use the app:

python app.py

• In the directory Code there are:

  • model.py – CNN definition
  • dataloader.py – Custom dataloader
  • train_test.py – Training and testing loop
  • predict.py – Predictor function
  • GUI.py – Gradio web app

• In the directory Archive there are the images of the dataset

• In the directory Model there is the saved binary model

• In the directory Dataset there are the .csv files and a support function to create a usable dataset

• The model input images are resized to 128x128. The original size of the images is 512x512
• Early stopping patience is set to 3 epochs by default
• Typically, training for 9 to 10 epochs is enough to reach a test accuracy of at least 95%

• To achieve the best results, take the photo of the leaf on a sheet of paper.

_{last update 06/06/2025}